Merge modeling files into a single one to avoid relative import
Browse files- config.json +1 -1
- configuration_xgenmm.py +0 -159
- image_processing_blip_3.py +12 -1
- modeling_xgenmm.py +2033 -40
- utils.py +0 -383
- vlm.py +0 -1381
config.json
CHANGED
@@ -3,7 +3,7 @@
|
|
3 |
"XGenMMModelForConditionalGeneration"
|
4 |
],
|
5 |
"auto_map": {
|
6 |
-
"AutoConfig": "
|
7 |
"AutoModelForVision2Seq": "modeling_xgenmm.XGenMMModelForConditionalGeneration"
|
8 |
},
|
9 |
"model_type": "xgenmm",
|
|
|
3 |
"XGenMMModelForConditionalGeneration"
|
4 |
],
|
5 |
"auto_map": {
|
6 |
+
"AutoConfig": "modeling_xgenmm.XGenMMConfig",
|
7 |
"AutoModelForVision2Seq": "modeling_xgenmm.XGenMMModelForConditionalGeneration"
|
8 |
},
|
9 |
"model_type": "xgenmm",
|
configuration_xgenmm.py
DELETED
@@ -1,159 +0,0 @@
|
|
1 |
-
from transformers import PretrainedConfig
|
2 |
-
from transformers import logging
|
3 |
-
from transformers import CONFIG_MAPPING
|
4 |
-
|
5 |
-
logger = logging.get_logger(__name__)
|
6 |
-
|
7 |
-
class XGenMMVisionEncoderConfig(PretrainedConfig):
|
8 |
-
model_type = "xgenmm_vision_encoder"
|
9 |
-
|
10 |
-
def __init__(self,
|
11 |
-
model_name: str = 'google/siglip-so400m-patch14-384',
|
12 |
-
anyres_grids: list[int] = [[384, 768],[768, 384],[768, 768],[1152, 384],[384,1152]],
|
13 |
-
**kwargs):
|
14 |
-
self.model_name = model_name
|
15 |
-
self.anyres_grids = anyres_grids
|
16 |
-
super().__init__(**kwargs)
|
17 |
-
|
18 |
-
|
19 |
-
class XGenMMVisionTokenizerConfig(PretrainedConfig):
|
20 |
-
model_type = "xgenmm_vision_tokenizer"
|
21 |
-
|
22 |
-
def __init__(self,
|
23 |
-
vis_feature_dim: int = 1152,
|
24 |
-
lang_embedding_dim: int = 3072,
|
25 |
-
num_vis_tokens: int = 128,
|
26 |
-
image_aspect_ratio: str = 'anyres',
|
27 |
-
**kwargs):
|
28 |
-
self.vis_feature_dim = vis_feature_dim
|
29 |
-
self.lang_embedding_dim = lang_embedding_dim
|
30 |
-
self.num_vis_tokens = num_vis_tokens
|
31 |
-
self.image_aspect_ratio = image_aspect_ratio
|
32 |
-
super().__init__(**kwargs)
|
33 |
-
|
34 |
-
|
35 |
-
class XGenMMConfig(PretrainedConfig):
|
36 |
-
model_type = "xgenmm"
|
37 |
-
|
38 |
-
def __init__(self,
|
39 |
-
vision_encoder_config: dict = None,
|
40 |
-
vision_tokenizer_config: dict = None,
|
41 |
-
text_config: dict = None,
|
42 |
-
**kwargs):
|
43 |
-
|
44 |
-
if vision_encoder_config is None:
|
45 |
-
vision_encoder_config = {'image_aspect_ratio': 'anyres', 'anyres_patch_sampling': True}
|
46 |
-
logger.info("vision_encoder_config is None. initializing the XGenMMVisionEncoderConfig with default values.")
|
47 |
-
|
48 |
-
if vision_tokenizer_config is None:
|
49 |
-
vision_tokenizer_config = {}
|
50 |
-
logger.info("vision_tokenizer_config is None. Initializing the XGenMMVisionTokenizerConfig with default values.")
|
51 |
-
|
52 |
-
if text_config is None:
|
53 |
-
text_config = {
|
54 |
-
'initial_tokenizer_len':32012,
|
55 |
-
'pad_token_id':32011,
|
56 |
-
'bos_token_id':1,
|
57 |
-
'eos_token_id':32000,
|
58 |
-
'vocab_size': 32064,
|
59 |
-
'hidden_size': 3072,
|
60 |
-
'intermediate_size': 8192,
|
61 |
-
'num_hidden_layers': 32,
|
62 |
-
'num_attention_heads': 32,
|
63 |
-
'num_key_value_heads': 32,
|
64 |
-
'resid_pdrop': 0.0,
|
65 |
-
'embd_pdrop': 0.0,
|
66 |
-
'attention_dropout': 0.0,
|
67 |
-
'hidden_act': 'silu',
|
68 |
-
'max_position_embeddings': 4096,
|
69 |
-
'original_max_position_embeddings': 4096,
|
70 |
-
'initializer_range': 0.02,
|
71 |
-
'rms_norm_eps': 1e-05,
|
72 |
-
'use_cache': True,
|
73 |
-
'rope_theta': 10000.0,
|
74 |
-
'rope_scaling': None,
|
75 |
-
'sliding_window': 2047,
|
76 |
-
'return_dict': True,
|
77 |
-
'output_hidden_states': False,
|
78 |
-
'output_attentions': False,
|
79 |
-
'torchscript': False,
|
80 |
-
'torch_dtype': 'bfloat16',
|
81 |
-
'use_bfloat16': False,
|
82 |
-
'tf_legacy_loss': False,
|
83 |
-
'pruned_heads': {},
|
84 |
-
'tie_word_embeddings': False,
|
85 |
-
'chunk_size_feed_forward': 0,
|
86 |
-
'is_encoder_decoder': False,
|
87 |
-
'is_decoder': False,
|
88 |
-
'cross_attention_hidden_size': None,
|
89 |
-
'add_cross_attention': False,
|
90 |
-
'tie_encoder_decoder': False,
|
91 |
-
'max_length': 20,
|
92 |
-
'min_length': 0,
|
93 |
-
'do_sample': False,
|
94 |
-
'early_stopping': False,
|
95 |
-
'num_beams': 1,
|
96 |
-
'num_beam_groups': 1,
|
97 |
-
'diversity_penalty': 0.0,
|
98 |
-
'temperature': 1.0,
|
99 |
-
'top_k': 50,
|
100 |
-
'top_p': 1.0,
|
101 |
-
'typical_p': 1.0,
|
102 |
-
'repetition_penalty': 1.0,
|
103 |
-
'length_penalty': 1.0,
|
104 |
-
'no_repeat_ngram_size': 0,
|
105 |
-
'encoder_no_repeat_ngram_size': 0,
|
106 |
-
'bad_words_ids': None,
|
107 |
-
'num_return_sequences': 1,
|
108 |
-
'output_scores': False,
|
109 |
-
'return_dict_in_generate': False,
|
110 |
-
'forced_bos_token_id': None,
|
111 |
-
'forced_eos_token_id': None,
|
112 |
-
'remove_invalid_values': False,
|
113 |
-
'exponential_decay_length_penalty': None,
|
114 |
-
'suppress_tokens': None,
|
115 |
-
'begin_suppress_tokens': None,
|
116 |
-
'finetuning_task': None,
|
117 |
-
'id2label': {0: 'LABEL_0', 1: 'LABEL_1'},
|
118 |
-
'label2id': {'LABEL_0': 0, 'LABEL_1': 1},
|
119 |
-
'tokenizer_class': None,
|
120 |
-
'prefix': None,
|
121 |
-
'bos_token_id': 1,
|
122 |
-
'pad_token_id': 32000,
|
123 |
-
'eos_token_id': 32000,
|
124 |
-
'sep_token_id': None,
|
125 |
-
'decoder_start_token_id': None,
|
126 |
-
'task_specific_params': None,
|
127 |
-
'problem_type': None,
|
128 |
-
'model_type': 'phi3'
|
129 |
-
}
|
130 |
-
logger.info("text_config is None. Initializing the text config with default values (`Phi3Config`).")
|
131 |
-
|
132 |
-
self.vision_encoder_config = XGenMMVisionEncoderConfig(**vision_encoder_config)
|
133 |
-
|
134 |
-
self.vision_tokenizer_config = XGenMMVisionTokenizerConfig(**vision_tokenizer_config)
|
135 |
-
|
136 |
-
text_model_type = text_config["model_type"] if "model_type" in text_config else "phi3"
|
137 |
-
self.text_config = CONFIG_MAPPING[text_model_type](**text_config)
|
138 |
-
|
139 |
-
for key in ['initial_tokenizer_len', 'pad_token_id']:
|
140 |
-
if key not in self.text_config.to_dict():
|
141 |
-
raise ValueError(f"The key `{key}` is missing in the text_config.")
|
142 |
-
|
143 |
-
super().__init__(**kwargs)
|
144 |
-
|
145 |
-
@classmethod
|
146 |
-
def from_vision_encoder_vision_tokenizer_text_configs(
|
147 |
-
cls,
|
148 |
-
vision_encoder_config: XGenMMVisionEncoderConfig,
|
149 |
-
vision_tokenizer_config: XGenMMVisionTokenizerConfig,
|
150 |
-
text_config: PretrainedConfig,
|
151 |
-
**kwargs):
|
152 |
-
|
153 |
-
return cls(
|
154 |
-
vision_encoder_config=vision_encoder_config.to_dict(),
|
155 |
-
vision_tokenizer_config=vision_tokenizer_config.to_dict(),
|
156 |
-
text_config=text_config.to_dict(),
|
157 |
-
**kwargs,
|
158 |
-
)
|
159 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
image_processing_blip_3.py
CHANGED
@@ -13,7 +13,18 @@ from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
|
|
13 |
from transformers.image_utils import ImageInput
|
14 |
from transformers.utils import TensorType
|
15 |
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
|
18 |
|
19 |
class Blip3ImageProcessor(BaseImageProcessor):
|
|
|
13 |
from transformers.image_utils import ImageInput
|
14 |
from transformers.utils import TensorType
|
15 |
|
16 |
+
def expand2square(pil_img, background_color):
|
17 |
+
width, height = pil_img.size
|
18 |
+
if width == height:
|
19 |
+
return pil_img
|
20 |
+
elif width > height:
|
21 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
22 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
23 |
+
return result
|
24 |
+
else:
|
25 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
26 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
27 |
+
return result
|
28 |
|
29 |
|
30 |
class Blip3ImageProcessor(BaseImageProcessor):
|
modeling_xgenmm.py
CHANGED
@@ -1,29 +1,2013 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
2 |
import torch
|
3 |
-
import
|
|
|
|
|
4 |
from typing import List, Optional, Tuple, Union
|
5 |
-
|
6 |
-
from .
|
7 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
class XGenMMVisionEncoder(PreTrainedModel):
|
10 |
main_input_name = "pixel_values"
|
11 |
config_class = XGenMMVisionEncoderConfig
|
12 |
-
|
13 |
def __init__(self, config: XGenMMVisionEncoderConfig):
|
14 |
super().__init__(config)
|
15 |
-
if config.model_name !=
|
16 |
-
raise ValueError(
|
|
|
|
|
17 |
self.model = AutoModel.from_pretrained(config.model_name)
|
18 |
-
|
19 |
def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
|
20 |
# assert pixel_values.ndim == 4, f"Expected 4D tensor (bs, c, h, w), got {pixel_values.ndim}"
|
21 |
return self.model.encode_image(pixel_values)
|
22 |
-
|
23 |
|
24 |
-
|
|
|
25 |
class XGenMMVisionTokenizer(PreTrainedModel):
|
26 |
config_class = XGenMMVisionTokenizerConfig
|
|
|
27 |
def __init__(self, config: XGenMMVisionTokenizerConfig):
|
28 |
super().__init__(config)
|
29 |
self.model = PerceiverResampler(
|
@@ -31,50 +2015,58 @@ class XGenMMVisionTokenizer(PreTrainedModel):
|
|
31 |
dim_inner=config.lang_embedding_dim,
|
32 |
num_latents=config.num_vis_tokens,
|
33 |
)
|
34 |
-
|
35 |
-
def forward(self,
|
36 |
-
vision_features: torch.Tensor,
|
37 |
-
vision_attn_masks: torch.Tensor):
|
38 |
return self.model(vision_features, vision_attn_masks)
|
39 |
-
|
|
|
40 |
# XGenMM model
|
41 |
class XGenMMModelForConditionalGeneration(PreTrainedModel):
|
42 |
config_class = XGenMMConfig
|
43 |
-
|
44 |
def __init__(self, config: XGenMMConfig):
|
45 |
super().__init__(config)
|
46 |
-
|
47 |
# vision encoder initialization
|
48 |
-
vision_encoder = AutoModel.from_pretrained(
|
49 |
-
|
50 |
-
|
|
|
|
|
51 |
language_model = AutoModelForCausalLM.from_config(config.text_config)
|
52 |
check_embedding_fns(language_model)
|
53 |
# Update _tied_weights_keys using the base model used.
|
54 |
if language_model._tied_weights_keys is not None:
|
55 |
-
self._tied_weights_keys = [
|
56 |
-
|
|
|
|
|
57 |
# vision tokenizer initialization
|
58 |
-
if
|
|
|
|
|
|
|
59 |
overwrite = language_model.get_input_embeddings().weight.shape[1]
|
60 |
config.vision_tokenizer_config.lang_embedding_dim = overwrite
|
61 |
-
print(
|
62 |
-
|
|
|
|
|
63 |
vision_tokenizer = XGenMMVisionTokenizer(config.vision_tokenizer_config).model
|
64 |
|
65 |
self.vlm = XGenMMPerceiver(
|
66 |
vision_encoder=vision_encoder,
|
67 |
vision_tokenizer=vision_tokenizer,
|
68 |
lang_model=language_model,
|
69 |
-
initial_tokenizer_len
|
70 |
-
pad_token_id
|
71 |
-
image_aspect_ratio
|
72 |
-
anyres_patch_sampling
|
73 |
-
anyres_grids
|
74 |
)
|
75 |
# Initialize weights and apply final processing
|
76 |
self.post_init()
|
77 |
-
|
78 |
@torch.no_grad()
|
79 |
def generate(
|
80 |
self,
|
@@ -82,14 +2074,15 @@ class XGenMMModelForConditionalGeneration(PreTrainedModel):
|
|
82 |
input_ids: Optional[torch.LongTensor] = None,
|
83 |
attention_mask: Optional[torch.LongTensor] = None,
|
84 |
**generate_kwargs,
|
85 |
-
|
86 |
self.vlm = self.vlm.eval()
|
87 |
return self.vlm.generate(
|
88 |
-
vision_x
|
89 |
-
lang_x
|
90 |
-
attention_mask
|
91 |
-
**generate_kwargs
|
92 |
-
|
|
|
93 |
def update_special_tokens(self, tokenizer):
|
94 |
tokenizer.add_special_tokens(
|
95 |
{"additional_special_tokens": list(self.vlm.special_tokens.values())}
|
@@ -97,8 +2090,8 @@ class XGenMMModelForConditionalGeneration(PreTrainedModel):
|
|
97 |
self.vlm.lang_model.config.vocab_size = len(tokenizer)
|
98 |
self.vlm.set_special_token_ids(
|
99 |
{
|
100 |
-
v: tokenizer.convert_tokens_to_ids(v)
|
|
|
101 |
}
|
102 |
)
|
103 |
return tokenizer
|
104 |
-
|
|
|
1 |
+
import ast
|
2 |
+
import math
|
3 |
+
from einops import rearrange, repeat
|
4 |
+
from einops_exts import rearrange_many
|
5 |
+
from einops import rearrange
|
6 |
+
from PIL import Image
|
7 |
import torch
|
8 |
+
from torch import einsum, nn
|
9 |
+
|
10 |
+
|
11 |
from typing import List, Optional, Tuple, Union
|
12 |
+
import torch.nn.functional as F
|
13 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
14 |
+
from dataclasses import dataclass
|
15 |
+
from transformers import CLIPVisionModel
|
16 |
+
from transformers import PreTrainedModel, AutoModelForCausalLM, AutoModel
|
17 |
+
from transformers import PretrainedConfig, logging, CONFIG_MAPPING
|
18 |
+
from transformers.models.siglip.modeling_siglip import SiglipVisionTransformer
|
19 |
+
|
20 |
+
|
21 |
+
logger = logging.get_logger(__name__)
|
22 |
+
|
23 |
+
|
24 |
+
class XGenMMVisionEncoderConfig(PretrainedConfig):
|
25 |
+
model_type = "xgenmm_vision_encoder"
|
26 |
+
|
27 |
+
def __init__(
|
28 |
+
self,
|
29 |
+
model_name: str = "google/siglip-so400m-patch14-384",
|
30 |
+
anyres_grids: list[int] = [
|
31 |
+
[384, 768],
|
32 |
+
[768, 384],
|
33 |
+
[768, 768],
|
34 |
+
[1152, 384],
|
35 |
+
[384, 1152],
|
36 |
+
],
|
37 |
+
**kwargs,
|
38 |
+
):
|
39 |
+
self.model_name = model_name
|
40 |
+
self.anyres_grids = anyres_grids
|
41 |
+
super().__init__(**kwargs)
|
42 |
+
|
43 |
+
|
44 |
+
class XGenMMVisionTokenizerConfig(PretrainedConfig):
|
45 |
+
model_type = "xgenmm_vision_tokenizer"
|
46 |
+
|
47 |
+
def __init__(
|
48 |
+
self,
|
49 |
+
vis_feature_dim: int = 1152,
|
50 |
+
lang_embedding_dim: int = 3072,
|
51 |
+
num_vis_tokens: int = 128,
|
52 |
+
image_aspect_ratio: str = "anyres",
|
53 |
+
**kwargs,
|
54 |
+
):
|
55 |
+
self.vis_feature_dim = vis_feature_dim
|
56 |
+
self.lang_embedding_dim = lang_embedding_dim
|
57 |
+
self.num_vis_tokens = num_vis_tokens
|
58 |
+
self.image_aspect_ratio = image_aspect_ratio
|
59 |
+
super().__init__(**kwargs)
|
60 |
+
|
61 |
+
|
62 |
+
class XGenMMConfig(PretrainedConfig):
|
63 |
+
model_type = "xgenmm"
|
64 |
+
|
65 |
+
def __init__(
|
66 |
+
self,
|
67 |
+
vision_encoder_config: dict = None,
|
68 |
+
vision_tokenizer_config: dict = None,
|
69 |
+
text_config: dict = None,
|
70 |
+
**kwargs,
|
71 |
+
):
|
72 |
+
|
73 |
+
if vision_encoder_config is None:
|
74 |
+
vision_encoder_config = {
|
75 |
+
"image_aspect_ratio": "anyres",
|
76 |
+
"anyres_patch_sampling": True,
|
77 |
+
}
|
78 |
+
logger.info(
|
79 |
+
"vision_encoder_config is None. initializing the XGenMMVisionEncoderConfig with default values."
|
80 |
+
)
|
81 |
+
|
82 |
+
if vision_tokenizer_config is None:
|
83 |
+
vision_tokenizer_config = {}
|
84 |
+
logger.info(
|
85 |
+
"vision_tokenizer_config is None. Initializing the XGenMMVisionTokenizerConfig with default values."
|
86 |
+
)
|
87 |
+
|
88 |
+
if text_config is None:
|
89 |
+
text_config = {
|
90 |
+
"initial_tokenizer_len": 32012,
|
91 |
+
"pad_token_id": 32011,
|
92 |
+
"bos_token_id": 1,
|
93 |
+
"eos_token_id": 32000,
|
94 |
+
"vocab_size": 32064,
|
95 |
+
"hidden_size": 3072,
|
96 |
+
"intermediate_size": 8192,
|
97 |
+
"num_hidden_layers": 32,
|
98 |
+
"num_attention_heads": 32,
|
99 |
+
"num_key_value_heads": 32,
|
100 |
+
"resid_pdrop": 0.0,
|
101 |
+
"embd_pdrop": 0.0,
|
102 |
+
"attention_dropout": 0.0,
|
103 |
+
"hidden_act": "silu",
|
104 |
+
"max_position_embeddings": 4096,
|
105 |
+
"original_max_position_embeddings": 4096,
|
106 |
+
"initializer_range": 0.02,
|
107 |
+
"rms_norm_eps": 1e-05,
|
108 |
+
"use_cache": True,
|
109 |
+
"rope_theta": 10000.0,
|
110 |
+
"rope_scaling": None,
|
111 |
+
"sliding_window": 2047,
|
112 |
+
"return_dict": True,
|
113 |
+
"output_hidden_states": False,
|
114 |
+
"output_attentions": False,
|
115 |
+
"torchscript": False,
|
116 |
+
"torch_dtype": "bfloat16",
|
117 |
+
"use_bfloat16": False,
|
118 |
+
"tf_legacy_loss": False,
|
119 |
+
"pruned_heads": {},
|
120 |
+
"tie_word_embeddings": False,
|
121 |
+
"chunk_size_feed_forward": 0,
|
122 |
+
"is_encoder_decoder": False,
|
123 |
+
"is_decoder": False,
|
124 |
+
"cross_attention_hidden_size": None,
|
125 |
+
"add_cross_attention": False,
|
126 |
+
"tie_encoder_decoder": False,
|
127 |
+
"max_length": 20,
|
128 |
+
"min_length": 0,
|
129 |
+
"do_sample": False,
|
130 |
+
"early_stopping": False,
|
131 |
+
"num_beams": 1,
|
132 |
+
"num_beam_groups": 1,
|
133 |
+
"diversity_penalty": 0.0,
|
134 |
+
"temperature": 1.0,
|
135 |
+
"top_k": 50,
|
136 |
+
"top_p": 1.0,
|
137 |
+
"typical_p": 1.0,
|
138 |
+
"repetition_penalty": 1.0,
|
139 |
+
"length_penalty": 1.0,
|
140 |
+
"no_repeat_ngram_size": 0,
|
141 |
+
"encoder_no_repeat_ngram_size": 0,
|
142 |
+
"bad_words_ids": None,
|
143 |
+
"num_return_sequences": 1,
|
144 |
+
"output_scores": False,
|
145 |
+
"return_dict_in_generate": False,
|
146 |
+
"forced_bos_token_id": None,
|
147 |
+
"forced_eos_token_id": None,
|
148 |
+
"remove_invalid_values": False,
|
149 |
+
"exponential_decay_length_penalty": None,
|
150 |
+
"suppress_tokens": None,
|
151 |
+
"begin_suppress_tokens": None,
|
152 |
+
"finetuning_task": None,
|
153 |
+
"id2label": {0: "LABEL_0", 1: "LABEL_1"},
|
154 |
+
"label2id": {"LABEL_0": 0, "LABEL_1": 1},
|
155 |
+
"tokenizer_class": None,
|
156 |
+
"prefix": None,
|
157 |
+
"bos_token_id": 1,
|
158 |
+
"pad_token_id": 32000,
|
159 |
+
"eos_token_id": 32000,
|
160 |
+
"sep_token_id": None,
|
161 |
+
"decoder_start_token_id": None,
|
162 |
+
"task_specific_params": None,
|
163 |
+
"problem_type": None,
|
164 |
+
"model_type": "phi3",
|
165 |
+
}
|
166 |
+
logger.info(
|
167 |
+
"text_config is None. Initializing the text config with default values (`Phi3Config`)."
|
168 |
+
)
|
169 |
+
|
170 |
+
self.vision_encoder_config = XGenMMVisionEncoderConfig(**vision_encoder_config)
|
171 |
+
|
172 |
+
self.vision_tokenizer_config = XGenMMVisionTokenizerConfig(
|
173 |
+
**vision_tokenizer_config
|
174 |
+
)
|
175 |
+
|
176 |
+
text_model_type = (
|
177 |
+
text_config["model_type"] if "model_type" in text_config else "phi3"
|
178 |
+
)
|
179 |
+
self.text_config = CONFIG_MAPPING[text_model_type](**text_config)
|
180 |
+
|
181 |
+
for key in ["initial_tokenizer_len", "pad_token_id"]:
|
182 |
+
if key not in self.text_config.to_dict():
|
183 |
+
raise ValueError(f"The key `{key}` is missing in the text_config.")
|
184 |
+
|
185 |
+
super().__init__(**kwargs)
|
186 |
+
|
187 |
+
|
188 |
+
def hasattr_recursive(obj, att):
|
189 |
+
"""
|
190 |
+
Check if obj has nested attribute
|
191 |
+
Example: hasattr_recursive(obj, 'a.b.c') is equivalent to hasattr(obj, 'a') and hasattr(obj.a, 'b') and hasattr(obj.a.b, 'c')
|
192 |
+
"""
|
193 |
+
if att == "":
|
194 |
+
return True
|
195 |
+
i = att.find(".")
|
196 |
+
if i < 0:
|
197 |
+
return hasattr(obj, att)
|
198 |
+
else:
|
199 |
+
try:
|
200 |
+
return hasattr_recursive(getattr(obj, att[:i]), att[i + 1 :])
|
201 |
+
except:
|
202 |
+
return False
|
203 |
+
|
204 |
+
|
205 |
+
def getattr_recursive(obj, att):
|
206 |
+
"""
|
207 |
+
Return nested attribute of obj
|
208 |
+
Example: getattr_recursive(obj, 'a.b.c') is equivalent to obj.a.b.c
|
209 |
+
"""
|
210 |
+
if att == "":
|
211 |
+
return obj
|
212 |
+
i = att.find(".")
|
213 |
+
if i < 0:
|
214 |
+
return getattr(obj, att)
|
215 |
+
else:
|
216 |
+
return getattr_recursive(getattr(obj, att[:i]), att[i + 1 :])
|
217 |
+
|
218 |
+
|
219 |
+
def setattr_recursive(obj, att, val):
|
220 |
+
"""
|
221 |
+
Set nested attribute of obj
|
222 |
+
Example: setattr_recursive(obj, 'a.b.c', val) is equivalent to obj.a.b.c = val
|
223 |
+
"""
|
224 |
+
if "." in att:
|
225 |
+
obj = getattr_recursive(obj, ".".join(att.split(".")[:-1]))
|
226 |
+
setattr(obj, att.split(".")[-1], val)
|
227 |
+
|
228 |
+
|
229 |
+
def check_embedding_fns(lang_model):
|
230 |
+
"""Checks for and attempts to set {get/set}_{input/output}_embeddings functions to the model"""
|
231 |
+
if not has_fn(lang_model, "get_input_embeddings"):
|
232 |
+
if hasattr_recursive(lang_model, "transformer.wte"): # MPT
|
233 |
+
lang_model.get_input_embeddings = lambda: lang_model.transformer.wte
|
234 |
+
elif hasattr_recursive(lang_model, "model.decoder.embed_tokens"): # OPT
|
235 |
+
lang_model.get_input_embeddings = lambda: lang_model.decoder.embed_tokens
|
236 |
+
else:
|
237 |
+
raise ValueError(
|
238 |
+
"We require the language encoder to have a get_input_embeddings method but we couldn't determine the name of the input embeddings attribute. Please supply this manually in factory.py."
|
239 |
+
)
|
240 |
+
|
241 |
+
if not has_fn(lang_model, "set_input_embeddings"):
|
242 |
+
if hasattr_recursive(lang_model, "transformer.wte"): # MPT
|
243 |
+
lang_model.set_input_embeddings = lambda x: setattr_recursive(
|
244 |
+
lang_model, "transformer.wte", x
|
245 |
+
)
|
246 |
+
elif hasattr_recursive(lang_model, "model.decoder.embed_tokens"): # OPT
|
247 |
+
lang_model.set_input_embeddings = lambda x: setattr_recursive(
|
248 |
+
lang_model, "model.decoder.embed_tokens", x
|
249 |
+
)
|
250 |
+
else:
|
251 |
+
raise ValueError(
|
252 |
+
"We require the language encoder to have a set_input_embeddings method but we couldn't determine the name of the input embeddings attribute. Please supply this manually in factory.py."
|
253 |
+
)
|
254 |
+
|
255 |
+
if not has_fn(lang_model, "get_output_embeddings"):
|
256 |
+
if hasattr_recursive(lang_model, "lm_head"):
|
257 |
+
lang_model.get_output_embeddings = lambda: lang_model.lm_head
|
258 |
+
else:
|
259 |
+
raise ValueError(
|
260 |
+
"We require the language encoder to have a get_output_embeddings method but we couldn't determine the name of the output embeddings attribute. Please supply this manually in factory.py."
|
261 |
+
)
|
262 |
+
|
263 |
+
if not has_fn(lang_model, "set_output_embeddings"):
|
264 |
+
if hasattr_recursive(lang_model, "lm_head"):
|
265 |
+
lang_model.set_output_embeddings = lambda x: setattr_recursive(
|
266 |
+
lang_model, "lm_head", x
|
267 |
+
)
|
268 |
+
else:
|
269 |
+
raise ValueError(
|
270 |
+
"We require the language encoder to have a set_output_embeddings method but we couldn't determine the name of the output embeddings attribute. Please supply this manually in factory.py."
|
271 |
+
)
|
272 |
+
|
273 |
+
|
274 |
+
def has_fn(model, fn_name):
|
275 |
+
"""Check if model has a function fn_name"""
|
276 |
+
return callable(getattr(model, fn_name, None))
|
277 |
+
|
278 |
+
|
279 |
+
def stack_with_padding(list_of_tensors, padding_value=0, padding_side="right"):
|
280 |
+
"""
|
281 |
+
Stack a list of tensors with padding on one side
|
282 |
+
Args:
|
283 |
+
list_of_tensors (list[torch.Tensor]): List of tensors to stack
|
284 |
+
padding_value (int, optional): Value to pad with. Defaults to 0.
|
285 |
+
padding_side (str, optional): Side to pad on. Defaults to "right".
|
286 |
+
Returns:
|
287 |
+
torch.Tensor: Stacked tensors
|
288 |
+
"""
|
289 |
+
max_tokens = max(tensor.size(0) for tensor in list_of_tensors)
|
290 |
+
padded_tensors = []
|
291 |
+
for tensor in list_of_tensors:
|
292 |
+
num_tokens = tensor.size(0)
|
293 |
+
if len(tensor.size()) == 1:
|
294 |
+
padding = torch.full(
|
295 |
+
(max_tokens - num_tokens,),
|
296 |
+
padding_value,
|
297 |
+
dtype=tensor.dtype,
|
298 |
+
device=tensor.device,
|
299 |
+
)
|
300 |
+
else:
|
301 |
+
padding = torch.full(
|
302 |
+
(max_tokens - num_tokens, tensor.size(1)),
|
303 |
+
padding_value,
|
304 |
+
dtype=tensor.dtype,
|
305 |
+
device=tensor.device,
|
306 |
+
)
|
307 |
+
padded_tensor = (
|
308 |
+
torch.cat((tensor, padding), dim=0)
|
309 |
+
if padding_side == "right"
|
310 |
+
else torch.cat((padding, tensor), dim=0)
|
311 |
+
)
|
312 |
+
padded_tensors.append(padded_tensor)
|
313 |
+
return torch.stack(padded_tensors)
|
314 |
+
|
315 |
+
|
316 |
+
def unpad_image(tensor, original_size, keep_original_shape=False):
|
317 |
+
"""
|
318 |
+
Unpads a PyTorch tensor of a padded and resized image.
|
319 |
+
|
320 |
+
Args:
|
321 |
+
tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
|
322 |
+
original_size (tuple): The original size of the image (height, width).
|
323 |
+
|
324 |
+
Returns:
|
325 |
+
torch.Tensor: The unpadded image tensor.
|
326 |
+
"""
|
327 |
+
original_width, original_height = original_size
|
328 |
+
current_height, current_width = tensor.shape[1:]
|
329 |
+
|
330 |
+
original_aspect_ratio = original_width / original_height
|
331 |
+
current_aspect_ratio = current_width / current_height
|
332 |
+
|
333 |
+
if original_aspect_ratio > current_aspect_ratio:
|
334 |
+
scale_factor = current_width / original_width
|
335 |
+
new_height = int(original_height * scale_factor)
|
336 |
+
padding = (current_height - new_height) // 2
|
337 |
+
if keep_original_shape:
|
338 |
+
attention_mask = torch.ones(
|
339 |
+
(current_height, current_width), device=tensor.device
|
340 |
+
)
|
341 |
+
attention_mask[:padding, :] = 0
|
342 |
+
attention_mask[current_height - padding :, :] = 0
|
343 |
+
return tensor, attention_mask
|
344 |
+
else:
|
345 |
+
unpadded_tensor = tensor[:, padding : current_height - padding, :]
|
346 |
+
return unpadded_tensor, None
|
347 |
+
else:
|
348 |
+
scale_factor = current_height / original_height
|
349 |
+
new_width = int(original_width * scale_factor)
|
350 |
+
padding = (current_width - new_width) // 2
|
351 |
+
if keep_original_shape:
|
352 |
+
attention_mask = torch.ones(
|
353 |
+
(current_height, current_width), device=tensor.device
|
354 |
+
)
|
355 |
+
attention_mask[:, :padding] = 0
|
356 |
+
attention_mask[:, current_width - padding :] = 0
|
357 |
+
return tensor, attention_mask
|
358 |
+
else:
|
359 |
+
unpadded_tensor = tensor[:, :, padding : current_width - padding]
|
360 |
+
return unpadded_tensor, None
|
361 |
+
|
362 |
+
|
363 |
+
def select_best_resolution(original_size, possible_resolutions):
|
364 |
+
"""
|
365 |
+
Selects the best resolution from a list of possible resolutions based on the original size.
|
366 |
+
|
367 |
+
Args:
|
368 |
+
original_size (tuple): The original size of the image in the format (width, height).
|
369 |
+
possible_resolutions (list): A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
|
370 |
+
|
371 |
+
Returns:
|
372 |
+
tuple: The best fit resolution in the format (width, height).
|
373 |
+
"""
|
374 |
+
original_width, original_height = original_size
|
375 |
+
best_fit = None
|
376 |
+
max_effective_resolution = 0
|
377 |
+
min_wasted_resolution = float("inf")
|
378 |
+
|
379 |
+
for width, height in possible_resolutions:
|
380 |
+
scale = min(width / original_width, height / original_height)
|
381 |
+
downscaled_width, downscaled_height = int(original_width * scale), int(
|
382 |
+
original_height * scale
|
383 |
+
)
|
384 |
+
effective_resolution = min(
|
385 |
+
downscaled_width * downscaled_height, original_width * original_height
|
386 |
+
)
|
387 |
+
wasted_resolution = (width * height) - effective_resolution
|
388 |
+
|
389 |
+
if effective_resolution > max_effective_resolution or (
|
390 |
+
effective_resolution == max_effective_resolution
|
391 |
+
and wasted_resolution < min_wasted_resolution
|
392 |
+
):
|
393 |
+
max_effective_resolution = effective_resolution
|
394 |
+
min_wasted_resolution = wasted_resolution
|
395 |
+
best_fit = (width, height)
|
396 |
+
|
397 |
+
return best_fit
|
398 |
+
|
399 |
+
|
400 |
+
def resize_and_pad_image(image, target_resolution):
|
401 |
+
"""
|
402 |
+
Resize and pad an image to a target resolution while maintaining aspect ratio.
|
403 |
+
|
404 |
+
Args:
|
405 |
+
image (PIL.Image.Image): The input image.
|
406 |
+
target_resolution (tuple): The target resolution (width, height) of the image.
|
407 |
+
|
408 |
+
Returns:
|
409 |
+
PIL.Image.Image: The resized and padded image.
|
410 |
+
"""
|
411 |
+
original_width, original_height = image.size
|
412 |
+
target_width, target_height = target_resolution
|
413 |
+
|
414 |
+
scale_w = target_width / original_width
|
415 |
+
scale_h = target_height / original_height
|
416 |
+
|
417 |
+
if scale_w < scale_h:
|
418 |
+
new_width = target_width
|
419 |
+
new_height = min(math.ceil(original_height * scale_w), target_height)
|
420 |
+
else:
|
421 |
+
new_height = target_height
|
422 |
+
new_width = min(math.ceil(original_width * scale_h), target_width)
|
423 |
+
|
424 |
+
# Resize the image
|
425 |
+
resized_image = image.resize((new_width, new_height))
|
426 |
+
|
427 |
+
new_image = Image.new("RGB", (target_width, target_height), (0, 0, 0))
|
428 |
+
paste_x = (target_width - new_width) // 2
|
429 |
+
paste_y = (target_height - new_height) // 2
|
430 |
+
new_image.paste(resized_image, (paste_x, paste_y))
|
431 |
+
|
432 |
+
return new_image
|
433 |
+
|
434 |
+
|
435 |
+
def divide_to_patches(image, patch_size):
|
436 |
+
"""
|
437 |
+
Divides an image into patches of a specified size.
|
438 |
+
|
439 |
+
Args:
|
440 |
+
image (PIL.Image.Image): The input image.
|
441 |
+
patch_size (int): The size of each patch.
|
442 |
+
|
443 |
+
Returns:
|
444 |
+
list: A list of PIL.Image.Image objects representing the patches.
|
445 |
+
"""
|
446 |
+
patches = []
|
447 |
+
width, height = image.size
|
448 |
+
for i in range(0, height, patch_size):
|
449 |
+
for j in range(0, width, patch_size):
|
450 |
+
box = (j, i, j + patch_size, i + patch_size)
|
451 |
+
patch = image.crop(box)
|
452 |
+
patches.append(patch)
|
453 |
+
|
454 |
+
return patches
|
455 |
+
|
456 |
+
|
457 |
+
def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
|
458 |
+
"""
|
459 |
+
Calculate the shape of the image patch grid after the preprocessing for images of any resolution.
|
460 |
+
|
461 |
+
Args:
|
462 |
+
image_size (tuple): The size of the input image in the format (width, height).
|
463 |
+
grid_pinpoints (str): A string representation of a list of possible resolutions.
|
464 |
+
patch_size (int): The size of each image patch.
|
465 |
+
|
466 |
+
Returns:
|
467 |
+
tuple: The shape of the image patch grid in the format (width, height).
|
468 |
+
"""
|
469 |
+
if type(grid_pinpoints) is list:
|
470 |
+
possible_resolutions = grid_pinpoints
|
471 |
+
else:
|
472 |
+
possible_resolutions = ast.literal_eval(grid_pinpoints)
|
473 |
+
width, height = select_best_resolution(image_size, possible_resolutions)
|
474 |
+
return width // patch_size, height // patch_size
|
475 |
+
|
476 |
+
|
477 |
+
def process_anyres_image(image, processor, grid_pinpoints):
|
478 |
+
"""
|
479 |
+
Process an image with variable resolutions.
|
480 |
+
|
481 |
+
Args:
|
482 |
+
image (PIL.Image.Image): The input image to be processed.
|
483 |
+
processor: The image processor object.
|
484 |
+
grid_pinpoints (str): A string representation of a list of possible resolutions.
|
485 |
+
|
486 |
+
Returns:
|
487 |
+
torch.Tensor: A tensor containing the processed image patches.
|
488 |
+
"""
|
489 |
+
# FIXME: determine grid_pinpoints from image sizes.
|
490 |
+
if type(grid_pinpoints) is list:
|
491 |
+
possible_resolutions = grid_pinpoints
|
492 |
+
else:
|
493 |
+
possible_resolutions = ast.literal_eval(grid_pinpoints)
|
494 |
+
best_resolution = select_best_resolution(image.size, possible_resolutions)
|
495 |
+
image_padded = resize_and_pad_image(image, best_resolution)
|
496 |
+
|
497 |
+
processor_size = processor.transforms[0].size
|
498 |
+
patches = divide_to_patches(image_padded, processor_size[0])
|
499 |
+
|
500 |
+
image_original_resize = image.resize((processor_size[0], processor_size[0]))
|
501 |
+
|
502 |
+
image_patches = [image_original_resize] + patches
|
503 |
+
image_patches = [processor(image_patch) for image_patch in image_patches]
|
504 |
+
return torch.stack(image_patches, dim=0)
|
505 |
+
|
506 |
+
|
507 |
+
def expand2square(pil_img, background_color):
|
508 |
+
width, height = pil_img.size
|
509 |
+
if width == height:
|
510 |
+
return pil_img
|
511 |
+
elif width > height:
|
512 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
513 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
514 |
+
return result
|
515 |
+
else:
|
516 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
517 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
518 |
+
return result
|
519 |
+
|
520 |
+
|
521 |
+
class VisionTokenizer(nn.Module):
|
522 |
+
def __init__(self, dim_media, num_tokens_per_media):
|
523 |
+
super().__init__()
|
524 |
+
self.dim_media = dim_media
|
525 |
+
self.num_tokens_per_media = num_tokens_per_media
|
526 |
+
|
527 |
+
|
528 |
+
class PerceiverAttention(nn.Module):
|
529 |
+
def __init__(self, *, dim, dim_head=64, heads=8):
|
530 |
+
super().__init__()
|
531 |
+
self.scale = dim_head**-0.5
|
532 |
+
self.heads = heads
|
533 |
+
inner_dim = dim_head * heads
|
534 |
+
|
535 |
+
self.norm_media = nn.LayerNorm(dim)
|
536 |
+
self.norm_latents = nn.LayerNorm(dim)
|
537 |
+
|
538 |
+
self.to_q = nn.Linear(dim, inner_dim, bias=False)
|
539 |
+
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
|
540 |
+
self.to_out = nn.Linear(inner_dim, dim, bias=False)
|
541 |
+
|
542 |
+
def forward(self, x, latents, vision_attn_masks=None):
|
543 |
+
"""
|
544 |
+
Args:
|
545 |
+
x (torch.Tensor): image features
|
546 |
+
shape (b, T, n1, D)
|
547 |
+
latent (torch.Tensor): latent features
|
548 |
+
shape (b, T, n2, D)
|
549 |
+
"""
|
550 |
+
x = self.norm_media(x)
|
551 |
+
latents = self.norm_latents(latents)
|
552 |
+
|
553 |
+
h = self.heads
|
554 |
+
|
555 |
+
q = self.to_q(latents)
|
556 |
+
kv_input = torch.cat(
|
557 |
+
(x, latents), dim=-2
|
558 |
+
) # TODO: Change the shape of vision attention mask according to this.
|
559 |
+
if vision_attn_masks is not None:
|
560 |
+
vision_attn_masks = torch.cat(
|
561 |
+
(
|
562 |
+
vision_attn_masks,
|
563 |
+
torch.ones(
|
564 |
+
(latents.shape[0], latents.shape[-2]),
|
565 |
+
dtype=latents.dtype,
|
566 |
+
device=latents.device,
|
567 |
+
),
|
568 |
+
),
|
569 |
+
dim=-1,
|
570 |
+
)
|
571 |
+
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
|
572 |
+
q, k, v = rearrange_many((q, k, v), "b t n (h d) -> b h t n d", h=h)
|
573 |
+
q = q * self.scale
|
574 |
+
|
575 |
+
# attention
|
576 |
+
sim = einsum("... i d, ... j d -> ... i j", q, k)
|
577 |
+
# Apply vision attention mask here.
|
578 |
+
# Reference: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html#torch.nn.functional.scaled_dot_product_attention
|
579 |
+
if vision_attn_masks is not None:
|
580 |
+
attn_bias = torch.zeros(
|
581 |
+
(q.size(0), 1, 1, q.size(-2), k.size(-2)),
|
582 |
+
dtype=q.dtype,
|
583 |
+
device=q.device,
|
584 |
+
)
|
585 |
+
vision_attn_masks = repeat(
|
586 |
+
vision_attn_masks, "b n -> b 1 1 l n", l=q.size(-2)
|
587 |
+
)
|
588 |
+
attn_bias.masked_fill_(vision_attn_masks.logical_not(), float("-inf"))
|
589 |
+
sim += attn_bias
|
590 |
+
|
591 |
+
sim = sim - sim.amax(dim=-1, keepdim=True).detach()
|
592 |
+
attn = sim.softmax(dim=-1)
|
593 |
+
|
594 |
+
out = einsum("... i j, ... j d -> ... i d", attn, v)
|
595 |
+
out = rearrange(out, "b h t n d -> b t n (h d)", h=h)
|
596 |
+
return self.to_out(out)
|
597 |
+
|
598 |
+
|
599 |
+
def FeedForward(dim, mult=4):
|
600 |
+
inner_dim = int(dim * mult)
|
601 |
+
return nn.Sequential(
|
602 |
+
nn.LayerNorm(dim),
|
603 |
+
nn.Linear(dim, inner_dim, bias=False),
|
604 |
+
nn.GELU(),
|
605 |
+
nn.Linear(inner_dim, dim, bias=False),
|
606 |
+
)
|
607 |
+
|
608 |
+
|
609 |
+
def num_params(module, filter_to_trainable=False):
|
610 |
+
"""Returns the number of parameters in the module, or optionally only the trainable parameters"""
|
611 |
+
if filter_to_trainable:
|
612 |
+
return sum(p.numel() for p in module.parameters() if p.requires_grad)
|
613 |
+
else:
|
614 |
+
return sum(p.numel() for p in module.parameters())
|
615 |
+
|
616 |
+
|
617 |
+
class PerceiverResampler(VisionTokenizer):
|
618 |
+
def __init__(
|
619 |
+
self,
|
620 |
+
*,
|
621 |
+
dim,
|
622 |
+
dim_inner=None,
|
623 |
+
depth=6,
|
624 |
+
dim_head=96,
|
625 |
+
heads=16,
|
626 |
+
num_latents=128,
|
627 |
+
max_num_media=None,
|
628 |
+
max_num_frames=None,
|
629 |
+
ff_mult=4,
|
630 |
+
):
|
631 |
+
"""
|
632 |
+
Perceiver module which takes in image features and outputs image tokens.
|
633 |
+
Args:
|
634 |
+
dim (int): dimension of the incoming image features
|
635 |
+
dim_inner (int, optional): final dimension to project the incoming image features to;
|
636 |
+
also the final dimension of the outputted features. If None, no projection is used, and dim_inner = dim.
|
637 |
+
depth (int, optional): number of layers. Defaults to 6.
|
638 |
+
dim_head (int, optional): dimension of each head. Defaults to 64.
|
639 |
+
heads (int, optional): number of heads. Defaults to 8.
|
640 |
+
num_latents (int, optional): number of latent tokens to use in the Perceiver;
|
641 |
+
also corresponds to number of tokens per sequence to output. Defaults to 64.
|
642 |
+
max_num_media (int, optional): maximum number of media per sequence to input into the Perceiver
|
643 |
+
and keep positional embeddings for. If None, no positional embeddings are used.
|
644 |
+
max_num_frames (int, optional): maximum number of frames to input into the Perceiver
|
645 |
+
and keep positional embeddings for. If None, no positional embeddings are used.
|
646 |
+
ff_mult (int, optional): dimension multiplier for the feedforward network. Defaults to 4.
|
647 |
+
"""
|
648 |
+
if dim_inner is not None:
|
649 |
+
projection = nn.Linear(dim, dim_inner)
|
650 |
+
else:
|
651 |
+
projection = None
|
652 |
+
dim_inner = dim
|
653 |
+
super().__init__(dim_media=dim, num_tokens_per_media=num_latents)
|
654 |
+
self.projection = projection
|
655 |
+
self.latents = nn.Parameter(torch.randn(num_latents, dim))
|
656 |
+
|
657 |
+
# positional embeddings
|
658 |
+
self.frame_embs = (
|
659 |
+
nn.Parameter(torch.randn(max_num_frames, dim))
|
660 |
+
if exists(max_num_frames)
|
661 |
+
else None
|
662 |
+
)
|
663 |
+
self.media_time_embs = (
|
664 |
+
nn.Parameter(torch.randn(max_num_media, 1, dim))
|
665 |
+
if exists(max_num_media)
|
666 |
+
else None
|
667 |
+
)
|
668 |
+
|
669 |
+
self.layers = nn.ModuleList([])
|
670 |
+
for _ in range(depth):
|
671 |
+
self.layers.append(
|
672 |
+
nn.ModuleList(
|
673 |
+
[
|
674 |
+
PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
|
675 |
+
FeedForward(dim=dim, mult=ff_mult),
|
676 |
+
]
|
677 |
+
)
|
678 |
+
)
|
679 |
+
|
680 |
+
self.norm = nn.LayerNorm(dim)
|
681 |
+
|
682 |
+
def forward(self, x, vision_attn_masks):
|
683 |
+
"""
|
684 |
+
Args:
|
685 |
+
x (torch.Tensor): image features
|
686 |
+
shape (b, T, F, v, D)
|
687 |
+
vision_attn_masks (torch.Tensor): attention masks for padded visiont tokens (i.e., x)
|
688 |
+
shape (b, v)
|
689 |
+
Returns:
|
690 |
+
shape (b, T, n, D) where n is self.num_latents
|
691 |
+
"""
|
692 |
+
b, T, F, v = x.shape[:4]
|
693 |
+
|
694 |
+
# frame and media time embeddings
|
695 |
+
if exists(self.frame_embs):
|
696 |
+
frame_embs = repeat(self.frame_embs[:F], "F d -> b T F v d", b=b, T=T, v=v)
|
697 |
+
x = x + frame_embs
|
698 |
+
x = rearrange(
|
699 |
+
x, "b T F v d -> b T (F v) d"
|
700 |
+
) # flatten the frame and spatial dimensions
|
701 |
+
if exists(self.media_time_embs):
|
702 |
+
x = x + self.media_time_embs[:T]
|
703 |
+
|
704 |
+
# blocks
|
705 |
+
latents = self.latents
|
706 |
+
latents = repeat(latents, "n d -> b T n d", b=b, T=T)
|
707 |
+
for attn, ff in self.layers:
|
708 |
+
latents = attn(x, latents, vision_attn_masks) + latents
|
709 |
+
latents = ff(latents) + latents
|
710 |
+
|
711 |
+
if exists(self.projection):
|
712 |
+
return self.projection(self.norm(latents))
|
713 |
+
else:
|
714 |
+
return self.norm(latents)
|
715 |
+
|
716 |
+
|
717 |
+
class DecoupledEmbedding(nn.Embedding):
|
718 |
+
# Derived from https://pytorch.org/docs/stable/_modules/torch/nn/modules/sparse.html#Embedding
|
719 |
+
"""
|
720 |
+
Implements a decoupling of parameters to allow freezing (or not) a subset of the embeddings. In practise, the
|
721 |
+
regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `num_additional_embeddings` > 0,
|
722 |
+
then it will create `num_additional_embeddings` additional parameters that are always trained. If
|
723 |
+
`num_additional_embeddings=0`, then the module defaults back to the regular behavior of `nn.Embedding`.
|
724 |
+
"""
|
725 |
+
|
726 |
+
def __init__(
|
727 |
+
self,
|
728 |
+
max_original_id: int,
|
729 |
+
num_additional_embeddings: int = 0,
|
730 |
+
_weight: torch.Tensor = None,
|
731 |
+
num_original_embeddings: int = None,
|
732 |
+
embedding_dim: int = None,
|
733 |
+
partially_freeze=True,
|
734 |
+
device=None,
|
735 |
+
dtype=None,
|
736 |
+
pad_token_id=None,
|
737 |
+
) -> None:
|
738 |
+
"""
|
739 |
+
Args:
|
740 |
+
max_original_id (`int`):
|
741 |
+
The largest token id that should be embedded using the regular embedding (regular `weight`).
|
742 |
+
This is usually len(tokenizer) - 1 before additional tokens are added.
|
743 |
+
Note that this may not equal self.weight.shape[0]
|
744 |
+
num_additional_embeddings (`int`):
|
745 |
+
Number of additional tokens to initialize an Embedding matrix for (`additional_weight`).
|
746 |
+
_weight (`torch.Tensor`, *optional*, defaults to `None`): The regular weight tensor.
|
747 |
+
If provided, this sets the `num_original_embeddings` and `embedding_dim` parameters.
|
748 |
+
num_original_embeddings (`int`):
|
749 |
+
self.weight.shape[0]
|
750 |
+
embedding_dim (`int`):
|
751 |
+
The size of each embedding vector
|
752 |
+
partially_freeze: (`bool`, *optional*, defaults to `True`):
|
753 |
+
If `True`, the regular `weight` will be frozen. `additional_weight` is never frozen.
|
754 |
+
padding_idx (`int`, *optional*):
|
755 |
+
The padding index (needs to be less than num_embeddings)
|
756 |
+
|
757 |
+
Note: there are a lot of other parameters to initialize a standard `nn.Embedding` such as `padding_idx`,
|
758 |
+
`max_norm` or `norm_type`. We are not supporting these.
|
759 |
+
"""
|
760 |
+
# validate args
|
761 |
+
if pad_token_id is not None and pad_token_id > max_original_id:
|
762 |
+
raise ValueError(
|
763 |
+
f"pad_token_id must be <= max_original_id. Got {pad_token_id} and {max_original_id}."
|
764 |
+
+ "If the original tokenizer does not have a pad_token_id, use pad_token_id=None."
|
765 |
+
)
|
766 |
+
if _weight is not None:
|
767 |
+
assert (num_original_embeddings is None) or (
|
768 |
+
_weight.shape[0] == num_original_embeddings
|
769 |
+
), f"num_original_embeddings={num_original_embeddings} but _weight.shape[0]={_weight.shape[0]}"
|
770 |
+
assert (embedding_dim is None) or (
|
771 |
+
_weight.shape[1] == embedding_dim
|
772 |
+
), f"embedding_dim={embedding_dim} but _weight.shape[1]={_weight.shape[1]}"
|
773 |
+
num_original_embeddings = _weight.shape[0]
|
774 |
+
embedding_dim = _weight.shape[1]
|
775 |
+
else:
|
776 |
+
assert (
|
777 |
+
num_original_embeddings is not None
|
778 |
+
), "num_original_embeddings must be provided if _weight is not provided"
|
779 |
+
assert (
|
780 |
+
embedding_dim is not None
|
781 |
+
), "embedding_dim must be provided if _weight is not provided"
|
782 |
+
|
783 |
+
super().__init__(
|
784 |
+
num_embeddings=num_original_embeddings,
|
785 |
+
embedding_dim=embedding_dim,
|
786 |
+
device=device,
|
787 |
+
dtype=dtype,
|
788 |
+
padding_idx=pad_token_id,
|
789 |
+
_weight=_weight,
|
790 |
+
)
|
791 |
+
self.max_original_id = max_original_id
|
792 |
+
self.padding_idx = pad_token_id
|
793 |
+
self.num_additional_embeddings = num_additional_embeddings
|
794 |
+
if self.num_additional_embeddings > 0:
|
795 |
+
self.additional_embedding = nn.Embedding(
|
796 |
+
num_embeddings=self.num_additional_embeddings,
|
797 |
+
embedding_dim=embedding_dim,
|
798 |
+
device=device,
|
799 |
+
dtype=dtype,
|
800 |
+
)
|
801 |
+
self.set_requires_grad(
|
802 |
+
require_regular_grad=not partially_freeze, require_additional_grad=True
|
803 |
+
)
|
804 |
+
|
805 |
+
def set_requires_grad(self, require_regular_grad, require_additional_grad):
|
806 |
+
"""
|
807 |
+
Helper function to separately set the requires_grad flag for the regular weight and the additional weight.
|
808 |
+
"""
|
809 |
+
self.weight.requires_grad_(require_regular_grad)
|
810 |
+
self.additional_embedding.requires_grad_(require_additional_grad)
|
811 |
+
|
812 |
+
def forward(self, input_ids):
|
813 |
+
"""
|
814 |
+
we have 2 embeddings, with different indices - one pretrained self.weight and another
|
815 |
+
self.additional_embedding.weight that is being trained.
|
816 |
+
|
817 |
+
in order to make a lookup of the input ids, we:
|
818 |
+
1. find out the indices of the entries belonging to the 2nd embedding
|
819 |
+
2. extract those values while subtracting the size of the first embedding (num_embeddings), since the 2nd
|
820 |
+
embedding starts from 0 and not num_embeddings
|
821 |
+
3. perform the 2nd embedding lookup
|
822 |
+
4. now we handle the 1st embedding, we overwrite indices belonging to the 2nd embedding with a padding index
|
823 |
+
5. perform the 1st embedding lookup
|
824 |
+
6. now we overwrite the values in the 1st embedding lookup with the values of the 2nd embedding lookup
|
825 |
+
|
826 |
+
note: for the 1st embedding lookup we could have looked up only the low indices and not do the padding, but
|
827 |
+
then we have to create a new tensor and populate it with 2 tensors that are spread out across various indices -
|
828 |
+
i.e. not a simple concat - I haven't benchmarked the complex case if it's any faster, given that seqlens are
|
829 |
+
usually relatively short it's probably not faster or if faster not by much - but might be a good idea to
|
830 |
+
measure.
|
831 |
+
|
832 |
+
"""
|
833 |
+
if self.num_additional_embeddings == 0:
|
834 |
+
return F.embedding(input_ids, self.weight)
|
835 |
+
|
836 |
+
# Clone so that we don't modify the original input_ids later on
|
837 |
+
input_ids = input_ids.clone()
|
838 |
+
additional_vocab_indices = torch.where(input_ids > self.max_original_id)
|
839 |
+
input_ids_additional_vocab = input_ids[additional_vocab_indices]
|
840 |
+
additional_embeddings = self.additional_embedding(
|
841 |
+
input_ids_additional_vocab - self.max_original_id - 1
|
842 |
+
)
|
843 |
+
|
844 |
+
# for successful lookup replace input_ids with 0, the results of these will be discarded anyway
|
845 |
+
input_ids[additional_vocab_indices] = 0
|
846 |
+
full_vector = F.embedding(input_ids, self.weight)
|
847 |
+
|
848 |
+
# overwrite the records with high indices
|
849 |
+
full_vector[additional_vocab_indices] = additional_embeddings
|
850 |
+
|
851 |
+
return full_vector
|
852 |
+
|
853 |
+
def extra_repr(self) -> str:
|
854 |
+
return "num_original_embeddings={}, num_additional_embeddings={}, embedding_dim={}, partially_freeze={}".format(
|
855 |
+
self.max_original_id + 1,
|
856 |
+
self.num_additional_embeddings,
|
857 |
+
self.embedding_dim,
|
858 |
+
(not self.weight.requires_grad),
|
859 |
+
)
|
860 |
+
|
861 |
+
|
862 |
+
class DecoupledLinear(nn.Linear):
|
863 |
+
# Derived from https://pytorch.org/docs/stable/_modules/torch/nn/modules/linear.html#Linear
|
864 |
+
"""
|
865 |
+
Implements a decoupling of parameters to allow freezing (or not) a subset of the parameters. In practise, the
|
866 |
+
regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `additional_out_features` > 0,
|
867 |
+
then it will create `additional_out_features * in_features` additional parameters that are always trained. If
|
868 |
+
`additional_out_features=0`, then the module defaults back to the regular behavior of `nn.Linear`.
|
869 |
+
"""
|
870 |
+
|
871 |
+
def __init__(
|
872 |
+
self,
|
873 |
+
max_original_id: int,
|
874 |
+
additional_out_features: int = 0,
|
875 |
+
_weight: torch.Tensor = None,
|
876 |
+
_bias: torch.Tensor = None,
|
877 |
+
in_features: int = None,
|
878 |
+
original_out_features: int = None,
|
879 |
+
bias: bool = True,
|
880 |
+
partially_freeze: bool = True,
|
881 |
+
device=None,
|
882 |
+
dtype=None,
|
883 |
+
) -> None:
|
884 |
+
"""
|
885 |
+
Args:
|
886 |
+
max_original_id (`int`): The largest token id that should be extracted from the regular weight.
|
887 |
+
This is usually len(tokenizer) - 1 before additional tokens are added.
|
888 |
+
Note that this may not equal original_out_features - 1
|
889 |
+
_weight: torch.Tensor, *optional*, defaults to `None`. The regular weight tensor.
|
890 |
+
If provided, this sets the `in_features` and `original_out_features` parameters.
|
891 |
+
_bias: torch.Tensor, *optional*, defaults to `None`. The regular bias tensor.
|
892 |
+
in_features: int. Input hidden size.
|
893 |
+
original_out_features: int. Original out_features of the language model's get_output_embeddings() function.
|
894 |
+
additional_out_features: int. Number of additional trainable dimensions.
|
895 |
+
bias: bool. Whether to include a bias term.
|
896 |
+
partially_freeze: bool, *optional*, defaults to `True`): If `True`, the regular `weight` will be frozen.
|
897 |
+
"""
|
898 |
+
# argument validation
|
899 |
+
if _weight is not None:
|
900 |
+
assert (_weight.shape[0] == original_out_features) or (
|
901 |
+
original_out_features is None
|
902 |
+
), f"original_out_features={original_out_features} but _weight.shape[0]={_weight.shape[0]}"
|
903 |
+
assert (_weight.shape[1] == in_features) or (
|
904 |
+
in_features is None
|
905 |
+
), f"in_features={in_features} but _weight.shape[1]={_weight.shape[1]}"
|
906 |
+
in_features = _weight.shape[1]
|
907 |
+
original_out_features = _weight.shape[0]
|
908 |
+
else:
|
909 |
+
assert (
|
910 |
+
in_features is not None
|
911 |
+
), "in_features must be provided if _weight is not provided"
|
912 |
+
assert (
|
913 |
+
original_out_features is not None
|
914 |
+
), "original_out_features must be provided if _weight is not provided"
|
915 |
+
|
916 |
+
if _bias is not None:
|
917 |
+
assert bias is True, "bias must be True if _bias is provided"
|
918 |
+
|
919 |
+
# initialize original linear
|
920 |
+
super().__init__(in_features, original_out_features, bias, device, dtype)
|
921 |
+
|
922 |
+
# set weight and bias manually
|
923 |
+
if _weight is not None:
|
924 |
+
self.weight = nn.Parameter(_weight)
|
925 |
+
if _bias is not None:
|
926 |
+
self.bias = nn.Parameter(_bias)
|
927 |
+
|
928 |
+
self.in_features = in_features
|
929 |
+
self.original_out_features = original_out_features
|
930 |
+
self.max_original_id = max_original_id
|
931 |
+
|
932 |
+
# initialize additional linear
|
933 |
+
self.additional_out_features = additional_out_features
|
934 |
+
self.has_bias = bias
|
935 |
+
if additional_out_features > 0:
|
936 |
+
self.additional_fc = nn.Linear(
|
937 |
+
in_features=in_features,
|
938 |
+
out_features=additional_out_features,
|
939 |
+
bias=self.has_bias,
|
940 |
+
device=device,
|
941 |
+
dtype=dtype,
|
942 |
+
)
|
943 |
+
self.set_requires_grad(
|
944 |
+
require_regular_grad=not partially_freeze, require_additional_grad=True
|
945 |
+
)
|
946 |
+
|
947 |
+
def set_requires_grad(self, require_regular_grad, require_additional_grad):
|
948 |
+
"""
|
949 |
+
Helper function to separately set the requires_grad flag for the regular weight and the additional weight.
|
950 |
+
"""
|
951 |
+
self.weight.requires_grad_(require_regular_grad)
|
952 |
+
if self.has_bias:
|
953 |
+
self.bias.requires_grad_(require_regular_grad)
|
954 |
+
self.additional_fc.requires_grad_(require_additional_grad)
|
955 |
+
|
956 |
+
def forward(self, input: torch.Tensor) -> torch.Tensor:
|
957 |
+
output = F.linear(input, self.weight, self.bias)
|
958 |
+
output = output[..., : self.max_original_id + 1]
|
959 |
+
|
960 |
+
if self.additional_out_features > 0:
|
961 |
+
additional_features = F.linear(
|
962 |
+
input, self.additional_fc.weight, self.additional_fc.bias
|
963 |
+
)
|
964 |
+
output = torch.cat((output, additional_features), -1)
|
965 |
+
return output
|
966 |
+
|
967 |
+
def extra_repr(self) -> str:
|
968 |
+
"""Overwriting `nn.Linear.extra_repr` to include new parameters."""
|
969 |
+
return "in_features={}, out_features={}, additional_out_features={}, bias={}, partially_freeze={}".format(
|
970 |
+
self.in_features,
|
971 |
+
self.max_original_id + 1,
|
972 |
+
self.additional_out_features,
|
973 |
+
self.bias is not None,
|
974 |
+
(not self.weight.requires_grad or not self.bias.requires_grad),
|
975 |
+
)
|
976 |
+
|
977 |
+
|
978 |
+
class VLM(nn.Module):
|
979 |
+
"""
|
980 |
+
Generic vision-language model (VLM) class.
|
981 |
+
A VLM consists of four components:
|
982 |
+
1. A vision encoder that extracts features from pixels, e.g. CLIP
|
983 |
+
input: (B, T_img, F, C, H, W)
|
984 |
+
output: (B, T_img, F, v, d)
|
985 |
+
2. A vision tokenizer that converts these features to visual token-like embeddings, e.g. Perceiver, or a linear projection head
|
986 |
+
input: (B, T_img, F, v, d)
|
987 |
+
output: (B, T_img, n, d)
|
988 |
+
3. A fusion method that allows the language model to attend to these tokens, e.g. cross-attention, or placing the tokens directly in the language model's input sequence
|
989 |
+
4. A language model
|
990 |
+
"""
|
991 |
+
|
992 |
+
def __init__(
|
993 |
+
self,
|
994 |
+
vision_encoder: nn.Module,
|
995 |
+
vision_tokenizer: nn.Module,
|
996 |
+
lang_model: nn.Module,
|
997 |
+
initial_tokenizer_len: int,
|
998 |
+
pad_token_id: int,
|
999 |
+
gradient_checkpointing: bool = False,
|
1000 |
+
):
|
1001 |
+
"""
|
1002 |
+
Args:
|
1003 |
+
vision_encoder (nn.Module): e.g. CLIP
|
1004 |
+
vision_tokenizer (nn.Module): e.g. PerceiverResampler
|
1005 |
+
lang_model (nn.Module): e.g. MPT
|
1006 |
+
initial_tokenizer_len (int): size of the original tokenizer vocab
|
1007 |
+
pad_token_id (int): id of the pad token
|
1008 |
+
gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
|
1009 |
+
"""
|
1010 |
+
super().__init__()
|
1011 |
+
|
1012 |
+
# save dimension information
|
1013 |
+
self.lang_embedding_dim = lang_model.get_input_embeddings().weight.shape[1]
|
1014 |
+
if hasattr(lang_model.config, "d_model"):
|
1015 |
+
self.lang_hidden_dim = lang_model.config.d_model # mpt uses d_model
|
1016 |
+
else:
|
1017 |
+
self.lang_hidden_dim = lang_model.config.hidden_size
|
1018 |
+
self.vis_embedding_dim = vision_tokenizer.dim_media
|
1019 |
+
self.num_tokens_per_vis = vision_tokenizer.num_tokens_per_media
|
1020 |
+
|
1021 |
+
# core components
|
1022 |
+
self.vision_encoder = vision_encoder
|
1023 |
+
self.vision_tokenizer = vision_tokenizer
|
1024 |
+
self.lang_model = lang_model
|
1025 |
+
|
1026 |
+
# lm embeddings
|
1027 |
+
self.pad_token_id = pad_token_id
|
1028 |
+
self.initial_tokenizer_len = initial_tokenizer_len
|
1029 |
+
input_embeds = DecoupledEmbedding(
|
1030 |
+
max_original_id=initial_tokenizer_len - 1,
|
1031 |
+
num_additional_embeddings=len(self.special_tokens),
|
1032 |
+
_weight=self.lang_model.get_input_embeddings().weight,
|
1033 |
+
pad_token_id=self.pad_token_id,
|
1034 |
+
)
|
1035 |
+
if hasattr(input_embeds, "additional_embedding"):
|
1036 |
+
input_embeds.additional_embedding.weight.data.normal_(
|
1037 |
+
mean=0.0,
|
1038 |
+
std=(
|
1039 |
+
self.lang_model.config.initializer_range
|
1040 |
+
if hasattr(self.lang_model.config, "initializer_range")
|
1041 |
+
else 0.02
|
1042 |
+
),
|
1043 |
+
)
|
1044 |
+
self.lang_model.set_input_embeddings(input_embeds)
|
1045 |
+
|
1046 |
+
out_embeds = DecoupledLinear(
|
1047 |
+
max_original_id=initial_tokenizer_len - 1,
|
1048 |
+
additional_out_features=len(self.special_tokens),
|
1049 |
+
_weight=self.lang_model.get_output_embeddings().weight,
|
1050 |
+
_bias=(
|
1051 |
+
self.lang_model.get_output_embeddings().bias
|
1052 |
+
if hasattr(self.lang_model.get_output_embeddings(), "bias")
|
1053 |
+
else None
|
1054 |
+
),
|
1055 |
+
)
|
1056 |
+
if hasattr(out_embeds, "additional_fc"):
|
1057 |
+
out_embeds.additional_fc.weight.data.normal_(
|
1058 |
+
mean=0.0,
|
1059 |
+
std=(
|
1060 |
+
self.lang_model.config.initializer_range
|
1061 |
+
if hasattr(self.lang_model.config, "initializer_range")
|
1062 |
+
else 0.02
|
1063 |
+
),
|
1064 |
+
)
|
1065 |
+
self.lang_model.set_output_embeddings(out_embeds)
|
1066 |
+
|
1067 |
+
# gradient checkpointing
|
1068 |
+
self.vision_tokenizer._use_gradient_checkpointing = gradient_checkpointing
|
1069 |
+
|
1070 |
+
def forward(
|
1071 |
+
self,
|
1072 |
+
vision_x: Optional[torch.Tensor],
|
1073 |
+
lang_x: torch.Tensor,
|
1074 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1075 |
+
labels: Optional[torch.Tensor] = None,
|
1076 |
+
past_key_values: Optional[
|
1077 |
+
List[Union[torch.Tensor, Tuple[torch.Tensor]]]
|
1078 |
+
] = None,
|
1079 |
+
past_media_locations: Optional[torch.Tensor] = None,
|
1080 |
+
past_vision_tokens: Optional[torch.Tensor] = None,
|
1081 |
+
use_cache: Optional[bool] = False,
|
1082 |
+
**kwargs,
|
1083 |
+
):
|
1084 |
+
"""
|
1085 |
+
Args:
|
1086 |
+
vision_x: Vision input
|
1087 |
+
shape (B, T_img, F, C, H, W) with F=1
|
1088 |
+
only F = 1 is supported (single-frame videos)
|
1089 |
+
if T_img > the number of media tokens in the corresponding input_ids (lang_x),
|
1090 |
+
only the first number of media tokens in lang_x are used
|
1091 |
+
lang_x: Language input ids, with media tokens denoting where
|
1092 |
+
visual media should be inserted.
|
1093 |
+
shape (B, T_txt)
|
1094 |
+
attention_mask: Attention mask. Defaults to None.
|
1095 |
+
labels: Labels. Defaults to None.
|
1096 |
+
shape (B, T_txt)
|
1097 |
+
past_key_values (Tuple[torch.Tensor]], optional): Past key value pairs for each of the T_txt previous tokens in the language model. Defaults to None.
|
1098 |
+
list of length = number of decoder layers in the LM
|
1099 |
+
exact implementation depends on LM, see Hugging Face docs
|
1100 |
+
past_media_locations (torch.Tensor, optional): boolean mask denoting which of the previous T_txt tokens were media tokens. Defaults to None.
|
1101 |
+
shape (B, T_txt)
|
1102 |
+
past_vision_tokens (torch.Tensor, optional): Previous vision tokens. Defaults to None.
|
1103 |
+
use_cache (Optional[bool], optional): Whether to use cache. Defaults to False.
|
1104 |
+
If True, includes key_values, media_locations, and vision_tokens in the output.
|
1105 |
+
"""
|
1106 |
+
assert not (past_vision_tokens is None) ^ (
|
1107 |
+
past_media_locations is None
|
1108 |
+
), "past_vision_tokens and past_media_locations must both be None or both be not None"
|
1109 |
+
|
1110 |
+
# convert pixels to vision tokens
|
1111 |
+
if vision_x is not None:
|
1112 |
+
vision_features = self._encode_vision_x(vision_x=vision_x)
|
1113 |
+
vision_tokens = self.vision_tokenizer(vision_features)
|
1114 |
+
else:
|
1115 |
+
vision_tokens = None
|
1116 |
+
|
1117 |
+
# fuse the vision and language tokens
|
1118 |
+
new_inputs = self._prepare_inputs_for_forward(
|
1119 |
+
vision_tokens=vision_tokens,
|
1120 |
+
lang_x=lang_x,
|
1121 |
+
attention_mask=attention_mask,
|
1122 |
+
labels=labels,
|
1123 |
+
past_key_values=past_key_values,
|
1124 |
+
past_media_locations=past_media_locations,
|
1125 |
+
padding_side="right",
|
1126 |
+
past_vision_tokens=past_vision_tokens,
|
1127 |
+
)
|
1128 |
+
output = self.lang_model(
|
1129 |
+
**new_inputs,
|
1130 |
+
use_cache=use_cache,
|
1131 |
+
past_key_values=past_key_values,
|
1132 |
+
**kwargs,
|
1133 |
+
)
|
1134 |
+
|
1135 |
+
# postprocessing may be needed, e.g. to remove extra tokens from logits that were inserted into the language stream
|
1136 |
+
# or to add the past_vision_tokens and past_media_locations to the output
|
1137 |
+
output = self._postprocess_outputs_from_forward(
|
1138 |
+
output=output,
|
1139 |
+
lang_x=lang_x,
|
1140 |
+
vision_tokens=vision_tokens,
|
1141 |
+
use_cache=use_cache,
|
1142 |
+
past_vision_tokens=past_vision_tokens,
|
1143 |
+
past_media_locations=past_media_locations,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
# postforward hooks
|
1147 |
+
self._post_forward_hook()
|
1148 |
+
return output
|
1149 |
+
|
1150 |
+
def _encode_vision_x_anyres(self, samples, device):
|
1151 |
+
assert self.anyres_grids is not None
|
1152 |
+
image_raw = samples[
|
1153 |
+
"image"
|
1154 |
+
] # list of patch list in of shape [1, N_patch, C, H, W]
|
1155 |
+
image_sizes = samples["image_size"]
|
1156 |
+
|
1157 |
+
# Image_raw can be a list of list of patches, when a `samples` has multiple images.
|
1158 |
+
if isinstance(image_raw[0], list):
|
1159 |
+
images = [x.squeeze(0) for sample_img in image_raw for x in sample_img]
|
1160 |
+
image_sizes = [s for sample_sizes in image_sizes for s in sample_sizes]
|
1161 |
+
else:
|
1162 |
+
# assert isinstance(image_raw[0], torch.Tensor), f"Unkown image type: {image_raw[0]}"
|
1163 |
+
# concate list of patches into one big patch for any res encoding.
|
1164 |
+
images = [x.squeeze(0) for x in image_raw] # [N_patch, C, H, W]
|
1165 |
+
image = torch.cat(images, dim=0) # [\sum{B}{N_patch_i}, C, H, W]
|
1166 |
+
image = image.to(device)
|
1167 |
+
|
1168 |
+
with torch.no_grad():
|
1169 |
+
if self.vision_encoder.__class__.__name__ == "TimmModel":
|
1170 |
+
image_embeds = self.vision_encoder.trunk.forward_features(image)
|
1171 |
+
elif self.vision_encoder.__class__.__name__ in [
|
1172 |
+
"CLIPVisionModel",
|
1173 |
+
"SiglipVisionTransformer",
|
1174 |
+
]:
|
1175 |
+
image_embeds = self.vision_encoder(image).last_hidden_state
|
1176 |
+
else:
|
1177 |
+
image_embeds = self.vision_encoder(image)[1] # OpenCLIP returns tuples
|
1178 |
+
|
1179 |
+
if isinstance(self.vision_encoder, CLIPVisionModel) or isinstance(
|
1180 |
+
self.vision_encoder, SiglipVisionTransformer
|
1181 |
+
):
|
1182 |
+
base_img_size = self.vision_encoder.config.image_size
|
1183 |
+
else:
|
1184 |
+
base_img_size = self.vision_encoder.image_size[0]
|
1185 |
+
|
1186 |
+
if self.vision_encoder.__class__.__name__ == "TimmModel":
|
1187 |
+
grid_size = self.vision_encoder.trunk.patch_embed.grid_size
|
1188 |
+
elif self.vision_encoder.__class__.__name__ in [
|
1189 |
+
"CLIPVisionModel",
|
1190 |
+
"SiglipVisionTransformer",
|
1191 |
+
]:
|
1192 |
+
grid_size_base = (
|
1193 |
+
self.vision_encoder.config.image_size
|
1194 |
+
// self.vision_encoder.config.patch_size
|
1195 |
+
)
|
1196 |
+
grid_size = (grid_size_base, grid_size_base)
|
1197 |
+
else:
|
1198 |
+
grid_size = self.vision_encoder.grid_size
|
1199 |
+
height, width = grid_size
|
1200 |
+
|
1201 |
+
if not image_embeds.shape[1] == height * width:
|
1202 |
+
assert (
|
1203 |
+
image_embeds.shape[1] == height * width + 1
|
1204 |
+
) # For vision encoders that has [CLS] token.
|
1205 |
+
image_embeds = image_embeds[:, 1:, :] # Drop the cls token for each patch.
|
1206 |
+
n_vis_token_per_patch = image_embeds.shape[1]
|
1207 |
+
|
1208 |
+
# Split encoded patches and merge patch features
|
1209 |
+
# 1. Get the raw sizes from samples, and split the image embeds [\sum_{B}(N_patch_i), N_tok(16*16), C]
|
1210 |
+
split_sizes = [image.shape[0] for image in images]
|
1211 |
+
image_embeds = torch.split(image_embeds, split_sizes, dim=0)
|
1212 |
+
# 2. For each image (consist of a list of patches), merge the patches spatially (of shape [C, n_patch_height, n_patch_width])
|
1213 |
+
new_image_embeds = []
|
1214 |
+
patch_attn_masks = []
|
1215 |
+
max_n_img_token = -1
|
1216 |
+
for idx, patch_embeds in enumerate(image_embeds):
|
1217 |
+
if patch_embeds.shape[0] > 1:
|
1218 |
+
# 3. Flatten the patch features and get [C, n_patch_height * (n_patch_width+1)]
|
1219 |
+
base_patch_embeds = patch_embeds[
|
1220 |
+
0
|
1221 |
+
] # TODO: prepend the CLS token for th base patch embeds (of the resized entire image).
|
1222 |
+
patch_embeds = patch_embeds[1:]
|
1223 |
+
|
1224 |
+
assert height * width == base_patch_embeds.shape[0]
|
1225 |
+
|
1226 |
+
num_patch_width, num_patch_height = get_anyres_image_grid_shape(
|
1227 |
+
image_sizes[idx], self.anyres_grids, base_img_size
|
1228 |
+
) # Hardcoded grid_pinpoints.
|
1229 |
+
patch_embeds = patch_embeds.view(
|
1230 |
+
num_patch_height, num_patch_width, height, width, -1
|
1231 |
+
)
|
1232 |
+
|
1233 |
+
patch_embeds = patch_embeds.permute(4, 0, 2, 1, 3).contiguous()
|
1234 |
+
patch_embeds = patch_embeds.flatten(1, 2).flatten(2, 3)
|
1235 |
+
patch_embeds, patch_attn_mask = unpad_image(
|
1236 |
+
patch_embeds, image_sizes[idx], self.anyres_patch_sampling
|
1237 |
+
)
|
1238 |
+
if hasattr(self, "image_newline"):
|
1239 |
+
patch_embeds = torch.cat(
|
1240 |
+
(
|
1241 |
+
patch_embeds,
|
1242 |
+
self.image_newline[:, None, None].expand(
|
1243 |
+
*patch_embeds.shape[:-1], 1
|
1244 |
+
),
|
1245 |
+
),
|
1246 |
+
dim=-1,
|
1247 |
+
)
|
1248 |
+
if self.anyres_patch_sampling:
|
1249 |
+
patch_embeds = patch_embeds.view(
|
1250 |
+
-1, num_patch_height, num_patch_width, height * width
|
1251 |
+
)
|
1252 |
+
patch_embeds = patch_embeds.flatten(1, 2).permute(1, 2, 0)
|
1253 |
+
assert patch_attn_mask is not None
|
1254 |
+
patch_attn_mask = patch_attn_mask.view(
|
1255 |
+
num_patch_height, num_patch_width, height * width
|
1256 |
+
)
|
1257 |
+
patch_attn_mask = patch_attn_mask.flatten(0, 1)
|
1258 |
+
patch_embeds = torch.cat(
|
1259 |
+
(base_patch_embeds.unsqueeze(0), patch_embeds), dim=0
|
1260 |
+
)
|
1261 |
+
patch_attn_mask = torch.cat(
|
1262 |
+
(
|
1263 |
+
torch.ones(
|
1264 |
+
n_vis_token_per_patch, device=patch_embeds.device
|
1265 |
+
).unsqueeze(0),
|
1266 |
+
patch_attn_mask,
|
1267 |
+
),
|
1268 |
+
dim=0,
|
1269 |
+
)
|
1270 |
+
else:
|
1271 |
+
patch_embeds = patch_embeds.flatten(1, 2).transpose(0, 1)
|
1272 |
+
patch_embeds = torch.cat((base_patch_embeds, patch_embeds), dim=0)
|
1273 |
+
else:
|
1274 |
+
patch_embeds = (
|
1275 |
+
patch_embeds[0].unsqueeze(0)
|
1276 |
+
if self.anyres_patch_sampling
|
1277 |
+
else patch_embeds[0]
|
1278 |
+
)
|
1279 |
+
patch_attn_mask = (
|
1280 |
+
torch.ones(
|
1281 |
+
n_vis_token_per_patch, device=patch_embeds.device
|
1282 |
+
).unsqueeze(0)
|
1283 |
+
if self.anyres_patch_sampling
|
1284 |
+
else None
|
1285 |
+
)
|
1286 |
+
if hasattr(self, "image_newline"):
|
1287 |
+
patch_embeds = torch.cat(
|
1288 |
+
(patch_embeds, self.image_newline[None]), dim=0
|
1289 |
+
)
|
1290 |
+
if not self.anyres_patch_sampling:
|
1291 |
+
max_n_img_token = max(patch_embeds.shape[0], max_n_img_token)
|
1292 |
+
|
1293 |
+
new_image_embeds.append(patch_embeds)
|
1294 |
+
patch_attn_masks.append(patch_attn_mask)
|
1295 |
+
|
1296 |
+
if self.anyres_patch_sampling:
|
1297 |
+
# Return individual patches for independent token downsampling.
|
1298 |
+
return new_image_embeds, patch_attn_masks
|
1299 |
+
|
1300 |
+
# 4. Pad and concat the list of image_embeds [N_tok_i, C] together into a batch. Also modify the query attention mask.
|
1301 |
+
image_embeds = []
|
1302 |
+
image_atts = []
|
1303 |
+
for image_embed in new_image_embeds:
|
1304 |
+
n_img_token = image_embed.shape[0]
|
1305 |
+
img_attn = torch.ones(
|
1306 |
+
(max_n_img_token), dtype=torch.long, device=image_embed.device
|
1307 |
+
)
|
1308 |
+
if n_img_token < max_n_img_token:
|
1309 |
+
padded_embed = torch.zeros(
|
1310 |
+
(max_n_img_token, image_embed.shape[-1]),
|
1311 |
+
dtype=image_embed.dtype,
|
1312 |
+
device=image_embed.device,
|
1313 |
+
)
|
1314 |
+
padded_embed[:n_img_token, :] = image_embed
|
1315 |
+
img_attn[n_img_token:] = 0 # Mask out the padded entries.
|
1316 |
+
else:
|
1317 |
+
padded_embed = image_embed
|
1318 |
+
image_embeds.append(padded_embed)
|
1319 |
+
image_atts.append(img_attn)
|
1320 |
+
image_embeds = torch.stack(
|
1321 |
+
image_embeds, dim=0
|
1322 |
+
) # Shape [B, N_tok_longest, C_dim]
|
1323 |
+
image_atts = torch.stack(image_atts, dim=0) # Shape [B, N_tok_longest, C_dim]
|
1324 |
+
# TODO: reshape image_embeds and image_atts to "b T F v d"
|
1325 |
+
image_embeds = image_embeds[:, None, None, :, :]
|
1326 |
+
# image_atts = image_atts[:, None, None, :, :]
|
1327 |
+
|
1328 |
+
return image_embeds, image_atts
|
1329 |
+
|
1330 |
+
def _encode_vision_x(self, vision_x: torch.Tensor):
|
1331 |
+
"""
|
1332 |
+
Compute media tokens from vision input by passing it through vision encoder and conditioning language model.
|
1333 |
+
Args:
|
1334 |
+
vision_x: Vision input
|
1335 |
+
shape (B, T_img, F, C, H, W)
|
1336 |
+
Images in the same chunk are collated along T_img, and frames are collated along F
|
1337 |
+
Currently only F=1 is supported (single-frame videos)
|
1338 |
+
|
1339 |
+
rearrange code based on https://github.com/dhansmair/flamingo-mini
|
1340 |
+
"""
|
1341 |
+
assert vision_x.ndim == 6, "vision_x should be of shape (b, T_img, F, C, H, W)"
|
1342 |
+
b, T, F = vision_x.shape[:3]
|
1343 |
+
|
1344 |
+
vision_x = rearrange(vision_x, "b T F c h w -> (b T F) c h w")
|
1345 |
+
with torch.no_grad():
|
1346 |
+
if self.vision_encoder.__class__.__name__ == "TimmModel":
|
1347 |
+
vision_x = self.vision_encoder.trunk.forward_features(vision_x)
|
1348 |
+
elif self.vision_encoder.__class__.__name__ in [
|
1349 |
+
"CLIPVisionModel",
|
1350 |
+
"SiglipVisionTransformer",
|
1351 |
+
]:
|
1352 |
+
vision_x = self.vision_encoder(vision_x).last_hidden_state
|
1353 |
+
else:
|
1354 |
+
vision_x = self.vision_encoder(vision_x)[1] # OpenCLIP returns tuples
|
1355 |
+
vision_x = rearrange(vision_x, "(b T F) v d -> b T F v d", b=b, T=T, F=F)
|
1356 |
+
return vision_x
|
1357 |
+
|
1358 |
+
def _concat_vision_cache(
|
1359 |
+
self, lang_x, vision_tokens, past_vision_tokens, past_media_locations, use_cache
|
1360 |
+
):
|
1361 |
+
"""
|
1362 |
+
Helper function to include the past vision tokens and past media locations in the output.
|
1363 |
+
"""
|
1364 |
+
if use_cache:
|
1365 |
+
if past_media_locations is not None and past_vision_tokens is not None:
|
1366 |
+
if vision_tokens is not None:
|
1367 |
+
updated_vision_tokens = torch.cat(
|
1368 |
+
[
|
1369 |
+
past_vision_tokens,
|
1370 |
+
vision_tokens,
|
1371 |
+
],
|
1372 |
+
dim=1,
|
1373 |
+
)
|
1374 |
+
else:
|
1375 |
+
updated_vision_tokens = past_vision_tokens
|
1376 |
+
updated_media_locations = torch.cat(
|
1377 |
+
[
|
1378 |
+
past_media_locations,
|
1379 |
+
lang_x == self.media_token_id,
|
1380 |
+
],
|
1381 |
+
dim=1,
|
1382 |
+
)
|
1383 |
+
else:
|
1384 |
+
updated_vision_tokens = vision_tokens
|
1385 |
+
updated_media_locations = lang_x == self.media_token_id
|
1386 |
+
|
1387 |
+
else:
|
1388 |
+
updated_vision_tokens = None
|
1389 |
+
updated_media_locations = None
|
1390 |
+
|
1391 |
+
return updated_vision_tokens, updated_media_locations
|
1392 |
+
|
1393 |
+
def generate(
|
1394 |
+
self,
|
1395 |
+
vision_x: torch.Tensor,
|
1396 |
+
lang_x: torch.Tensor,
|
1397 |
+
attention_mask: torch.Tensor = None,
|
1398 |
+
past_key_values: Optional[
|
1399 |
+
List[Union[torch.Tensor, Tuple[torch.Tensor]]]
|
1400 |
+
] = None,
|
1401 |
+
past_media_locations: Optional[torch.Tensor] = None,
|
1402 |
+
past_vision_tokens: Optional[torch.Tensor] = None,
|
1403 |
+
**kwargs,
|
1404 |
+
):
|
1405 |
+
"""
|
1406 |
+
Generate text conditioned on vision and language inputs.
|
1407 |
+
Args:
|
1408 |
+
vision_x (torch.Tensor): Vision input
|
1409 |
+
shape (B, T_img, F, C, H, W)
|
1410 |
+
see documentation for forward
|
1411 |
+
lang_x (torch.Tensor): Language input
|
1412 |
+
shape (B, T_txt)
|
1413 |
+
attention_mask (torch.Tensor, optional): Attention mask. Defaults to None.
|
1414 |
+
**kwargs: see generate documentation in Hugging Face CausalLM models.
|
1415 |
+
Returns:
|
1416 |
+
torch.Tensor: lang_x with generated tokens appended to it
|
1417 |
+
"""
|
1418 |
+
num_beams = kwargs.pop("num_beams", 1)
|
1419 |
+
|
1420 |
+
# convert pixels to vision tokens
|
1421 |
+
if vision_x is not None:
|
1422 |
+
vision_features = self._encode_vision_x(vision_x=vision_x)
|
1423 |
+
vision_tokens = self.vision_tokenizer(vision_features)
|
1424 |
+
else:
|
1425 |
+
vision_tokens = None
|
1426 |
+
|
1427 |
+
# fuse the vision and language tokens
|
1428 |
+
# for xattn, vision_x and media_location are repeat_interleaved s.t.
|
1429 |
+
# the total batch size is B * num_beams
|
1430 |
+
new_inputs = self._prepare_inputs_for_forward(
|
1431 |
+
vision_tokens=vision_tokens,
|
1432 |
+
lang_x=lang_x,
|
1433 |
+
attention_mask=attention_mask,
|
1434 |
+
past_key_values=past_key_values,
|
1435 |
+
past_media_locations=past_media_locations,
|
1436 |
+
past_vision_tokens=past_vision_tokens,
|
1437 |
+
padding_side="left",
|
1438 |
+
num_beams=num_beams,
|
1439 |
+
)
|
1440 |
+
output = self.lang_model.generate(
|
1441 |
+
**new_inputs,
|
1442 |
+
past_key_values=past_key_values,
|
1443 |
+
num_beams=num_beams,
|
1444 |
+
use_cache=True,
|
1445 |
+
**kwargs,
|
1446 |
+
)
|
1447 |
+
self._post_forward_hook()
|
1448 |
+
return output
|
1449 |
+
|
1450 |
+
@property
|
1451 |
+
def num_trainable_params(self):
|
1452 |
+
"""Print the number of trainable parameters"""
|
1453 |
+
return num_params(self, filter_to_trainable=True)
|
1454 |
+
|
1455 |
+
def set_trainable(self):
|
1456 |
+
"""
|
1457 |
+
Freeze appropriate parameters in the model.
|
1458 |
+
"""
|
1459 |
+
raise NotImplementedError
|
1460 |
+
|
1461 |
+
def group_params_by_weight_decay(self):
|
1462 |
+
"""
|
1463 |
+
Return a tuple of (params to optimize w/ weight decay, params to optimize w/o weight decay)
|
1464 |
+
"""
|
1465 |
+
params_with_wd, params_without_wd = [], []
|
1466 |
+
for n, p in self.named_parameters():
|
1467 |
+
if p.requires_grad:
|
1468 |
+
if self._should_apply_weight_decay(n):
|
1469 |
+
params_with_wd.append(p)
|
1470 |
+
else:
|
1471 |
+
params_without_wd.append(p)
|
1472 |
+
return params_with_wd, params_without_wd
|
1473 |
+
|
1474 |
+
def _should_apply_weight_decay(self, parameter_name):
|
1475 |
+
"""
|
1476 |
+
Return whether weight decay should be applied to a parameter.
|
1477 |
+
"""
|
1478 |
+
raise NotImplementedError
|
1479 |
+
|
1480 |
+
@property
|
1481 |
+
def special_tokens(self):
|
1482 |
+
"""
|
1483 |
+
Returns a dict mapping from the attribute name of a special token to its string format,
|
1484 |
+
e.g. "media_token": "<image>"
|
1485 |
+
"""
|
1486 |
+
assert (
|
1487 |
+
"media_token" in self._special_tokens
|
1488 |
+
), "VLMs need to request that the tokenizer add a media_token and call set_special_token_ids to set self.media_token_id"
|
1489 |
+
return self._special_tokens
|
1490 |
+
|
1491 |
+
@property
|
1492 |
+
def special_token_ids(self):
|
1493 |
+
"""
|
1494 |
+
Returns a list of the special token ids
|
1495 |
+
"""
|
1496 |
+
return [getattr(self, f"{att_name}_id") for att_name in self.special_tokens]
|
1497 |
+
|
1498 |
+
def set_special_token_ids(self, string_to_ids):
|
1499 |
+
"""
|
1500 |
+
Args:
|
1501 |
+
string_to_ids (dict): mapping from token string to id
|
1502 |
+
"""
|
1503 |
+
assert set(self.special_tokens.values()).issubset(set(string_to_ids.keys()))
|
1504 |
+
for att_name, token_str in self.special_tokens.items():
|
1505 |
+
token_id = string_to_ids[token_str]
|
1506 |
+
setattr(self, f"{att_name}_id", token_id)
|
1507 |
+
setattr(self.lang_model, f"{att_name}_id", token_id)
|
1508 |
+
|
1509 |
+
def init_gradient_checkpointing(self):
|
1510 |
+
from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
|
1511 |
+
checkpoint_wrapper,
|
1512 |
+
CheckpointWrapper,
|
1513 |
+
CheckpointImpl,
|
1514 |
+
apply_activation_checkpointing,
|
1515 |
+
)
|
1516 |
+
from functools import partial
|
1517 |
+
|
1518 |
+
non_reentrant_wrapper = partial(
|
1519 |
+
checkpoint_wrapper,
|
1520 |
+
checkpoint_impl=CheckpointImpl.NO_REENTRANT,
|
1521 |
+
)
|
1522 |
+
apply_activation_checkpointing(
|
1523 |
+
self,
|
1524 |
+
checkpoint_wrapper_fn=non_reentrant_wrapper,
|
1525 |
+
check_fn=lambda m: getattr(m, "_use_gradient_checkpointing", False)
|
1526 |
+
and not isinstance(m, CheckpointWrapper),
|
1527 |
+
)
|
1528 |
+
|
1529 |
+
|
1530 |
+
@dataclass
|
1531 |
+
class VLMOutputWithPast(CausalLMOutputWithPast):
|
1532 |
+
"""
|
1533 |
+
VLMOutputWithPast is a wrapper around CausalLMOutputWithPast that adds the following attributes:
|
1534 |
+
past_media_locations: Optional[torch.Tensor] = None,
|
1535 |
+
past_vision_tokens: Optional[torch.Tensor] = None,
|
1536 |
+
"""
|
1537 |
+
|
1538 |
+
past_media_locations: Optional[torch.Tensor] = None
|
1539 |
+
past_vision_tokens: Optional[torch.Tensor] = None
|
1540 |
+
|
1541 |
+
|
1542 |
+
def exists(val):
|
1543 |
+
return val is not None
|
1544 |
+
|
1545 |
+
|
1546 |
+
def FeedForward(dim, mult=4):
|
1547 |
+
inner_dim = int(dim * mult)
|
1548 |
+
return nn.Sequential(
|
1549 |
+
nn.LayerNorm(dim),
|
1550 |
+
nn.Linear(dim, inner_dim, bias=False),
|
1551 |
+
nn.GELU(),
|
1552 |
+
nn.Linear(inner_dim, dim, bias=False),
|
1553 |
+
)
|
1554 |
+
|
1555 |
+
|
1556 |
+
class VLMWithLanguageStream(VLM):
|
1557 |
+
"""
|
1558 |
+
VLM that fuses modalities by inserting vision tokens directly into the language stream.
|
1559 |
+
"""
|
1560 |
+
|
1561 |
+
def __init__(
|
1562 |
+
self,
|
1563 |
+
vision_encoder: nn.Module,
|
1564 |
+
vision_tokenizer: nn.Module,
|
1565 |
+
lang_model: nn.Module,
|
1566 |
+
initial_tokenizer_len: int,
|
1567 |
+
pad_token_id: int,
|
1568 |
+
decoder_layers_attr_name: str = None,
|
1569 |
+
gradient_checkpointing: bool = False,
|
1570 |
+
):
|
1571 |
+
super().__init__(
|
1572 |
+
vision_encoder=vision_encoder,
|
1573 |
+
vision_tokenizer=vision_tokenizer,
|
1574 |
+
lang_model=lang_model,
|
1575 |
+
initial_tokenizer_len=initial_tokenizer_len,
|
1576 |
+
pad_token_id=pad_token_id,
|
1577 |
+
gradient_checkpointing=gradient_checkpointing,
|
1578 |
+
)
|
1579 |
+
self.decoder_layers_attr_name = decoder_layers_attr_name
|
1580 |
+
if decoder_layers_attr_name is not None:
|
1581 |
+
for block in getattr_recursive(
|
1582 |
+
self.lang_model, self.decoder_layers_attr_name
|
1583 |
+
):
|
1584 |
+
block._use_gradient_checkpointing = gradient_checkpointing
|
1585 |
+
|
1586 |
+
def _prepare_inputs_for_forward(
|
1587 |
+
self,
|
1588 |
+
vision_tokens: torch.Tensor,
|
1589 |
+
lang_x: torch.Tensor,
|
1590 |
+
attention_mask: torch.Tensor,
|
1591 |
+
labels: torch.Tensor = None,
|
1592 |
+
past_key_values=None,
|
1593 |
+
vision_attention_mask: Optional[torch.Tensor] = None,
|
1594 |
+
past_media_locations: torch.Tensor = None,
|
1595 |
+
past_vision_tokens: torch.Tensor = None,
|
1596 |
+
padding_side: str = "left",
|
1597 |
+
num_beams: int = 1,
|
1598 |
+
):
|
1599 |
+
"""
|
1600 |
+
Insert the vision tokens directly into the language stream/
|
1601 |
+
This requires us to modify the input_ids, attention_mask, and labels.
|
1602 |
+
"""
|
1603 |
+
if past_key_values is not None:
|
1604 |
+
past_len = past_key_values[0][0].shape[2]
|
1605 |
+
assert attention_mask.shape[1] == past_len + lang_x.shape[1], (
|
1606 |
+
"Attention_mask must be as long as the entire past len (including image tokens) and current input IDs. "
|
1607 |
+
+ "Check that you've expanded the attention mask to account for past image tokens."
|
1608 |
+
)
|
1609 |
+
|
1610 |
+
if vision_tokens is None:
|
1611 |
+
return {
|
1612 |
+
"input_ids": lang_x,
|
1613 |
+
"attention_mask": attention_mask,
|
1614 |
+
"labels": labels,
|
1615 |
+
}
|
1616 |
+
|
1617 |
+
# get the language embeddings
|
1618 |
+
lang_embeds = self.lang_model.get_input_embeddings()(lang_x)
|
1619 |
+
|
1620 |
+
# build up the multimodal embeddings
|
1621 |
+
B = lang_x.shape[0]
|
1622 |
+
has_labels = labels is not None
|
1623 |
+
multimodal_embeds = []
|
1624 |
+
multimodal_attention_mask = []
|
1625 |
+
multimodal_labels = [] if has_labels else None
|
1626 |
+
for i in range(B):
|
1627 |
+
# get index of <image> tokens in lang_x[i]
|
1628 |
+
image_token_idxs = torch.where(lang_x[i] == self.media_token_id)[0]
|
1629 |
+
|
1630 |
+
if len(image_token_idxs) == 0:
|
1631 |
+
multimodal_embeds.append(lang_embeds[i].clone())
|
1632 |
+
multimodal_attention_mask.append(attention_mask[i].clone())
|
1633 |
+
if has_labels:
|
1634 |
+
multimodal_labels.append(labels[i].clone())
|
1635 |
+
continue
|
1636 |
+
|
1637 |
+
# loop through the image_token_idxs and insert the vision tokens
|
1638 |
+
new_embed = lang_embeds[i].clone()
|
1639 |
+
new_attention_mask = (
|
1640 |
+
attention_mask[i].clone() if attention_mask is not None else None
|
1641 |
+
)
|
1642 |
+
if has_labels:
|
1643 |
+
new_label = labels[i].clone()
|
1644 |
+
|
1645 |
+
for img_num, img_idx in enumerate(image_token_idxs):
|
1646 |
+
# Get vision token attention mask for padded llava-style any resolution image tokens.
|
1647 |
+
if self.image_aspect_ratio == "anyres":
|
1648 |
+
num_vis_tokens = vision_tokens[i][img_num].shape[0]
|
1649 |
+
if vision_attention_mask is not None:
|
1650 |
+
vis_attention_mask = vision_attention_mask[i]
|
1651 |
+
else:
|
1652 |
+
vis_attention_mask = torch.ones(
|
1653 |
+
num_vis_tokens, dtype=torch.long
|
1654 |
+
).to(attention_mask.device)
|
1655 |
+
else:
|
1656 |
+
assert (
|
1657 |
+
vision_tokens[i][img_num].shape[0] == self.num_tokens_per_vis
|
1658 |
+
), f"vision token number mismatch: image embedding ({vision_tokens[i][img_num].shape[0]}) \
|
1659 |
+
vs. model.num_tokens_per_vis ({self.num_tokens_per_vis})"
|
1660 |
+
# By default, vision tokens are not padded.
|
1661 |
+
num_vis_tokens = self.num_tokens_per_vis
|
1662 |
+
vis_attention_mask = torch.ones(
|
1663 |
+
num_vis_tokens, dtype=torch.long
|
1664 |
+
).to(attention_mask.device)
|
1665 |
+
|
1666 |
+
new_embed = torch.cat(
|
1667 |
+
(
|
1668 |
+
new_embed[:img_idx],
|
1669 |
+
vision_tokens[i][img_num],
|
1670 |
+
new_embed[img_idx + 1 :],
|
1671 |
+
),
|
1672 |
+
dim=0,
|
1673 |
+
)
|
1674 |
+
new_attention_mask = torch.cat(
|
1675 |
+
(
|
1676 |
+
new_attention_mask[:img_idx],
|
1677 |
+
vis_attention_mask,
|
1678 |
+
new_attention_mask[img_idx + 1 :],
|
1679 |
+
),
|
1680 |
+
dim=0,
|
1681 |
+
)
|
1682 |
+
if has_labels:
|
1683 |
+
new_label = torch.cat(
|
1684 |
+
(
|
1685 |
+
new_label[:img_idx],
|
1686 |
+
torch.ones(num_vis_tokens, dtype=torch.long).to(
|
1687 |
+
labels.device
|
1688 |
+
)
|
1689 |
+
* -100,
|
1690 |
+
new_label[img_idx + 1 :],
|
1691 |
+
),
|
1692 |
+
dim=0,
|
1693 |
+
)
|
1694 |
+
multimodal_embeds.append(new_embed)
|
1695 |
+
multimodal_attention_mask.append(new_attention_mask)
|
1696 |
+
if has_labels:
|
1697 |
+
multimodal_labels.append(new_label)
|
1698 |
+
|
1699 |
+
# stack
|
1700 |
+
multimodal_embeds = stack_with_padding(
|
1701 |
+
multimodal_embeds,
|
1702 |
+
padding_value=self.pad_token_id,
|
1703 |
+
padding_side=padding_side,
|
1704 |
+
)
|
1705 |
+
multimodal_attention_mask = stack_with_padding(
|
1706 |
+
multimodal_attention_mask,
|
1707 |
+
padding_value=0,
|
1708 |
+
padding_side=padding_side,
|
1709 |
+
)
|
1710 |
+
if has_labels:
|
1711 |
+
multimodal_labels = stack_with_padding(
|
1712 |
+
multimodal_labels,
|
1713 |
+
padding_value=-100,
|
1714 |
+
padding_side=padding_side,
|
1715 |
+
)
|
1716 |
+
|
1717 |
+
return {
|
1718 |
+
"inputs_embeds": multimodal_embeds,
|
1719 |
+
"attention_mask": multimodal_attention_mask,
|
1720 |
+
"labels": multimodal_labels,
|
1721 |
+
}
|
1722 |
+
|
1723 |
+
def _postprocess_outputs_from_forward(
|
1724 |
+
self,
|
1725 |
+
output: CausalLMOutputWithPast,
|
1726 |
+
lang_x: torch.Tensor,
|
1727 |
+
vision_tokens: torch.Tensor,
|
1728 |
+
past_vision_tokens: torch.Tensor,
|
1729 |
+
past_media_locations: torch.Tensor,
|
1730 |
+
use_cache: bool = False,
|
1731 |
+
):
|
1732 |
+
# Include the past vision tokens and past media locations in the output
|
1733 |
+
updated_vision_tokens, updated_media_locations = self._concat_vision_cache(
|
1734 |
+
lang_x=lang_x,
|
1735 |
+
vision_tokens=vision_tokens,
|
1736 |
+
past_vision_tokens=past_vision_tokens,
|
1737 |
+
past_media_locations=past_media_locations,
|
1738 |
+
use_cache=use_cache,
|
1739 |
+
)
|
1740 |
+
|
1741 |
+
# return logits that are the same shape as the original input_ids
|
1742 |
+
logits = output.logits
|
1743 |
+
batch_logits = []
|
1744 |
+
B, T_txt = lang_x.shape
|
1745 |
+
for i in range(B):
|
1746 |
+
sequence_logits = []
|
1747 |
+
logits_j = 0
|
1748 |
+
for j in range(T_txt):
|
1749 |
+
if lang_x[i, j] != self.media_token_id:
|
1750 |
+
sequence_logits.append(logits[i, logits_j])
|
1751 |
+
logits_j += 1
|
1752 |
+
else:
|
1753 |
+
# append the logit for the first image token, then skip over the rest
|
1754 |
+
# note: the model actually learns to predict <im_patch>, not <image>
|
1755 |
+
sequence_logits.append(logits[i, logits_j])
|
1756 |
+
logits_j += self.num_tokens_per_vis
|
1757 |
+
sequence_logits = torch.stack(sequence_logits, dim=0) # (B, vocab_size)
|
1758 |
+
batch_logits.append(sequence_logits)
|
1759 |
+
|
1760 |
+
batch_logits = torch.stack(batch_logits, dim=0) # (B, T_txt, vocab_size)
|
1761 |
+
# The final logits shape should be the same as the original input_ids shape
|
1762 |
+
assert batch_logits.shape[:2] == (B, T_txt)
|
1763 |
+
|
1764 |
+
# assemble the output
|
1765 |
+
output = VLMOutputWithPast(
|
1766 |
+
loss=output.loss,
|
1767 |
+
logits=batch_logits,
|
1768 |
+
past_key_values=output.past_key_values,
|
1769 |
+
hidden_states=output.hidden_states,
|
1770 |
+
attentions=output.attentions,
|
1771 |
+
past_media_locations=updated_media_locations,
|
1772 |
+
past_vision_tokens=updated_vision_tokens,
|
1773 |
+
)
|
1774 |
+
|
1775 |
+
return output
|
1776 |
+
|
1777 |
+
def _post_forward_hook(self):
|
1778 |
+
pass
|
1779 |
+
|
1780 |
+
@property
|
1781 |
+
def num_params_per_module(self):
|
1782 |
+
"""Print the number of parameters per module in the model"""
|
1783 |
+
return "\n".join(
|
1784 |
+
[
|
1785 |
+
f"Vision encoder: {num_params(self.vision_encoder):,} parameters",
|
1786 |
+
f"Vision tokenizer: {num_params(self.vision_tokenizer):,} parameters",
|
1787 |
+
f"Language model: {num_params(self.lang_model):,} parameters",
|
1788 |
+
]
|
1789 |
+
)
|
1790 |
+
|
1791 |
+
@property
|
1792 |
+
def num_trainable_params_per_module(self):
|
1793 |
+
"""Print the number of trainable parameters per module in the model"""
|
1794 |
+
return "\n".join(
|
1795 |
+
[
|
1796 |
+
f"Vision encoder: {num_params(self.vision_encoder, filter_to_trainable=True):,} trainable parameters",
|
1797 |
+
f"Vision tokenizer: {num_params(self.vision_tokenizer, filter_to_trainable=True):,} trainable parameters",
|
1798 |
+
f"Language model: {num_params(self.lang_model, filter_to_trainable=True):,} trainable parameters",
|
1799 |
+
]
|
1800 |
+
)
|
1801 |
+
|
1802 |
+
|
1803 |
+
class XGenMMPerceiver(VLMWithLanguageStream):
|
1804 |
+
def __init__(
|
1805 |
+
self,
|
1806 |
+
vision_encoder: nn.Module,
|
1807 |
+
vision_tokenizer: nn.Module,
|
1808 |
+
lang_model: nn.Module,
|
1809 |
+
initial_tokenizer_len: int,
|
1810 |
+
pad_token_id: int,
|
1811 |
+
decoder_layers_attr_name: str = None,
|
1812 |
+
gradient_checkpointing: bool = False,
|
1813 |
+
image_aspect_ratio: str = "anyres",
|
1814 |
+
anyres_patch_sampling: bool = True,
|
1815 |
+
anyres_grids: list[int] = None,
|
1816 |
+
):
|
1817 |
+
"""
|
1818 |
+
Args:
|
1819 |
+
vision_encoder (nn.Module): HF CLIPModel
|
1820 |
+
lang_encoder (nn.Module): HF causal language model
|
1821 |
+
vis_feature_dim (int): final dimension of the visual features outputted by the vision_encoder
|
1822 |
+
initial_tokenizer_len (int): size of the tokenizer vocab
|
1823 |
+
padding_token_id (int): id of the padding token. None if no padding token; then a padding token
|
1824 |
+
will be inserted into self.special_tokens, which factory.py fills after creating new tokens
|
1825 |
+
decoder_layers_attr_name (str, optional): name of the decoder layers attribute. Defaults to None.
|
1826 |
+
gradient_checkpointing (bool, optional): whether to use gradient checkpointing. Defaults to False.
|
1827 |
+
"""
|
1828 |
+
self._special_tokens = {
|
1829 |
+
"media_token": "<image>",
|
1830 |
+
"image_placeholder_token": "<image placeholder>",
|
1831 |
+
"end_of_trunk_token": "<|endofchunk|>",
|
1832 |
+
}
|
1833 |
+
lang_embedding_dim = lang_model.get_input_embeddings().weight.shape[1]
|
1834 |
+
super().__init__(
|
1835 |
+
vision_encoder=vision_encoder,
|
1836 |
+
vision_tokenizer=vision_tokenizer,
|
1837 |
+
lang_model=lang_model,
|
1838 |
+
initial_tokenizer_len=initial_tokenizer_len,
|
1839 |
+
gradient_checkpointing=gradient_checkpointing,
|
1840 |
+
decoder_layers_attr_name=decoder_layers_attr_name,
|
1841 |
+
pad_token_id=pad_token_id,
|
1842 |
+
)
|
1843 |
+
self.image_aspect_ratio = image_aspect_ratio
|
1844 |
+
self.anyres_patch_sampling = anyres_patch_sampling
|
1845 |
+
self.anyres_grids = anyres_grids
|
1846 |
+
|
1847 |
+
def set_trainable(self):
|
1848 |
+
"""
|
1849 |
+
Unfreeze everything except the vision_encoder
|
1850 |
+
"""
|
1851 |
+
self.requires_grad_(True)
|
1852 |
+
self.vision_encoder.requires_grad_(False)
|
1853 |
+
|
1854 |
+
def _should_apply_weight_decay(self, parameter_name):
|
1855 |
+
"""
|
1856 |
+
Kosmos applies 0.01 weight deacy to everything
|
1857 |
+
"""
|
1858 |
+
return True
|
1859 |
+
|
1860 |
+
def generate(
|
1861 |
+
self,
|
1862 |
+
vision_x: torch.Tensor,
|
1863 |
+
lang_x: torch.Tensor,
|
1864 |
+
image_size: Optional[Tuple] = None,
|
1865 |
+
attention_mask: torch.Tensor = None,
|
1866 |
+
past_key_values: Optional[
|
1867 |
+
List[Union[torch.Tensor, Tuple[torch.Tensor]]]
|
1868 |
+
] = None,
|
1869 |
+
past_media_locations: Optional[torch.Tensor] = None,
|
1870 |
+
past_vision_tokens: Optional[torch.Tensor] = None,
|
1871 |
+
**kwargs,
|
1872 |
+
):
|
1873 |
+
"""
|
1874 |
+
Generate text conditioned on vision and language inputs.
|
1875 |
+
Args:
|
1876 |
+
vision_x (torch.Tensor): Vision input
|
1877 |
+
shape (B, T_img, F, C, H, W)
|
1878 |
+
see documentation for forward
|
1879 |
+
lang_x (torch.Tensor): Language input
|
1880 |
+
shape (B, T_txt)
|
1881 |
+
attention_mask (torch.Tensor, optional): Attention mask. Defaults to None.
|
1882 |
+
**kwargs: see generate documentation in Hugging Face CausalLM models.
|
1883 |
+
Returns:
|
1884 |
+
torch.Tensor: lang_x with generated tokens appended to it
|
1885 |
+
"""
|
1886 |
+
num_beams = kwargs.pop("num_beams", 1)
|
1887 |
+
|
1888 |
+
# convert pixels to vision tokens
|
1889 |
+
vision_attention_mask = None
|
1890 |
+
if vision_x is not None:
|
1891 |
+
if self.image_aspect_ratio == "anyres":
|
1892 |
+
input_dict = dict(image=vision_x, image_size=image_size)
|
1893 |
+
vision_features, vision_attn_masks = self._encode_vision_x_anyres(
|
1894 |
+
input_dict, lang_x.device
|
1895 |
+
)
|
1896 |
+
else:
|
1897 |
+
vision_features = self._encode_vision_x(vision_x=vision_x)
|
1898 |
+
vision_attn_masks = None
|
1899 |
+
# If doing patch sampling, then flatten patches of shape [b, Np_i, v, d] -> [b*Np, v, d]
|
1900 |
+
# Same for attention masks: [b, Np, v] -> [b*Np, v]
|
1901 |
+
if self.anyres_patch_sampling:
|
1902 |
+
split_sizes = [feature.shape[0] for feature in vision_features]
|
1903 |
+
# Nested splits for multi-image samples.
|
1904 |
+
if isinstance(vision_x[0], list):
|
1905 |
+
nt_images = [len(images) for images in vision_x]
|
1906 |
+
split_split_sizes = []
|
1907 |
+
img_id = 0
|
1908 |
+
for nt in nt_images:
|
1909 |
+
split_split_sizes.append(split_sizes[img_id : img_id + nt])
|
1910 |
+
img_id += nt
|
1911 |
+
else:
|
1912 |
+
nt_images = [1] * len(vision_x)
|
1913 |
+
split_split_sizes = split_sizes
|
1914 |
+
vision_features = torch.cat(vision_features, dim=0)
|
1915 |
+
vision_features = vision_features[
|
1916 |
+
:, None, None, :, :
|
1917 |
+
] # Expand dimensions.
|
1918 |
+
vision_attn_masks = torch.cat(vision_attn_masks, dim=0)
|
1919 |
+
vision_tokens = self.vision_tokenizer(vision_features, vision_attn_masks)
|
1920 |
+
|
1921 |
+
# Post-processing: Split the batches into groups of patches and concatenate them together.
|
1922 |
+
if self.anyres_patch_sampling:
|
1923 |
+
assert isinstance(vision_x, list)
|
1924 |
+
if isinstance(vision_x[0], list):
|
1925 |
+
vision_token_groups = torch.split(
|
1926 |
+
vision_tokens,
|
1927 |
+
list(sum(nt_img) for nt_img in split_split_sizes),
|
1928 |
+
dim=0,
|
1929 |
+
)
|
1930 |
+
vision_tokens = []
|
1931 |
+
|
1932 |
+
for sample_id, patch_vis_tokens in enumerate(vision_token_groups):
|
1933 |
+
patch_vis_token_groups = torch.split(
|
1934 |
+
patch_vis_tokens, split_split_sizes[sample_id], dim=0
|
1935 |
+
) # [Np*nt, 1, v, d] -> [[Np_t, 1, v, d], ...]
|
1936 |
+
flatten_vision_tokens = []
|
1937 |
+
for image_vis_token in patch_vis_token_groups:
|
1938 |
+
image_vis_token = image_vis_token.flatten(
|
1939 |
+
0, 2
|
1940 |
+
) # [Np, 1, v, d] -> [Np*v, d]
|
1941 |
+
flatten_vision_tokens.append(image_vis_token)
|
1942 |
+
vision_tokens_i = flatten_vision_tokens
|
1943 |
+
vision_tokens.append(vision_tokens_i)
|
1944 |
+
else:
|
1945 |
+
vision_token_groups = torch.split(vision_tokens, split_sizes, dim=0)
|
1946 |
+
vision_tokens = []
|
1947 |
+
for patch_vis_tokens in vision_token_groups:
|
1948 |
+
patch_vis_tokens = patch_vis_tokens.flatten(
|
1949 |
+
0, 2
|
1950 |
+
) # [Np, 1, v, d] -> [Np*v, d]
|
1951 |
+
vision_tokens.append(
|
1952 |
+
patch_vis_tokens.unsqueeze(0)
|
1953 |
+
) # Add the nt dimension.
|
1954 |
+
else:
|
1955 |
+
vision_tokens = None
|
1956 |
+
|
1957 |
+
# fuse the vision and language tokens
|
1958 |
+
# for xattn, vision_x and media_location are repeat_interleaved s.t.
|
1959 |
+
# the total batch size is B * num_beams
|
1960 |
+
new_inputs = self._prepare_inputs_for_forward(
|
1961 |
+
vision_tokens=vision_tokens,
|
1962 |
+
lang_x=lang_x,
|
1963 |
+
attention_mask=attention_mask,
|
1964 |
+
vision_attention_mask=vision_attention_mask,
|
1965 |
+
past_key_values=past_key_values,
|
1966 |
+
past_media_locations=past_media_locations,
|
1967 |
+
past_vision_tokens=past_vision_tokens,
|
1968 |
+
padding_side="left",
|
1969 |
+
num_beams=num_beams,
|
1970 |
+
)
|
1971 |
+
if past_key_values is not None:
|
1972 |
+
output = self.lang_model.generate(
|
1973 |
+
**new_inputs,
|
1974 |
+
past_key_values=past_key_values,
|
1975 |
+
num_beams=num_beams,
|
1976 |
+
use_cache=True,
|
1977 |
+
**kwargs,
|
1978 |
+
)
|
1979 |
+
else:
|
1980 |
+
output = self.lang_model.generate(
|
1981 |
+
**new_inputs,
|
1982 |
+
num_beams=num_beams,
|
1983 |
+
use_cache=True,
|
1984 |
+
**kwargs,
|
1985 |
+
)
|
1986 |
+
self._post_forward_hook()
|
1987 |
+
return output
|
1988 |
+
|
1989 |
|
1990 |
class XGenMMVisionEncoder(PreTrainedModel):
|
1991 |
main_input_name = "pixel_values"
|
1992 |
config_class = XGenMMVisionEncoderConfig
|
1993 |
+
|
1994 |
def __init__(self, config: XGenMMVisionEncoderConfig):
|
1995 |
super().__init__(config)
|
1996 |
+
if config.model_name != "google/siglip-so400m-patch14-384":
|
1997 |
+
raise ValueError(
|
1998 |
+
f"Unsupported model {config.model_name}. New vision models will be added soon."
|
1999 |
+
)
|
2000 |
self.model = AutoModel.from_pretrained(config.model_name)
|
2001 |
+
|
2002 |
def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
|
2003 |
# assert pixel_values.ndim == 4, f"Expected 4D tensor (bs, c, h, w), got {pixel_values.ndim}"
|
2004 |
return self.model.encode_image(pixel_values)
|
|
|
2005 |
|
2006 |
+
|
2007 |
+
# vision tokenizer
|
2008 |
class XGenMMVisionTokenizer(PreTrainedModel):
|
2009 |
config_class = XGenMMVisionTokenizerConfig
|
2010 |
+
|
2011 |
def __init__(self, config: XGenMMVisionTokenizerConfig):
|
2012 |
super().__init__(config)
|
2013 |
self.model = PerceiverResampler(
|
|
|
2015 |
dim_inner=config.lang_embedding_dim,
|
2016 |
num_latents=config.num_vis_tokens,
|
2017 |
)
|
2018 |
+
|
2019 |
+
def forward(self, vision_features: torch.Tensor, vision_attn_masks: torch.Tensor):
|
|
|
|
|
2020 |
return self.model(vision_features, vision_attn_masks)
|
2021 |
+
|
2022 |
+
|
2023 |
# XGenMM model
|
2024 |
class XGenMMModelForConditionalGeneration(PreTrainedModel):
|
2025 |
config_class = XGenMMConfig
|
2026 |
+
|
2027 |
def __init__(self, config: XGenMMConfig):
|
2028 |
super().__init__(config)
|
2029 |
+
|
2030 |
# vision encoder initialization
|
2031 |
+
vision_encoder = AutoModel.from_pretrained(
|
2032 |
+
config.vision_encoder_config.model_name
|
2033 |
+
).vision_model
|
2034 |
+
|
2035 |
+
# language model initialization
|
2036 |
language_model = AutoModelForCausalLM.from_config(config.text_config)
|
2037 |
check_embedding_fns(language_model)
|
2038 |
# Update _tied_weights_keys using the base model used.
|
2039 |
if language_model._tied_weights_keys is not None:
|
2040 |
+
self._tied_weights_keys = [
|
2041 |
+
f"language_model.{k}" for k in language_model._tied_weights_keys
|
2042 |
+
]
|
2043 |
+
|
2044 |
# vision tokenizer initialization
|
2045 |
+
if (
|
2046 |
+
config.vision_tokenizer_config.lang_embedding_dim
|
2047 |
+
!= language_model.get_input_embeddings().weight.shape[1]
|
2048 |
+
):
|
2049 |
overwrite = language_model.get_input_embeddings().weight.shape[1]
|
2050 |
config.vision_tokenizer_config.lang_embedding_dim = overwrite
|
2051 |
+
print(
|
2052 |
+
f"Warning: The language embedding dimension in the vision tokenizer config is different from the language model's embedding dimension. Overwriting the language embedding dimension in the vision tokenizer config to {overwrite}."
|
2053 |
+
)
|
2054 |
+
|
2055 |
vision_tokenizer = XGenMMVisionTokenizer(config.vision_tokenizer_config).model
|
2056 |
|
2057 |
self.vlm = XGenMMPerceiver(
|
2058 |
vision_encoder=vision_encoder,
|
2059 |
vision_tokenizer=vision_tokenizer,
|
2060 |
lang_model=language_model,
|
2061 |
+
initial_tokenizer_len=config.text_config.initial_tokenizer_len,
|
2062 |
+
pad_token_id=config.text_config.pad_token_id,
|
2063 |
+
image_aspect_ratio=config.vision_encoder_config.image_aspect_ratio,
|
2064 |
+
anyres_patch_sampling=config.vision_encoder_config.anyres_patch_sampling,
|
2065 |
+
anyres_grids=config.vision_encoder_config.anyres_grids,
|
2066 |
)
|
2067 |
# Initialize weights and apply final processing
|
2068 |
self.post_init()
|
2069 |
+
|
2070 |
@torch.no_grad()
|
2071 |
def generate(
|
2072 |
self,
|
|
|
2074 |
input_ids: Optional[torch.LongTensor] = None,
|
2075 |
attention_mask: Optional[torch.LongTensor] = None,
|
2076 |
**generate_kwargs,
|
2077 |
+
) -> torch.LongTensor:
|
2078 |
self.vlm = self.vlm.eval()
|
2079 |
return self.vlm.generate(
|
2080 |
+
vision_x=pixel_values,
|
2081 |
+
lang_x=input_ids,
|
2082 |
+
attention_mask=attention_mask,
|
2083 |
+
**generate_kwargs,
|
2084 |
+
)
|
2085 |
+
|
2086 |
def update_special_tokens(self, tokenizer):
|
2087 |
tokenizer.add_special_tokens(
|
2088 |
{"additional_special_tokens": list(self.vlm.special_tokens.values())}
|
|
|
2090 |
self.vlm.lang_model.config.vocab_size = len(tokenizer)
|
2091 |
self.vlm.set_special_token_ids(
|
2092 |
{
|
2093 |
+
v: tokenizer.convert_tokens_to_ids(v)
|
2094 |
+
for v in self.vlm.special_tokens.values()
|
2095 |
}
|
2096 |
)
|
2097 |
return tokenizer
|
|
utils.py
DELETED
@@ -1,383 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
import ast
|
3 |
-
import math
|
4 |
-
from PIL import Image
|
5 |
-
from packaging.version import Version
|
6 |
-
|
7 |
-
def has_fn(model, fn_name):
|
8 |
-
"""Check if model has a function fn_name"""
|
9 |
-
return callable(getattr(model, fn_name, None))
|
10 |
-
|
11 |
-
def exists(val):
|
12 |
-
return val is not None
|
13 |
-
|
14 |
-
def num_params(module, filter_to_trainable=False):
|
15 |
-
"""Returns the number of parameters in the module, or optionally only the trainable parameters"""
|
16 |
-
if filter_to_trainable:
|
17 |
-
return sum(p.numel() for p in module.parameters() if p.requires_grad)
|
18 |
-
else:
|
19 |
-
return sum(p.numel() for p in module.parameters())
|
20 |
-
|
21 |
-
def hasattr_recursive(obj, att):
|
22 |
-
"""
|
23 |
-
Check if obj has nested attribute
|
24 |
-
Example: hasattr_recursive(obj, 'a.b.c') is equivalent to hasattr(obj, 'a') and hasattr(obj.a, 'b') and hasattr(obj.a.b, 'c')
|
25 |
-
"""
|
26 |
-
if att == "":
|
27 |
-
return True
|
28 |
-
i = att.find(".")
|
29 |
-
if i < 0:
|
30 |
-
return hasattr(obj, att)
|
31 |
-
else:
|
32 |
-
try:
|
33 |
-
return hasattr_recursive(getattr(obj, att[:i]), att[i + 1 :])
|
34 |
-
except:
|
35 |
-
return False
|
36 |
-
|
37 |
-
def getattr_recursive(obj, att):
|
38 |
-
"""
|
39 |
-
Return nested attribute of obj
|
40 |
-
Example: getattr_recursive(obj, 'a.b.c') is equivalent to obj.a.b.c
|
41 |
-
"""
|
42 |
-
if att == "":
|
43 |
-
return obj
|
44 |
-
i = att.find(".")
|
45 |
-
if i < 0:
|
46 |
-
return getattr(obj, att)
|
47 |
-
else:
|
48 |
-
return getattr_recursive(getattr(obj, att[:i]), att[i + 1 :])
|
49 |
-
|
50 |
-
|
51 |
-
def setattr_recursive(obj, att, val):
|
52 |
-
"""
|
53 |
-
Set nested attribute of obj
|
54 |
-
Example: setattr_recursive(obj, 'a.b.c', val) is equivalent to obj.a.b.c = val
|
55 |
-
"""
|
56 |
-
if "." in att:
|
57 |
-
obj = getattr_recursive(obj, ".".join(att.split(".")[:-1]))
|
58 |
-
setattr(obj, att.split(".")[-1], val)
|
59 |
-
|
60 |
-
|
61 |
-
def stack_with_padding(list_of_tensors, padding_value=0, padding_side="right"):
|
62 |
-
"""
|
63 |
-
Stack a list of tensors with padding on one side
|
64 |
-
Args:
|
65 |
-
list_of_tensors (list[torch.Tensor]): List of tensors to stack
|
66 |
-
padding_value (int, optional): Value to pad with. Defaults to 0.
|
67 |
-
padding_side (str, optional): Side to pad on. Defaults to "right".
|
68 |
-
Returns:
|
69 |
-
torch.Tensor: Stacked tensors
|
70 |
-
"""
|
71 |
-
max_tokens = max(tensor.size(0) for tensor in list_of_tensors)
|
72 |
-
padded_tensors = []
|
73 |
-
for tensor in list_of_tensors:
|
74 |
-
num_tokens = tensor.size(0)
|
75 |
-
if len(tensor.size()) == 1:
|
76 |
-
padding = torch.full(
|
77 |
-
(max_tokens - num_tokens,),
|
78 |
-
padding_value,
|
79 |
-
dtype=tensor.dtype,
|
80 |
-
device=tensor.device,
|
81 |
-
)
|
82 |
-
else:
|
83 |
-
padding = torch.full(
|
84 |
-
(max_tokens - num_tokens, tensor.size(1)),
|
85 |
-
padding_value,
|
86 |
-
dtype=tensor.dtype,
|
87 |
-
device=tensor.device,
|
88 |
-
)
|
89 |
-
padded_tensor = (
|
90 |
-
torch.cat((tensor, padding), dim=0)
|
91 |
-
if padding_side == "right"
|
92 |
-
else torch.cat((padding, tensor), dim=0)
|
93 |
-
)
|
94 |
-
padded_tensors.append(padded_tensor)
|
95 |
-
return torch.stack(padded_tensors)
|
96 |
-
|
97 |
-
|
98 |
-
def check_embedding_fns(lang_model):
|
99 |
-
"""Checks for and attempts to set {get/set}_{input/output}_embeddings functions to the model"""
|
100 |
-
if not has_fn(lang_model, "get_input_embeddings"):
|
101 |
-
if hasattr_recursive(lang_model, "transformer.wte"): # MPT
|
102 |
-
lang_model.get_input_embeddings = lambda: lang_model.transformer.wte
|
103 |
-
elif hasattr_recursive(lang_model, "model.decoder.embed_tokens"): # OPT
|
104 |
-
lang_model.get_input_embeddings = lambda: lang_model.decoder.embed_tokens
|
105 |
-
else:
|
106 |
-
raise ValueError(
|
107 |
-
"We require the language encoder to have a get_input_embeddings method but we couldn't determine the name of the input embeddings attribute. Please supply this manually in factory.py."
|
108 |
-
)
|
109 |
-
|
110 |
-
if not has_fn(lang_model, "set_input_embeddings"):
|
111 |
-
if hasattr_recursive(lang_model, "transformer.wte"): # MPT
|
112 |
-
lang_model.set_input_embeddings = lambda x: setattr_recursive(
|
113 |
-
lang_model, "transformer.wte", x
|
114 |
-
)
|
115 |
-
elif hasattr_recursive(lang_model, "model.decoder.embed_tokens"): # OPT
|
116 |
-
lang_model.set_input_embeddings = lambda x: setattr_recursive(
|
117 |
-
lang_model, "model.decoder.embed_tokens", x
|
118 |
-
)
|
119 |
-
else:
|
120 |
-
raise ValueError(
|
121 |
-
"We require the language encoder to have a set_input_embeddings method but we couldn't determine the name of the input embeddings attribute. Please supply this manually in factory.py."
|
122 |
-
)
|
123 |
-
|
124 |
-
if not has_fn(lang_model, "get_output_embeddings"):
|
125 |
-
if hasattr_recursive(lang_model, "lm_head"):
|
126 |
-
lang_model.get_output_embeddings = lambda: lang_model.lm_head
|
127 |
-
else:
|
128 |
-
raise ValueError(
|
129 |
-
"We require the language encoder to have a get_output_embeddings method but we couldn't determine the name of the output embeddings attribute. Please supply this manually in factory.py."
|
130 |
-
)
|
131 |
-
|
132 |
-
if not has_fn(lang_model, "set_output_embeddings"):
|
133 |
-
if hasattr_recursive(lang_model, "lm_head"):
|
134 |
-
lang_model.set_output_embeddings = lambda x: setattr_recursive(
|
135 |
-
lang_model, "lm_head", x
|
136 |
-
)
|
137 |
-
else:
|
138 |
-
raise ValueError(
|
139 |
-
"We require the language encoder to have a set_output_embeddings method but we couldn't determine the name of the output embeddings attribute. Please supply this manually in factory.py."
|
140 |
-
)
|
141 |
-
|
142 |
-
|
143 |
-
def has_fn(model, fn_name):
|
144 |
-
"""Check if model has a function fn_name"""
|
145 |
-
return callable(getattr(model, fn_name, None))
|
146 |
-
|
147 |
-
|
148 |
-
# Adopted from https://github.com/haotian-liu/LLaVA. Below is the original copyright:
|
149 |
-
#
|
150 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
151 |
-
# you may not use this file except in compliance with the License.
|
152 |
-
# You may obtain a copy of the License at
|
153 |
-
#
|
154 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
155 |
-
#
|
156 |
-
# Unless required by applicable law or agreed to in writing, software
|
157 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
158 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
159 |
-
# See the License for the specific language governing permissions and
|
160 |
-
# limitations under the License.
|
161 |
-
|
162 |
-
def unpad_image(tensor, original_size, keep_original_shape=False):
|
163 |
-
"""
|
164 |
-
Unpads a PyTorch tensor of a padded and resized image.
|
165 |
-
|
166 |
-
Args:
|
167 |
-
tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
|
168 |
-
original_size (tuple): The original size of the image (height, width).
|
169 |
-
|
170 |
-
Returns:
|
171 |
-
torch.Tensor: The unpadded image tensor.
|
172 |
-
"""
|
173 |
-
original_width, original_height = original_size
|
174 |
-
current_height, current_width = tensor.shape[1:]
|
175 |
-
|
176 |
-
original_aspect_ratio = original_width / original_height
|
177 |
-
current_aspect_ratio = current_width / current_height
|
178 |
-
|
179 |
-
if original_aspect_ratio > current_aspect_ratio:
|
180 |
-
scale_factor = current_width / original_width
|
181 |
-
new_height = int(original_height * scale_factor)
|
182 |
-
padding = (current_height - new_height) // 2
|
183 |
-
if keep_original_shape:
|
184 |
-
attention_mask = torch.ones((current_height, current_width), device=tensor.device)
|
185 |
-
attention_mask[:padding, :] = 0
|
186 |
-
attention_mask[current_height - padding:, :] = 0
|
187 |
-
return tensor, attention_mask
|
188 |
-
else:
|
189 |
-
unpadded_tensor = tensor[:, padding:current_height - padding, :]
|
190 |
-
return unpadded_tensor, None
|
191 |
-
else:
|
192 |
-
scale_factor = current_height / original_height
|
193 |
-
new_width = int(original_width * scale_factor)
|
194 |
-
padding = (current_width - new_width) // 2
|
195 |
-
if keep_original_shape:
|
196 |
-
attention_mask = torch.ones((current_height, current_width), device=tensor.device)
|
197 |
-
attention_mask[:, :padding] = 0
|
198 |
-
attention_mask[:, current_width - padding:] = 0
|
199 |
-
return tensor, attention_mask
|
200 |
-
else:
|
201 |
-
unpadded_tensor = tensor[:, :, padding:current_width - padding]
|
202 |
-
return unpadded_tensor, None
|
203 |
-
|
204 |
-
|
205 |
-
def select_best_resolution(original_size, possible_resolutions):
|
206 |
-
"""
|
207 |
-
Selects the best resolution from a list of possible resolutions based on the original size.
|
208 |
-
|
209 |
-
Args:
|
210 |
-
original_size (tuple): The original size of the image in the format (width, height).
|
211 |
-
possible_resolutions (list): A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
|
212 |
-
|
213 |
-
Returns:
|
214 |
-
tuple: The best fit resolution in the format (width, height).
|
215 |
-
"""
|
216 |
-
original_width, original_height = original_size
|
217 |
-
best_fit = None
|
218 |
-
max_effective_resolution = 0
|
219 |
-
min_wasted_resolution = float('inf')
|
220 |
-
|
221 |
-
for width, height in possible_resolutions:
|
222 |
-
scale = min(width / original_width, height / original_height)
|
223 |
-
downscaled_width, downscaled_height = int(original_width * scale), int(original_height * scale)
|
224 |
-
effective_resolution = min(downscaled_width * downscaled_height, original_width * original_height)
|
225 |
-
wasted_resolution = (width * height) - effective_resolution
|
226 |
-
|
227 |
-
if effective_resolution > max_effective_resolution or (effective_resolution == max_effective_resolution and wasted_resolution < min_wasted_resolution):
|
228 |
-
max_effective_resolution = effective_resolution
|
229 |
-
min_wasted_resolution = wasted_resolution
|
230 |
-
best_fit = (width, height)
|
231 |
-
|
232 |
-
return best_fit
|
233 |
-
|
234 |
-
|
235 |
-
def resize_and_pad_image(image, target_resolution):
|
236 |
-
"""
|
237 |
-
Resize and pad an image to a target resolution while maintaining aspect ratio.
|
238 |
-
|
239 |
-
Args:
|
240 |
-
image (PIL.Image.Image): The input image.
|
241 |
-
target_resolution (tuple): The target resolution (width, height) of the image.
|
242 |
-
|
243 |
-
Returns:
|
244 |
-
PIL.Image.Image: The resized and padded image.
|
245 |
-
"""
|
246 |
-
original_width, original_height = image.size
|
247 |
-
target_width, target_height = target_resolution
|
248 |
-
|
249 |
-
scale_w = target_width / original_width
|
250 |
-
scale_h = target_height / original_height
|
251 |
-
|
252 |
-
if scale_w < scale_h:
|
253 |
-
new_width = target_width
|
254 |
-
new_height = min(math.ceil(original_height * scale_w), target_height)
|
255 |
-
else:
|
256 |
-
new_height = target_height
|
257 |
-
new_width = min(math.ceil(original_width * scale_h), target_width)
|
258 |
-
|
259 |
-
# Resize the image
|
260 |
-
resized_image = image.resize((new_width, new_height))
|
261 |
-
|
262 |
-
new_image = Image.new('RGB', (target_width, target_height), (0, 0, 0))
|
263 |
-
paste_x = (target_width - new_width) // 2
|
264 |
-
paste_y = (target_height - new_height) // 2
|
265 |
-
new_image.paste(resized_image, (paste_x, paste_y))
|
266 |
-
|
267 |
-
return new_image
|
268 |
-
|
269 |
-
|
270 |
-
def divide_to_patches(image, patch_size):
|
271 |
-
"""
|
272 |
-
Divides an image into patches of a specified size.
|
273 |
-
|
274 |
-
Args:
|
275 |
-
image (PIL.Image.Image): The input image.
|
276 |
-
patch_size (int): The size of each patch.
|
277 |
-
|
278 |
-
Returns:
|
279 |
-
list: A list of PIL.Image.Image objects representing the patches.
|
280 |
-
"""
|
281 |
-
patches = []
|
282 |
-
width, height = image.size
|
283 |
-
for i in range(0, height, patch_size):
|
284 |
-
for j in range(0, width, patch_size):
|
285 |
-
box = (j, i, j + patch_size, i + patch_size)
|
286 |
-
patch = image.crop(box)
|
287 |
-
patches.append(patch)
|
288 |
-
|
289 |
-
return patches
|
290 |
-
|
291 |
-
|
292 |
-
def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
|
293 |
-
"""
|
294 |
-
Calculate the shape of the image patch grid after the preprocessing for images of any resolution.
|
295 |
-
|
296 |
-
Args:
|
297 |
-
image_size (tuple): The size of the input image in the format (width, height).
|
298 |
-
grid_pinpoints (str): A string representation of a list of possible resolutions.
|
299 |
-
patch_size (int): The size of each image patch.
|
300 |
-
|
301 |
-
Returns:
|
302 |
-
tuple: The shape of the image patch grid in the format (width, height).
|
303 |
-
"""
|
304 |
-
if type(grid_pinpoints) is list:
|
305 |
-
possible_resolutions = grid_pinpoints
|
306 |
-
else:
|
307 |
-
possible_resolutions = ast.literal_eval(grid_pinpoints)
|
308 |
-
width, height = select_best_resolution(image_size, possible_resolutions)
|
309 |
-
return width // patch_size, height // patch_size
|
310 |
-
|
311 |
-
|
312 |
-
def process_anyres_image(image, processor, grid_pinpoints):
|
313 |
-
"""
|
314 |
-
Process an image with variable resolutions.
|
315 |
-
|
316 |
-
Args:
|
317 |
-
image (PIL.Image.Image): The input image to be processed.
|
318 |
-
processor: The image processor object.
|
319 |
-
grid_pinpoints (str): A string representation of a list of possible resolutions.
|
320 |
-
|
321 |
-
Returns:
|
322 |
-
torch.Tensor: A tensor containing the processed image patches.
|
323 |
-
"""
|
324 |
-
# FIXME: determine grid_pinpoints from image sizes.
|
325 |
-
if type(grid_pinpoints) is list:
|
326 |
-
possible_resolutions = grid_pinpoints
|
327 |
-
else:
|
328 |
-
possible_resolutions = ast.literal_eval(grid_pinpoints)
|
329 |
-
best_resolution = select_best_resolution(image.size, possible_resolutions)
|
330 |
-
image_padded = resize_and_pad_image(image, best_resolution)
|
331 |
-
|
332 |
-
processor_size = processor.transforms[0].size
|
333 |
-
patches = divide_to_patches(image_padded, processor_size[0])
|
334 |
-
|
335 |
-
image_original_resize = image.resize((processor_size[0], processor_size[0]))
|
336 |
-
|
337 |
-
image_patches = [image_original_resize] + patches
|
338 |
-
image_patches = [processor(image_patch)
|
339 |
-
for image_patch in image_patches]
|
340 |
-
return torch.stack(image_patches, dim=0)
|
341 |
-
|
342 |
-
|
343 |
-
def expand2square(pil_img, background_color):
|
344 |
-
width, height = pil_img.size
|
345 |
-
if width == height:
|
346 |
-
return pil_img
|
347 |
-
elif width > height:
|
348 |
-
result = Image.new(pil_img.mode, (width, width), background_color)
|
349 |
-
result.paste(pil_img, (0, (width - height) // 2))
|
350 |
-
return result
|
351 |
-
else:
|
352 |
-
result = Image.new(pil_img.mode, (height, height), background_color)
|
353 |
-
result.paste(pil_img, ((height - width) // 2, 0))
|
354 |
-
return result
|
355 |
-
|
356 |
-
|
357 |
-
def process_images(images, image_processor, model_cfg):
|
358 |
-
image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
|
359 |
-
new_images = []
|
360 |
-
if image_aspect_ratio == 'pad':
|
361 |
-
for image in images:
|
362 |
-
image = expand2square(image, tuple(int(x*255) for x in image_processor.transforms[-1].mean))
|
363 |
-
image = image_processor(image)
|
364 |
-
new_images.append(image)
|
365 |
-
elif image_aspect_ratio in ["anyres", "anyres-legacy"]:
|
366 |
-
base_img_size = image_processor.transforms[0].size[0]
|
367 |
-
for image in images:
|
368 |
-
image = process_anyres_image(image, image_processor, [[base_img_size,base_img_size*2],
|
369 |
-
[base_img_size*2,base_img_size],
|
370 |
-
[base_img_size*2,base_img_size*2],
|
371 |
-
[base_img_size*3,base_img_size],
|
372 |
-
[base_img_size,base_img_size*3]])
|
373 |
-
|
374 |
-
# Debug any res inference by only using 672x672.
|
375 |
-
# image = process_anyres_image(image, image_processor, [[base_img_size*2,base_img_size*2]])
|
376 |
-
new_images.append(image)
|
377 |
-
else:
|
378 |
-
return image_processor(images)
|
379 |
-
if all(x.shape == new_images[0].shape for x in new_images):
|
380 |
-
new_images = torch.stack(new_images, dim=0)
|
381 |
-
return new_images
|
382 |
-
|
383 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vlm.py
DELETED
@@ -1,1381 +0,0 @@
|
|
1 |
-
|
2 |
-
import torch
|
3 |
-
from torch import einsum, nn
|
4 |
-
from einops import rearrange, repeat
|
5 |
-
from einops_exts import rearrange_many
|
6 |
-
from einops import rearrange
|
7 |
-
from typing import List, Optional, Tuple, Union
|
8 |
-
import torch.nn.functional as F
|
9 |
-
from transformers.modeling_outputs import CausalLMOutputWithPast
|
10 |
-
from dataclasses import dataclass
|
11 |
-
from transformers import CLIPVisionModel
|
12 |
-
from transformers.models.siglip.modeling_siglip import SiglipVisionTransformer
|
13 |
-
|
14 |
-
import transformers
|
15 |
-
from packaging.version import Version
|
16 |
-
|
17 |
-
from .utils import num_params, getattr_recursive, stack_with_padding, get_anyres_image_grid_shape, unpad_image
|
18 |
-
|
19 |
-
|
20 |
-
class VisionTokenizer(nn.Module):
|
21 |
-
def __init__(self, dim_media, num_tokens_per_media):
|
22 |
-
super().__init__()
|
23 |
-
self.dim_media = dim_media
|
24 |
-
self.num_tokens_per_media = num_tokens_per_media
|
25 |
-
|
26 |
-
class PerceiverAttention(nn.Module):
|
27 |
-
def __init__(self, *, dim, dim_head=64, heads=8):
|
28 |
-
super().__init__()
|
29 |
-
self.scale = dim_head**-0.5
|
30 |
-
self.heads = heads
|
31 |
-
inner_dim = dim_head * heads
|
32 |
-
|
33 |
-
self.norm_media = nn.LayerNorm(dim)
|
34 |
-
self.norm_latents = nn.LayerNorm(dim)
|
35 |
-
|
36 |
-
self.to_q = nn.Linear(dim, inner_dim, bias=False)
|
37 |
-
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
|
38 |
-
self.to_out = nn.Linear(inner_dim, dim, bias=False)
|
39 |
-
|
40 |
-
def forward(self, x, latents, vision_attn_masks=None):
|
41 |
-
"""
|
42 |
-
Args:
|
43 |
-
x (torch.Tensor): image features
|
44 |
-
shape (b, T, n1, D)
|
45 |
-
latent (torch.Tensor): latent features
|
46 |
-
shape (b, T, n2, D)
|
47 |
-
"""
|
48 |
-
x = self.norm_media(x)
|
49 |
-
latents = self.norm_latents(latents)
|
50 |
-
|
51 |
-
h = self.heads
|
52 |
-
|
53 |
-
q = self.to_q(latents)
|
54 |
-
kv_input = torch.cat((x, latents), dim=-2) # TODO: Change the shape of vision attention mask according to this.
|
55 |
-
if vision_attn_masks is not None:
|
56 |
-
vision_attn_masks = torch.cat((vision_attn_masks,
|
57 |
-
torch.ones((latents.shape[0], latents.shape[-2]), dtype=latents.dtype, device=latents.device)),
|
58 |
-
dim=-1)
|
59 |
-
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
|
60 |
-
q, k, v = rearrange_many((q, k, v), "b t n (h d) -> b h t n d", h=h)
|
61 |
-
q = q * self.scale
|
62 |
-
|
63 |
-
# attention
|
64 |
-
sim = einsum("... i d, ... j d -> ... i j", q, k)
|
65 |
-
# Apply vision attention mask here.
|
66 |
-
# Reference: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html#torch.nn.functional.scaled_dot_product_attention
|
67 |
-
if vision_attn_masks is not None:
|
68 |
-
attn_bias = torch.zeros((q.size(0), 1, 1, q.size(-2), k.size(-2)), dtype=q.dtype, device=q.device)
|
69 |
-
vision_attn_masks = repeat(vision_attn_masks, 'b n -> b 1 1 l n', l=q.size(-2))
|
70 |
-
attn_bias.masked_fill_(vision_attn_masks.logical_not(), float("-inf"))
|
71 |
-
sim += attn_bias
|
72 |
-
|
73 |
-
sim = sim - sim.amax(dim=-1, keepdim=True).detach()
|
74 |
-
attn = sim.softmax(dim=-1)
|
75 |
-
|
76 |
-
|
77 |
-
out = einsum("... i j, ... j d -> ... i d", attn, v)
|
78 |
-
out = rearrange(out, "b h t n d -> b t n (h d)", h=h)
|
79 |
-
return self.to_out(out)
|
80 |
-
|
81 |
-
|
82 |
-
def FeedForward(dim, mult=4):
|
83 |
-
inner_dim = int(dim * mult)
|
84 |
-
return nn.Sequential(
|
85 |
-
nn.LayerNorm(dim),
|
86 |
-
nn.Linear(dim, inner_dim, bias=False),
|
87 |
-
nn.GELU(),
|
88 |
-
nn.Linear(inner_dim, dim, bias=False),
|
89 |
-
)
|
90 |
-
|
91 |
-
|
92 |
-
class PerceiverResampler(VisionTokenizer):
|
93 |
-
def __init__(
|
94 |
-
self,
|
95 |
-
*,
|
96 |
-
dim,
|
97 |
-
dim_inner=None,
|
98 |
-
depth=6,
|
99 |
-
dim_head=96,
|
100 |
-
heads=16,
|
101 |
-
num_latents=128,
|
102 |
-
max_num_media=None,
|
103 |
-
max_num_frames=None,
|
104 |
-
ff_mult=4,
|
105 |
-
):
|
106 |
-
"""
|
107 |
-
Perceiver module which takes in image features and outputs image tokens.
|
108 |
-
Args:
|
109 |
-
dim (int): dimension of the incoming image features
|
110 |
-
dim_inner (int, optional): final dimension to project the incoming image features to;
|
111 |
-
also the final dimension of the outputted features. If None, no projection is used, and dim_inner = dim.
|
112 |
-
depth (int, optional): number of layers. Defaults to 6.
|
113 |
-
dim_head (int, optional): dimension of each head. Defaults to 64.
|
114 |
-
heads (int, optional): number of heads. Defaults to 8.
|
115 |
-
num_latents (int, optional): number of latent tokens to use in the Perceiver;
|
116 |
-
also corresponds to number of tokens per sequence to output. Defaults to 64.
|
117 |
-
max_num_media (int, optional): maximum number of media per sequence to input into the Perceiver
|
118 |
-
and keep positional embeddings for. If None, no positional embeddings are used.
|
119 |
-
max_num_frames (int, optional): maximum number of frames to input into the Perceiver
|
120 |
-
and keep positional embeddings for. If None, no positional embeddings are used.
|
121 |
-
ff_mult (int, optional): dimension multiplier for the feedforward network. Defaults to 4.
|
122 |
-
"""
|
123 |
-
if dim_inner is not None:
|
124 |
-
projection = nn.Linear(dim, dim_inner)
|
125 |
-
else:
|
126 |
-
projection = None
|
127 |
-
dim_inner = dim
|
128 |
-
super().__init__(dim_media=dim, num_tokens_per_media=num_latents)
|
129 |
-
self.projection = projection
|
130 |
-
self.latents = nn.Parameter(torch.randn(num_latents, dim))
|
131 |
-
|
132 |
-
# positional embeddings
|
133 |
-
self.frame_embs = (
|
134 |
-
nn.Parameter(torch.randn(max_num_frames, dim))
|
135 |
-
if exists(max_num_frames)
|
136 |
-
else None
|
137 |
-
)
|
138 |
-
self.media_time_embs = (
|
139 |
-
nn.Parameter(torch.randn(max_num_media, 1, dim))
|
140 |
-
if exists(max_num_media)
|
141 |
-
else None
|
142 |
-
)
|
143 |
-
|
144 |
-
self.layers = nn.ModuleList([])
|
145 |
-
for _ in range(depth):
|
146 |
-
self.layers.append(
|
147 |
-
nn.ModuleList(
|
148 |
-
[
|
149 |
-
PerceiverAttention(
|
150 |
-
dim=dim, dim_head=dim_head, heads=heads
|
151 |
-
),
|
152 |
-
FeedForward(dim=dim, mult=ff_mult),
|
153 |
-
]
|
154 |
-
)
|
155 |
-
)
|
156 |
-
|
157 |
-
self.norm = nn.LayerNorm(dim)
|
158 |
-
|
159 |
-
def forward(self, x, vision_attn_masks):
|
160 |
-
"""
|
161 |
-
Args:
|
162 |
-
x (torch.Tensor): image features
|
163 |
-
shape (b, T, F, v, D)
|
164 |
-
vision_attn_masks (torch.Tensor): attention masks for padded visiont tokens (i.e., x)
|
165 |
-
shape (b, v)
|
166 |
-
Returns:
|
167 |
-
shape (b, T, n, D) where n is self.num_latents
|
168 |
-
"""
|
169 |
-
b, T, F, v = x.shape[:4]
|
170 |
-
|
171 |
-
# frame and media time embeddings
|
172 |
-
if exists(self.frame_embs):
|
173 |
-
frame_embs = repeat(self.frame_embs[:F], "F d -> b T F v d", b=b, T=T, v=v)
|
174 |
-
x = x + frame_embs
|
175 |
-
x = rearrange(
|
176 |
-
x, "b T F v d -> b T (F v) d"
|
177 |
-
) # flatten the frame and spatial dimensions
|
178 |
-
if exists(self.media_time_embs):
|
179 |
-
x = x + self.media_time_embs[:T]
|
180 |
-
|
181 |
-
# blocks
|
182 |
-
latents = self.latents
|
183 |
-
latents = repeat(latents, "n d -> b T n d", b=b, T=T)
|
184 |
-
for attn, ff in self.layers:
|
185 |
-
latents = attn(x, latents, vision_attn_masks) + latents
|
186 |
-
latents = ff(latents) + latents
|
187 |
-
|
188 |
-
if exists(self.projection):
|
189 |
-
return self.projection(self.norm(latents))
|
190 |
-
else:
|
191 |
-
return self.norm(latents)
|
192 |
-
|
193 |
-
|
194 |
-
class DecoupledEmbedding(nn.Embedding):
|
195 |
-
# Derived from https://pytorch.org/docs/stable/_modules/torch/nn/modules/sparse.html#Embedding
|
196 |
-
"""
|
197 |
-
Implements a decoupling of parameters to allow freezing (or not) a subset of the embeddings. In practise, the
|
198 |
-
regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `num_additional_embeddings` > 0,
|
199 |
-
then it will create `num_additional_embeddings` additional parameters that are always trained. If
|
200 |
-
`num_additional_embeddings=0`, then the module defaults back to the regular behavior of `nn.Embedding`.
|
201 |
-
"""
|
202 |
-
|
203 |
-
def __init__(
|
204 |
-
self,
|
205 |
-
max_original_id: int,
|
206 |
-
num_additional_embeddings: int = 0,
|
207 |
-
_weight: torch.Tensor = None,
|
208 |
-
num_original_embeddings: int = None,
|
209 |
-
embedding_dim: int = None,
|
210 |
-
partially_freeze=True,
|
211 |
-
device=None,
|
212 |
-
dtype=None,
|
213 |
-
pad_token_id=None,
|
214 |
-
) -> None:
|
215 |
-
"""
|
216 |
-
Args:
|
217 |
-
max_original_id (`int`):
|
218 |
-
The largest token id that should be embedded using the regular embedding (regular `weight`).
|
219 |
-
This is usually len(tokenizer) - 1 before additional tokens are added.
|
220 |
-
Note that this may not equal self.weight.shape[0]
|
221 |
-
num_additional_embeddings (`int`):
|
222 |
-
Number of additional tokens to initialize an Embedding matrix for (`additional_weight`).
|
223 |
-
_weight (`torch.Tensor`, *optional*, defaults to `None`): The regular weight tensor.
|
224 |
-
If provided, this sets the `num_original_embeddings` and `embedding_dim` parameters.
|
225 |
-
num_original_embeddings (`int`):
|
226 |
-
self.weight.shape[0]
|
227 |
-
embedding_dim (`int`):
|
228 |
-
The size of each embedding vector
|
229 |
-
partially_freeze: (`bool`, *optional*, defaults to `True`):
|
230 |
-
If `True`, the regular `weight` will be frozen. `additional_weight` is never frozen.
|
231 |
-
padding_idx (`int`, *optional*):
|
232 |
-
The padding index (needs to be less than num_embeddings)
|
233 |
-
|
234 |
-
Note: there are a lot of other parameters to initialize a standard `nn.Embedding` such as `padding_idx`,
|
235 |
-
`max_norm` or `norm_type`. We are not supporting these.
|
236 |
-
"""
|
237 |
-
# validate args
|
238 |
-
if pad_token_id is not None and pad_token_id > max_original_id:
|
239 |
-
raise ValueError(
|
240 |
-
f"pad_token_id must be <= max_original_id. Got {pad_token_id} and {max_original_id}."
|
241 |
-
+ "If the original tokenizer does not have a pad_token_id, use pad_token_id=None."
|
242 |
-
)
|
243 |
-
if _weight is not None:
|
244 |
-
assert (num_original_embeddings is None) or (
|
245 |
-
_weight.shape[0] == num_original_embeddings
|
246 |
-
), f"num_original_embeddings={num_original_embeddings} but _weight.shape[0]={_weight.shape[0]}"
|
247 |
-
assert (embedding_dim is None) or (
|
248 |
-
_weight.shape[1] == embedding_dim
|
249 |
-
), f"embedding_dim={embedding_dim} but _weight.shape[1]={_weight.shape[1]}"
|
250 |
-
num_original_embeddings = _weight.shape[0]
|
251 |
-
embedding_dim = _weight.shape[1]
|
252 |
-
else:
|
253 |
-
assert (
|
254 |
-
num_original_embeddings is not None
|
255 |
-
), "num_original_embeddings must be provided if _weight is not provided"
|
256 |
-
assert (
|
257 |
-
embedding_dim is not None
|
258 |
-
), "embedding_dim must be provided if _weight is not provided"
|
259 |
-
|
260 |
-
super().__init__(
|
261 |
-
num_embeddings=num_original_embeddings,
|
262 |
-
embedding_dim=embedding_dim,
|
263 |
-
device=device,
|
264 |
-
dtype=dtype,
|
265 |
-
padding_idx=pad_token_id,
|
266 |
-
_weight=_weight,
|
267 |
-
)
|
268 |
-
self.max_original_id = max_original_id
|
269 |
-
self.padding_idx = pad_token_id
|
270 |
-
self.num_additional_embeddings = num_additional_embeddings
|
271 |
-
if self.num_additional_embeddings > 0:
|
272 |
-
self.additional_embedding = nn.Embedding(
|
273 |
-
num_embeddings=self.num_additional_embeddings,
|
274 |
-
embedding_dim=embedding_dim,
|
275 |
-
device=device,
|
276 |
-
dtype=dtype,
|
277 |
-
)
|
278 |
-
self.set_requires_grad(
|
279 |
-
require_regular_grad=not partially_freeze, require_additional_grad=True
|
280 |
-
)
|
281 |
-
|
282 |
-
def set_requires_grad(self, require_regular_grad, require_additional_grad):
|
283 |
-
"""
|
284 |
-
Helper function to separately set the requires_grad flag for the regular weight and the additional weight.
|
285 |
-
"""
|
286 |
-
self.weight.requires_grad_(require_regular_grad)
|
287 |
-
self.additional_embedding.requires_grad_(require_additional_grad)
|
288 |
-
|
289 |
-
def forward(self, input_ids):
|
290 |
-
"""
|
291 |
-
we have 2 embeddings, with different indices - one pretrained self.weight and another
|
292 |
-
self.additional_embedding.weight that is being trained.
|
293 |
-
|
294 |
-
in order to make a lookup of the input ids, we:
|
295 |
-
1. find out the indices of the entries belonging to the 2nd embedding
|
296 |
-
2. extract those values while subtracting the size of the first embedding (num_embeddings), since the 2nd
|
297 |
-
embedding starts from 0 and not num_embeddings
|
298 |
-
3. perform the 2nd embedding lookup
|
299 |
-
4. now we handle the 1st embedding, we overwrite indices belonging to the 2nd embedding with a padding index
|
300 |
-
5. perform the 1st embedding lookup
|
301 |
-
6. now we overwrite the values in the 1st embedding lookup with the values of the 2nd embedding lookup
|
302 |
-
|
303 |
-
note: for the 1st embedding lookup we could have looked up only the low indices and not do the padding, but
|
304 |
-
then we have to create a new tensor and populate it with 2 tensors that are spread out across various indices -
|
305 |
-
i.e. not a simple concat - I haven't benchmarked the complex case if it's any faster, given that seqlens are
|
306 |
-
usually relatively short it's probably not faster or if faster not by much - but might be a good idea to
|
307 |
-
measure.
|
308 |
-
|
309 |
-
"""
|
310 |
-
if self.num_additional_embeddings == 0:
|
311 |
-
return F.embedding(input_ids, self.weight)
|
312 |
-
|
313 |
-
# Clone so that we don't modify the original input_ids later on
|
314 |
-
input_ids = input_ids.clone()
|
315 |
-
additional_vocab_indices = torch.where(input_ids > self.max_original_id)
|
316 |
-
input_ids_additional_vocab = input_ids[additional_vocab_indices]
|
317 |
-
additional_embeddings = self.additional_embedding(
|
318 |
-
input_ids_additional_vocab - self.max_original_id - 1
|
319 |
-
)
|
320 |
-
|
321 |
-
# for successful lookup replace input_ids with 0, the results of these will be discarded anyway
|
322 |
-
input_ids[additional_vocab_indices] = 0
|
323 |
-
full_vector = F.embedding(input_ids, self.weight)
|
324 |
-
|
325 |
-
# overwrite the records with high indices
|
326 |
-
full_vector[additional_vocab_indices] = additional_embeddings
|
327 |
-
|
328 |
-
return full_vector
|
329 |
-
|
330 |
-
def extra_repr(self) -> str:
|
331 |
-
return "num_original_embeddings={}, num_additional_embeddings={}, embedding_dim={}, partially_freeze={}".format(
|
332 |
-
self.max_original_id + 1,
|
333 |
-
self.num_additional_embeddings,
|
334 |
-
self.embedding_dim,
|
335 |
-
(not self.weight.requires_grad),
|
336 |
-
)
|
337 |
-
|
338 |
-
|
339 |
-
class DecoupledLinear(nn.Linear):
|
340 |
-
# Derived from https://pytorch.org/docs/stable/_modules/torch/nn/modules/linear.html#Linear
|
341 |
-
"""
|
342 |
-
Implements a decoupling of parameters to allow freezing (or not) a subset of the parameters. In practise, the
|
343 |
-
regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `additional_out_features` > 0,
|
344 |
-
then it will create `additional_out_features * in_features` additional parameters that are always trained. If
|
345 |
-
`additional_out_features=0`, then the module defaults back to the regular behavior of `nn.Linear`.
|
346 |
-
"""
|
347 |
-
|
348 |
-
def __init__(
|
349 |
-
self,
|
350 |
-
max_original_id: int,
|
351 |
-
additional_out_features: int = 0,
|
352 |
-
_weight: torch.Tensor = None,
|
353 |
-
_bias: torch.Tensor = None,
|
354 |
-
in_features: int = None,
|
355 |
-
original_out_features: int = None,
|
356 |
-
bias: bool = True,
|
357 |
-
partially_freeze: bool = True,
|
358 |
-
device=None,
|
359 |
-
dtype=None,
|
360 |
-
) -> None:
|
361 |
-
"""
|
362 |
-
Args:
|
363 |
-
max_original_id (`int`): The largest token id that should be extracted from the regular weight.
|
364 |
-
This is usually len(tokenizer) - 1 before additional tokens are added.
|
365 |
-
Note that this may not equal original_out_features - 1
|
366 |
-
_weight: torch.Tensor, *optional*, defaults to `None`. The regular weight tensor.
|
367 |
-
If provided, this sets the `in_features` and `original_out_features` parameters.
|
368 |
-
_bias: torch.Tensor, *optional*, defaults to `None`. The regular bias tensor.
|
369 |
-
in_features: int. Input hidden size.
|
370 |
-
original_out_features: int. Original out_features of the language model's get_output_embeddings() function.
|
371 |
-
additional_out_features: int. Number of additional trainable dimensions.
|
372 |
-
bias: bool. Whether to include a bias term.
|
373 |
-
partially_freeze: bool, *optional*, defaults to `True`): If `True`, the regular `weight` will be frozen.
|
374 |
-
"""
|
375 |
-
# argument validation
|
376 |
-
if _weight is not None:
|
377 |
-
assert (_weight.shape[0] == original_out_features) or (
|
378 |
-
original_out_features is None
|
379 |
-
), f"original_out_features={original_out_features} but _weight.shape[0]={_weight.shape[0]}"
|
380 |
-
assert (_weight.shape[1] == in_features) or (
|
381 |
-
in_features is None
|
382 |
-
), f"in_features={in_features} but _weight.shape[1]={_weight.shape[1]}"
|
383 |
-
in_features = _weight.shape[1]
|
384 |
-
original_out_features = _weight.shape[0]
|
385 |
-
else:
|
386 |
-
assert (
|
387 |
-
in_features is not None
|
388 |
-
), "in_features must be provided if _weight is not provided"
|
389 |
-
assert (
|
390 |
-
original_out_features is not None
|
391 |
-
), "original_out_features must be provided if _weight is not provided"
|
392 |
-
|
393 |
-
if _bias is not None:
|
394 |
-
assert bias is True, "bias must be True if _bias is provided"
|
395 |
-
|
396 |
-
# initialize original linear
|
397 |
-
super().__init__(
|
398 |
-
in_features,
|
399 |
-
original_out_features,
|
400 |
-
bias,
|
401 |
-
device,
|
402 |
-
dtype)
|
403 |
-
|
404 |
-
# set weight and bias manually
|
405 |
-
if _weight is not None:
|
406 |
-
self.weight = nn.Parameter(_weight)
|
407 |
-
if _bias is not None:
|
408 |
-
self.bias = nn.Parameter(_bias)
|
409 |
-
|
410 |
-
self.in_features = in_features
|
411 |
-
self.original_out_features = original_out_features
|
412 |
-
self.max_original_id = max_original_id
|
413 |
-
|
414 |
-
# initialize additional linear
|
415 |
-
self.additional_out_features = additional_out_features
|
416 |
-
self.has_bias = bias
|
417 |
-
if additional_out_features > 0:
|
418 |
-
self.additional_fc = nn.Linear(
|
419 |
-
in_features=in_features,
|
420 |
-
out_features=additional_out_features,
|
421 |
-
bias=self.has_bias,
|
422 |
-
device=device,
|
423 |
-
dtype=dtype,
|
424 |
-
)
|
425 |
-
self.set_requires_grad(
|
426 |
-
require_regular_grad=not partially_freeze, require_additional_grad=True
|
427 |
-
)
|
428 |
-
|
429 |
-
def set_requires_grad(self, require_regular_grad, require_additional_grad):
|
430 |
-
"""
|
431 |
-
Helper function to separately set the requires_grad flag for the regular weight and the additional weight.
|
432 |
-
"""
|
433 |
-
self.weight.requires_grad_(require_regular_grad)
|
434 |
-
if self.has_bias:
|
435 |
-
self.bias.requires_grad_(require_regular_grad)
|
436 |
-
self.additional_fc.requires_grad_(require_additional_grad)
|
437 |
-
|
438 |
-
def forward(self, input: torch.Tensor) -> torch.Tensor:
|
439 |
-
output = F.linear(input, self.weight, self.bias)
|
440 |
-
output = output[..., : self.max_original_id + 1]
|
441 |
-
|
442 |
-
if self.additional_out_features > 0:
|
443 |
-
additional_features = F.linear(
|
444 |
-
input, self.additional_fc.weight, self.additional_fc.bias
|
445 |
-
)
|
446 |
-
output = torch.cat((output, additional_features), -1)
|
447 |
-
return output
|
448 |
-
|
449 |
-
def extra_repr(self) -> str:
|
450 |
-
"""Overwriting `nn.Linear.extra_repr` to include new parameters."""
|
451 |
-
return "in_features={}, out_features={}, additional_out_features={}, bias={}, partially_freeze={}".format(
|
452 |
-
self.in_features,
|
453 |
-
self.max_original_id + 1,
|
454 |
-
self.additional_out_features,
|
455 |
-
self.bias is not None,
|
456 |
-
(not self.weight.requires_grad or not self.bias.requires_grad),
|
457 |
-
)
|
458 |
-
|
459 |
-
class VLM(nn.Module):
|
460 |
-
"""
|
461 |
-
Generic vision-language model (VLM) class.
|
462 |
-
A VLM consists of four components:
|
463 |
-
1. A vision encoder that extracts features from pixels, e.g. CLIP
|
464 |
-
input: (B, T_img, F, C, H, W)
|
465 |
-
output: (B, T_img, F, v, d)
|
466 |
-
2. A vision tokenizer that converts these features to visual token-like embeddings, e.g. Perceiver, or a linear projection head
|
467 |
-
input: (B, T_img, F, v, d)
|
468 |
-
output: (B, T_img, n, d)
|
469 |
-
3. A fusion method that allows the language model to attend to these tokens, e.g. cross-attention, or placing the tokens directly in the language model's input sequence
|
470 |
-
4. A language model
|
471 |
-
"""
|
472 |
-
|
473 |
-
def __init__(
|
474 |
-
self,
|
475 |
-
vision_encoder: nn.Module,
|
476 |
-
vision_tokenizer: nn.Module,
|
477 |
-
lang_model: nn.Module,
|
478 |
-
initial_tokenizer_len: int,
|
479 |
-
pad_token_id: int,
|
480 |
-
gradient_checkpointing: bool = False,
|
481 |
-
):
|
482 |
-
"""
|
483 |
-
Args:
|
484 |
-
vision_encoder (nn.Module): e.g. CLIP
|
485 |
-
vision_tokenizer (nn.Module): e.g. PerceiverResampler
|
486 |
-
lang_model (nn.Module): e.g. MPT
|
487 |
-
initial_tokenizer_len (int): size of the original tokenizer vocab
|
488 |
-
pad_token_id (int): id of the pad token
|
489 |
-
gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
|
490 |
-
"""
|
491 |
-
super().__init__()
|
492 |
-
|
493 |
-
# save dimension information
|
494 |
-
self.lang_embedding_dim = lang_model.get_input_embeddings().weight.shape[1]
|
495 |
-
if hasattr(lang_model.config, "d_model"):
|
496 |
-
self.lang_hidden_dim = lang_model.config.d_model # mpt uses d_model
|
497 |
-
else:
|
498 |
-
self.lang_hidden_dim = lang_model.config.hidden_size
|
499 |
-
self.vis_embedding_dim = vision_tokenizer.dim_media
|
500 |
-
self.num_tokens_per_vis = vision_tokenizer.num_tokens_per_media
|
501 |
-
|
502 |
-
# core components
|
503 |
-
self.vision_encoder = vision_encoder
|
504 |
-
self.vision_tokenizer = vision_tokenizer
|
505 |
-
self.lang_model = lang_model
|
506 |
-
|
507 |
-
# lm embeddings
|
508 |
-
self.pad_token_id = pad_token_id
|
509 |
-
self.initial_tokenizer_len = initial_tokenizer_len
|
510 |
-
input_embeds = DecoupledEmbedding(
|
511 |
-
max_original_id=initial_tokenizer_len - 1,
|
512 |
-
num_additional_embeddings=len(self.special_tokens),
|
513 |
-
_weight=self.lang_model.get_input_embeddings().weight,
|
514 |
-
pad_token_id=self.pad_token_id,
|
515 |
-
)
|
516 |
-
if hasattr(input_embeds, "additional_embedding"):
|
517 |
-
input_embeds.additional_embedding.weight.data.normal_(
|
518 |
-
mean=0.0,
|
519 |
-
std=self.lang_model.config.initializer_range
|
520 |
-
if hasattr(self.lang_model.config, "initializer_range")
|
521 |
-
else 0.02,
|
522 |
-
)
|
523 |
-
self.lang_model.set_input_embeddings(input_embeds)
|
524 |
-
|
525 |
-
out_embeds = DecoupledLinear(
|
526 |
-
max_original_id=initial_tokenizer_len - 1,
|
527 |
-
additional_out_features=len(self.special_tokens),
|
528 |
-
_weight=self.lang_model.get_output_embeddings().weight,
|
529 |
-
_bias=self.lang_model.get_output_embeddings().bias if hasattr(self.lang_model.get_output_embeddings(), "bias") else None,
|
530 |
-
)
|
531 |
-
if hasattr(out_embeds, "additional_fc"):
|
532 |
-
out_embeds.additional_fc.weight.data.normal_(
|
533 |
-
mean=0.0,
|
534 |
-
std=self.lang_model.config.initializer_range
|
535 |
-
if hasattr(self.lang_model.config, "initializer_range")
|
536 |
-
else 0.02,
|
537 |
-
)
|
538 |
-
self.lang_model.set_output_embeddings(out_embeds)
|
539 |
-
|
540 |
-
# gradient checkpointing
|
541 |
-
self.vision_tokenizer._use_gradient_checkpointing = gradient_checkpointing
|
542 |
-
|
543 |
-
def forward(
|
544 |
-
self,
|
545 |
-
vision_x: Optional[torch.Tensor],
|
546 |
-
lang_x: torch.Tensor,
|
547 |
-
attention_mask: Optional[torch.Tensor] = None,
|
548 |
-
labels: Optional[torch.Tensor] = None,
|
549 |
-
past_key_values: Optional[
|
550 |
-
List[Union[torch.Tensor, Tuple[torch.Tensor]]]
|
551 |
-
] = None,
|
552 |
-
past_media_locations: Optional[torch.Tensor] = None,
|
553 |
-
past_vision_tokens: Optional[torch.Tensor] = None,
|
554 |
-
use_cache: Optional[bool] = False,
|
555 |
-
**kwargs,
|
556 |
-
):
|
557 |
-
"""
|
558 |
-
Args:
|
559 |
-
vision_x: Vision input
|
560 |
-
shape (B, T_img, F, C, H, W) with F=1
|
561 |
-
only F = 1 is supported (single-frame videos)
|
562 |
-
if T_img > the number of media tokens in the corresponding input_ids (lang_x),
|
563 |
-
only the first number of media tokens in lang_x are used
|
564 |
-
lang_x: Language input ids, with media tokens denoting where
|
565 |
-
visual media should be inserted.
|
566 |
-
shape (B, T_txt)
|
567 |
-
attention_mask: Attention mask. Defaults to None.
|
568 |
-
labels: Labels. Defaults to None.
|
569 |
-
shape (B, T_txt)
|
570 |
-
past_key_values (Tuple[torch.Tensor]], optional): Past key value pairs for each of the T_txt previous tokens in the language model. Defaults to None.
|
571 |
-
list of length = number of decoder layers in the LM
|
572 |
-
exact implementation depends on LM, see Hugging Face docs
|
573 |
-
past_media_locations (torch.Tensor, optional): boolean mask denoting which of the previous T_txt tokens were media tokens. Defaults to None.
|
574 |
-
shape (B, T_txt)
|
575 |
-
past_vision_tokens (torch.Tensor, optional): Previous vision tokens. Defaults to None.
|
576 |
-
use_cache (Optional[bool], optional): Whether to use cache. Defaults to False.
|
577 |
-
If True, includes key_values, media_locations, and vision_tokens in the output.
|
578 |
-
"""
|
579 |
-
assert not (past_vision_tokens is None) ^ (
|
580 |
-
past_media_locations is None
|
581 |
-
), "past_vision_tokens and past_media_locations must both be None or both be not None"
|
582 |
-
|
583 |
-
# convert pixels to vision tokens
|
584 |
-
if vision_x is not None:
|
585 |
-
vision_features = self._encode_vision_x(vision_x=vision_x)
|
586 |
-
vision_tokens = self.vision_tokenizer(vision_features)
|
587 |
-
else:
|
588 |
-
vision_tokens = None
|
589 |
-
|
590 |
-
# fuse the vision and language tokens
|
591 |
-
new_inputs = self._prepare_inputs_for_forward(
|
592 |
-
vision_tokens=vision_tokens,
|
593 |
-
lang_x=lang_x,
|
594 |
-
attention_mask=attention_mask,
|
595 |
-
labels=labels,
|
596 |
-
past_key_values=past_key_values,
|
597 |
-
past_media_locations=past_media_locations,
|
598 |
-
padding_side="right",
|
599 |
-
past_vision_tokens=past_vision_tokens,
|
600 |
-
)
|
601 |
-
output = self.lang_model(
|
602 |
-
**new_inputs,
|
603 |
-
use_cache=use_cache,
|
604 |
-
past_key_values=past_key_values,
|
605 |
-
**kwargs,
|
606 |
-
)
|
607 |
-
|
608 |
-
# postprocessing may be needed, e.g. to remove extra tokens from logits that were inserted into the language stream
|
609 |
-
# or to add the past_vision_tokens and past_media_locations to the output
|
610 |
-
output = self._postprocess_outputs_from_forward(
|
611 |
-
output=output,
|
612 |
-
lang_x=lang_x,
|
613 |
-
vision_tokens=vision_tokens,
|
614 |
-
use_cache=use_cache,
|
615 |
-
past_vision_tokens=past_vision_tokens,
|
616 |
-
past_media_locations=past_media_locations,
|
617 |
-
)
|
618 |
-
|
619 |
-
# postforward hooks
|
620 |
-
self._post_forward_hook()
|
621 |
-
return output
|
622 |
-
|
623 |
-
def _encode_vision_x_anyres(self, samples, device):
|
624 |
-
assert self.anyres_grids is not None
|
625 |
-
image_raw = samples["image"] # list of patch list in of shape [1, N_patch, C, H, W]
|
626 |
-
image_sizes = samples["image_size"]
|
627 |
-
|
628 |
-
# Image_raw can be a list of list of patches, when a `samples` has multiple images.
|
629 |
-
if isinstance(image_raw[0], list):
|
630 |
-
images = [x.squeeze(0) for sample_img in image_raw for x in sample_img]
|
631 |
-
image_sizes = [s for sample_sizes in image_sizes for s in sample_sizes]
|
632 |
-
else:
|
633 |
-
# assert isinstance(image_raw[0], torch.Tensor), f"Unkown image type: {image_raw[0]}"
|
634 |
-
# concate list of patches into one big patch for any res encoding.
|
635 |
-
images = [x.squeeze(0) for x in image_raw] # [N_patch, C, H, W]
|
636 |
-
image = torch.cat(images, dim=0) # [\sum{B}{N_patch_i}, C, H, W]
|
637 |
-
image = image.to(device)
|
638 |
-
|
639 |
-
with torch.no_grad():
|
640 |
-
if self.vision_encoder.__class__.__name__ == "TimmModel":
|
641 |
-
image_embeds = self.vision_encoder.trunk.forward_features(image)
|
642 |
-
elif self.vision_encoder.__class__.__name__ in ['CLIPVisionModel', 'SiglipVisionTransformer']:
|
643 |
-
image_embeds = self.vision_encoder(image).last_hidden_state
|
644 |
-
else:
|
645 |
-
image_embeds = self.vision_encoder(image)[1] # OpenCLIP returns tuples
|
646 |
-
|
647 |
-
if isinstance(self.vision_encoder, CLIPVisionModel) or isinstance(self.vision_encoder, SiglipVisionTransformer):
|
648 |
-
base_img_size = self.vision_encoder.config.image_size
|
649 |
-
else:
|
650 |
-
base_img_size = self.vision_encoder.image_size[0]
|
651 |
-
|
652 |
-
if self.vision_encoder.__class__.__name__ == "TimmModel":
|
653 |
-
grid_size = self.vision_encoder.trunk.patch_embed.grid_size
|
654 |
-
elif self.vision_encoder.__class__.__name__ in ['CLIPVisionModel', 'SiglipVisionTransformer']:
|
655 |
-
grid_size_base = self.vision_encoder.config.image_size // self.vision_encoder.config.patch_size
|
656 |
-
grid_size = (grid_size_base, grid_size_base)
|
657 |
-
else:
|
658 |
-
grid_size = self.vision_encoder.grid_size
|
659 |
-
height, width = grid_size
|
660 |
-
|
661 |
-
if not image_embeds.shape[1] == height * width:
|
662 |
-
assert image_embeds.shape[1] == height * width + 1 # For vision encoders that has [CLS] token.
|
663 |
-
image_embeds = image_embeds[:, 1:, :] # Drop the cls token for each patch.
|
664 |
-
n_vis_token_per_patch = image_embeds.shape[1]
|
665 |
-
|
666 |
-
# Split encoded patches and merge patch features
|
667 |
-
# 1. Get the raw sizes from samples, and split the image embeds [\sum_{B}(N_patch_i), N_tok(16*16), C]
|
668 |
-
split_sizes = [image.shape[0] for image in images]
|
669 |
-
image_embeds = torch.split(image_embeds, split_sizes, dim=0)
|
670 |
-
# 2. For each image (consist of a list of patches), merge the patches spatially (of shape [C, n_patch_height, n_patch_width])
|
671 |
-
new_image_embeds = []
|
672 |
-
patch_attn_masks = []
|
673 |
-
max_n_img_token = -1
|
674 |
-
for idx, patch_embeds in enumerate(image_embeds):
|
675 |
-
if patch_embeds.shape[0] > 1:
|
676 |
-
# 3. Flatten the patch features and get [C, n_patch_height * (n_patch_width+1)]
|
677 |
-
base_patch_embeds = patch_embeds[0] # TODO: prepend the CLS token for th base patch embeds (of the resized entire image).
|
678 |
-
patch_embeds = patch_embeds[1:]
|
679 |
-
|
680 |
-
assert height * width == base_patch_embeds.shape[0]
|
681 |
-
|
682 |
-
num_patch_width, num_patch_height = get_anyres_image_grid_shape(image_sizes[idx],
|
683 |
-
self.anyres_grids,
|
684 |
-
base_img_size) # Hardcoded grid_pinpoints.
|
685 |
-
patch_embeds = patch_embeds.view(num_patch_height, num_patch_width, height, width, -1)
|
686 |
-
|
687 |
-
patch_embeds = patch_embeds.permute(4, 0, 2, 1, 3).contiguous()
|
688 |
-
patch_embeds = patch_embeds.flatten(1, 2).flatten(2, 3)
|
689 |
-
patch_embeds, patch_attn_mask = unpad_image(patch_embeds, image_sizes[idx], self.anyres_patch_sampling)
|
690 |
-
if hasattr(self, 'image_newline'):
|
691 |
-
patch_embeds = torch.cat((
|
692 |
-
patch_embeds,
|
693 |
-
self.image_newline[:, None, None].expand(*patch_embeds.shape[:-1], 1)
|
694 |
-
), dim=-1)
|
695 |
-
if self.anyres_patch_sampling:
|
696 |
-
patch_embeds = patch_embeds.view(-1, num_patch_height, num_patch_width, height*width)
|
697 |
-
patch_embeds = patch_embeds.flatten(1, 2).permute(1, 2, 0)
|
698 |
-
assert patch_attn_mask is not None
|
699 |
-
patch_attn_mask = patch_attn_mask.view(num_patch_height, num_patch_width, height*width)
|
700 |
-
patch_attn_mask = patch_attn_mask.flatten(0, 1)
|
701 |
-
patch_embeds = torch.cat((base_patch_embeds.unsqueeze(0), patch_embeds), dim=0)
|
702 |
-
patch_attn_mask = torch.cat((torch.ones(n_vis_token_per_patch, device=patch_embeds.device).unsqueeze(0), patch_attn_mask), dim=0)
|
703 |
-
else:
|
704 |
-
patch_embeds = patch_embeds.flatten(1, 2).transpose(0, 1)
|
705 |
-
patch_embeds = torch.cat((base_patch_embeds, patch_embeds), dim=0)
|
706 |
-
else:
|
707 |
-
patch_embeds = patch_embeds[0].unsqueeze(0) if self.anyres_patch_sampling else patch_embeds[0]
|
708 |
-
patch_attn_mask = torch.ones(n_vis_token_per_patch, device=patch_embeds.device).unsqueeze(0) if self.anyres_patch_sampling else None
|
709 |
-
if hasattr(self, 'image_newline'):
|
710 |
-
patch_embeds = torch.cat((
|
711 |
-
patch_embeds,
|
712 |
-
self.image_newline[None]
|
713 |
-
), dim=0)
|
714 |
-
if not self.anyres_patch_sampling:
|
715 |
-
max_n_img_token = max(patch_embeds.shape[0], max_n_img_token)
|
716 |
-
|
717 |
-
new_image_embeds.append(patch_embeds)
|
718 |
-
patch_attn_masks.append(patch_attn_mask)
|
719 |
-
|
720 |
-
if self.anyres_patch_sampling:
|
721 |
-
# Return individual patches for independent token downsampling.
|
722 |
-
return new_image_embeds, patch_attn_masks
|
723 |
-
|
724 |
-
# 4. Pad and concat the list of image_embeds [N_tok_i, C] together into a batch. Also modify the query attention mask.
|
725 |
-
image_embeds = []
|
726 |
-
image_atts = []
|
727 |
-
for image_embed in new_image_embeds:
|
728 |
-
n_img_token = image_embed.shape[0]
|
729 |
-
img_attn = torch.ones((max_n_img_token), dtype=torch.long, device=image_embed.device)
|
730 |
-
if n_img_token < max_n_img_token:
|
731 |
-
padded_embed = torch.zeros((max_n_img_token, image_embed.shape[-1]), dtype=image_embed.dtype, device=image_embed.device)
|
732 |
-
padded_embed[:n_img_token, :] = image_embed
|
733 |
-
img_attn[n_img_token:] = 0 # Mask out the padded entries.
|
734 |
-
else:
|
735 |
-
padded_embed = image_embed
|
736 |
-
image_embeds.append(padded_embed)
|
737 |
-
image_atts.append(img_attn)
|
738 |
-
image_embeds = torch.stack(image_embeds, dim=0) # Shape [B, N_tok_longest, C_dim]
|
739 |
-
image_atts = torch.stack(image_atts, dim=0) # Shape [B, N_tok_longest, C_dim]
|
740 |
-
# TODO: reshape image_embeds and image_atts to "b T F v d"
|
741 |
-
image_embeds = image_embeds[:, None, None, :, :]
|
742 |
-
# image_atts = image_atts[:, None, None, :, :]
|
743 |
-
|
744 |
-
return image_embeds, image_atts
|
745 |
-
|
746 |
-
def _encode_vision_x(self, vision_x: torch.Tensor):
|
747 |
-
"""
|
748 |
-
Compute media tokens from vision input by passing it through vision encoder and conditioning language model.
|
749 |
-
Args:
|
750 |
-
vision_x: Vision input
|
751 |
-
shape (B, T_img, F, C, H, W)
|
752 |
-
Images in the same chunk are collated along T_img, and frames are collated along F
|
753 |
-
Currently only F=1 is supported (single-frame videos)
|
754 |
-
|
755 |
-
rearrange code based on https://github.com/dhansmair/flamingo-mini
|
756 |
-
"""
|
757 |
-
assert vision_x.ndim == 6, "vision_x should be of shape (b, T_img, F, C, H, W)"
|
758 |
-
b, T, F = vision_x.shape[:3]
|
759 |
-
|
760 |
-
vision_x = rearrange(vision_x, "b T F c h w -> (b T F) c h w")
|
761 |
-
with torch.no_grad():
|
762 |
-
if self.vision_encoder.__class__.__name__ == "TimmModel":
|
763 |
-
vision_x = self.vision_encoder.trunk.forward_features(vision_x)
|
764 |
-
elif self.vision_encoder.__class__.__name__ in ['CLIPVisionModel', 'SiglipVisionTransformer']:
|
765 |
-
vision_x = self.vision_encoder(vision_x).last_hidden_state
|
766 |
-
else:
|
767 |
-
vision_x = self.vision_encoder(vision_x)[1] # OpenCLIP returns tuples
|
768 |
-
vision_x = rearrange(vision_x, "(b T F) v d -> b T F v d", b=b, T=T, F=F)
|
769 |
-
return vision_x
|
770 |
-
|
771 |
-
def _concat_vision_cache(
|
772 |
-
self, lang_x, vision_tokens, past_vision_tokens, past_media_locations, use_cache
|
773 |
-
):
|
774 |
-
"""
|
775 |
-
Helper function to include the past vision tokens and past media locations in the output.
|
776 |
-
"""
|
777 |
-
if use_cache:
|
778 |
-
if past_media_locations is not None and past_vision_tokens is not None:
|
779 |
-
if vision_tokens is not None:
|
780 |
-
updated_vision_tokens = torch.cat(
|
781 |
-
[
|
782 |
-
past_vision_tokens,
|
783 |
-
vision_tokens,
|
784 |
-
],
|
785 |
-
dim=1,
|
786 |
-
)
|
787 |
-
else:
|
788 |
-
updated_vision_tokens = past_vision_tokens
|
789 |
-
updated_media_locations = torch.cat(
|
790 |
-
[
|
791 |
-
past_media_locations,
|
792 |
-
lang_x == self.media_token_id,
|
793 |
-
],
|
794 |
-
dim=1,
|
795 |
-
)
|
796 |
-
else:
|
797 |
-
updated_vision_tokens = vision_tokens
|
798 |
-
updated_media_locations = lang_x == self.media_token_id
|
799 |
-
|
800 |
-
else:
|
801 |
-
updated_vision_tokens = None
|
802 |
-
updated_media_locations = None
|
803 |
-
|
804 |
-
return updated_vision_tokens, updated_media_locations
|
805 |
-
|
806 |
-
def generate(
|
807 |
-
self,
|
808 |
-
vision_x: torch.Tensor,
|
809 |
-
lang_x: torch.Tensor,
|
810 |
-
attention_mask: torch.Tensor = None,
|
811 |
-
past_key_values: Optional[
|
812 |
-
List[Union[torch.Tensor, Tuple[torch.Tensor]]]
|
813 |
-
] = None,
|
814 |
-
past_media_locations: Optional[torch.Tensor] = None,
|
815 |
-
past_vision_tokens: Optional[torch.Tensor] = None,
|
816 |
-
**kwargs,
|
817 |
-
):
|
818 |
-
"""
|
819 |
-
Generate text conditioned on vision and language inputs.
|
820 |
-
Args:
|
821 |
-
vision_x (torch.Tensor): Vision input
|
822 |
-
shape (B, T_img, F, C, H, W)
|
823 |
-
see documentation for forward
|
824 |
-
lang_x (torch.Tensor): Language input
|
825 |
-
shape (B, T_txt)
|
826 |
-
attention_mask (torch.Tensor, optional): Attention mask. Defaults to None.
|
827 |
-
**kwargs: see generate documentation in Hugging Face CausalLM models.
|
828 |
-
Returns:
|
829 |
-
torch.Tensor: lang_x with generated tokens appended to it
|
830 |
-
"""
|
831 |
-
num_beams = kwargs.pop("num_beams", 1)
|
832 |
-
|
833 |
-
# convert pixels to vision tokens
|
834 |
-
if vision_x is not None:
|
835 |
-
vision_features = self._encode_vision_x(vision_x=vision_x)
|
836 |
-
vision_tokens = self.vision_tokenizer(vision_features)
|
837 |
-
else:
|
838 |
-
vision_tokens = None
|
839 |
-
|
840 |
-
# fuse the vision and language tokens
|
841 |
-
# for xattn, vision_x and media_location are repeat_interleaved s.t.
|
842 |
-
# the total batch size is B * num_beams
|
843 |
-
new_inputs = self._prepare_inputs_for_forward(
|
844 |
-
vision_tokens=vision_tokens,
|
845 |
-
lang_x=lang_x,
|
846 |
-
attention_mask=attention_mask,
|
847 |
-
past_key_values=past_key_values,
|
848 |
-
past_media_locations=past_media_locations,
|
849 |
-
past_vision_tokens=past_vision_tokens,
|
850 |
-
padding_side="left",
|
851 |
-
num_beams=num_beams,
|
852 |
-
)
|
853 |
-
output = self.lang_model.generate(
|
854 |
-
**new_inputs,
|
855 |
-
past_key_values=past_key_values,
|
856 |
-
num_beams=num_beams,
|
857 |
-
use_cache=True,
|
858 |
-
**kwargs,
|
859 |
-
)
|
860 |
-
self._post_forward_hook()
|
861 |
-
return output
|
862 |
-
|
863 |
-
@property
|
864 |
-
def num_trainable_params(self):
|
865 |
-
"""Print the number of trainable parameters"""
|
866 |
-
return num_params(self, filter_to_trainable=True)
|
867 |
-
|
868 |
-
def set_trainable(self):
|
869 |
-
"""
|
870 |
-
Freeze appropriate parameters in the model.
|
871 |
-
"""
|
872 |
-
raise NotImplementedError
|
873 |
-
|
874 |
-
def group_params_by_weight_decay(self):
|
875 |
-
"""
|
876 |
-
Return a tuple of (params to optimize w/ weight decay, params to optimize w/o weight decay)
|
877 |
-
"""
|
878 |
-
params_with_wd, params_without_wd = [], []
|
879 |
-
for n, p in self.named_parameters():
|
880 |
-
if p.requires_grad:
|
881 |
-
if self._should_apply_weight_decay(n):
|
882 |
-
params_with_wd.append(p)
|
883 |
-
else:
|
884 |
-
params_without_wd.append(p)
|
885 |
-
return params_with_wd, params_without_wd
|
886 |
-
|
887 |
-
def _should_apply_weight_decay(self, parameter_name):
|
888 |
-
"""
|
889 |
-
Return whether weight decay should be applied to a parameter.
|
890 |
-
"""
|
891 |
-
raise NotImplementedError
|
892 |
-
|
893 |
-
@property
|
894 |
-
def special_tokens(self):
|
895 |
-
"""
|
896 |
-
Returns a dict mapping from the attribute name of a special token to its string format,
|
897 |
-
e.g. "media_token": "<image>"
|
898 |
-
"""
|
899 |
-
assert (
|
900 |
-
"media_token" in self._special_tokens
|
901 |
-
), "VLMs need to request that the tokenizer add a media_token and call set_special_token_ids to set self.media_token_id"
|
902 |
-
return self._special_tokens
|
903 |
-
|
904 |
-
@property
|
905 |
-
def special_token_ids(self):
|
906 |
-
"""
|
907 |
-
Returns a list of the special token ids
|
908 |
-
"""
|
909 |
-
return [getattr(self, f"{att_name}_id") for att_name in self.special_tokens]
|
910 |
-
|
911 |
-
def set_special_token_ids(self, string_to_ids):
|
912 |
-
"""
|
913 |
-
Args:
|
914 |
-
string_to_ids (dict): mapping from token string to id
|
915 |
-
"""
|
916 |
-
assert set(self.special_tokens.values()).issubset(set(string_to_ids.keys()))
|
917 |
-
for att_name, token_str in self.special_tokens.items():
|
918 |
-
token_id = string_to_ids[token_str]
|
919 |
-
setattr(self, f"{att_name}_id", token_id)
|
920 |
-
setattr(self.lang_model, f"{att_name}_id", token_id)
|
921 |
-
|
922 |
-
def init_gradient_checkpointing(self):
|
923 |
-
from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
|
924 |
-
checkpoint_wrapper,
|
925 |
-
CheckpointWrapper,
|
926 |
-
CheckpointImpl,
|
927 |
-
apply_activation_checkpointing,
|
928 |
-
)
|
929 |
-
from functools import partial
|
930 |
-
|
931 |
-
non_reentrant_wrapper = partial(
|
932 |
-
checkpoint_wrapper,
|
933 |
-
checkpoint_impl=CheckpointImpl.NO_REENTRANT,
|
934 |
-
)
|
935 |
-
apply_activation_checkpointing(
|
936 |
-
self,
|
937 |
-
checkpoint_wrapper_fn=non_reentrant_wrapper,
|
938 |
-
check_fn=lambda m: getattr(m, "_use_gradient_checkpointing", False)
|
939 |
-
and not isinstance(m, CheckpointWrapper),
|
940 |
-
)
|
941 |
-
|
942 |
-
@dataclass
|
943 |
-
class VLMOutputWithPast(CausalLMOutputWithPast):
|
944 |
-
"""
|
945 |
-
VLMOutputWithPast is a wrapper around CausalLMOutputWithPast that adds the following attributes:
|
946 |
-
past_media_locations: Optional[torch.Tensor] = None,
|
947 |
-
past_vision_tokens: Optional[torch.Tensor] = None,
|
948 |
-
"""
|
949 |
-
|
950 |
-
past_media_locations: Optional[torch.Tensor] = None
|
951 |
-
past_vision_tokens: Optional[torch.Tensor] = None
|
952 |
-
|
953 |
-
|
954 |
-
def exists(val):
|
955 |
-
return val is not None
|
956 |
-
|
957 |
-
|
958 |
-
def FeedForward(dim, mult=4):
|
959 |
-
inner_dim = int(dim * mult)
|
960 |
-
return nn.Sequential(
|
961 |
-
nn.LayerNorm(dim),
|
962 |
-
nn.Linear(dim, inner_dim, bias=False),
|
963 |
-
nn.GELU(),
|
964 |
-
nn.Linear(inner_dim, dim, bias=False),
|
965 |
-
)
|
966 |
-
|
967 |
-
class VLMWithLanguageStream(VLM):
|
968 |
-
"""
|
969 |
-
VLM that fuses modalities by inserting vision tokens directly into the language stream.
|
970 |
-
"""
|
971 |
-
|
972 |
-
def __init__(
|
973 |
-
self,
|
974 |
-
vision_encoder: nn.Module,
|
975 |
-
vision_tokenizer: nn.Module,
|
976 |
-
lang_model: nn.Module,
|
977 |
-
initial_tokenizer_len: int,
|
978 |
-
pad_token_id: int,
|
979 |
-
decoder_layers_attr_name: str = None,
|
980 |
-
gradient_checkpointing: bool = False,
|
981 |
-
):
|
982 |
-
super().__init__(
|
983 |
-
vision_encoder=vision_encoder,
|
984 |
-
vision_tokenizer=vision_tokenizer,
|
985 |
-
lang_model=lang_model,
|
986 |
-
initial_tokenizer_len=initial_tokenizer_len,
|
987 |
-
pad_token_id=pad_token_id,
|
988 |
-
gradient_checkpointing=gradient_checkpointing,
|
989 |
-
)
|
990 |
-
self.decoder_layers_attr_name = decoder_layers_attr_name
|
991 |
-
if decoder_layers_attr_name is not None:
|
992 |
-
for block in getattr_recursive(self.lang_model, self.decoder_layers_attr_name):
|
993 |
-
block._use_gradient_checkpointing = gradient_checkpointing
|
994 |
-
|
995 |
-
def _prepare_inputs_for_forward(
|
996 |
-
self,
|
997 |
-
vision_tokens: torch.Tensor,
|
998 |
-
lang_x: torch.Tensor,
|
999 |
-
attention_mask: torch.Tensor,
|
1000 |
-
labels: torch.Tensor = None,
|
1001 |
-
past_key_values=None,
|
1002 |
-
vision_attention_mask: Optional[torch.Tensor] = None,
|
1003 |
-
past_media_locations: torch.Tensor = None,
|
1004 |
-
past_vision_tokens: torch.Tensor = None,
|
1005 |
-
padding_side: str = "left",
|
1006 |
-
num_beams: int = 1,
|
1007 |
-
):
|
1008 |
-
"""
|
1009 |
-
Insert the vision tokens directly into the language stream/
|
1010 |
-
This requires us to modify the input_ids, attention_mask, and labels.
|
1011 |
-
"""
|
1012 |
-
if past_key_values is not None:
|
1013 |
-
past_len = past_key_values[0][0].shape[2]
|
1014 |
-
assert attention_mask.shape[1] == past_len + lang_x.shape[1], (
|
1015 |
-
"Attention_mask must be as long as the entire past len (including image tokens) and current input IDs. "
|
1016 |
-
+ "Check that you've expanded the attention mask to account for past image tokens."
|
1017 |
-
)
|
1018 |
-
|
1019 |
-
if vision_tokens is None:
|
1020 |
-
return {
|
1021 |
-
"input_ids": lang_x,
|
1022 |
-
"attention_mask": attention_mask,
|
1023 |
-
"labels": labels,
|
1024 |
-
}
|
1025 |
-
|
1026 |
-
# get the language embeddings
|
1027 |
-
lang_embeds = self.lang_model.get_input_embeddings()(lang_x)
|
1028 |
-
|
1029 |
-
# build up the multimodal embeddings
|
1030 |
-
B = lang_x.shape[0]
|
1031 |
-
has_labels = labels is not None
|
1032 |
-
multimodal_embeds = []
|
1033 |
-
multimodal_attention_mask = []
|
1034 |
-
multimodal_labels = [] if has_labels else None
|
1035 |
-
for i in range(B):
|
1036 |
-
# get index of <image> tokens in lang_x[i]
|
1037 |
-
image_token_idxs = torch.where(lang_x[i] == self.media_token_id)[0]
|
1038 |
-
|
1039 |
-
if len(image_token_idxs) == 0:
|
1040 |
-
multimodal_embeds.append(lang_embeds[i].clone())
|
1041 |
-
multimodal_attention_mask.append(attention_mask[i].clone())
|
1042 |
-
if has_labels:
|
1043 |
-
multimodal_labels.append(labels[i].clone())
|
1044 |
-
continue
|
1045 |
-
|
1046 |
-
# loop through the image_token_idxs and insert the vision tokens
|
1047 |
-
new_embed = lang_embeds[i].clone()
|
1048 |
-
new_attention_mask = (
|
1049 |
-
attention_mask[i].clone() if attention_mask is not None else None
|
1050 |
-
)
|
1051 |
-
if has_labels:
|
1052 |
-
new_label = labels[i].clone()
|
1053 |
-
|
1054 |
-
for img_num, img_idx in enumerate(image_token_idxs):
|
1055 |
-
# Get vision token attention mask for padded llava-style any resolution image tokens.
|
1056 |
-
if self.image_aspect_ratio =='anyres':
|
1057 |
-
num_vis_tokens = vision_tokens[i][img_num].shape[0]
|
1058 |
-
if vision_attention_mask is not None:
|
1059 |
-
vis_attention_mask = vision_attention_mask[i]
|
1060 |
-
else:
|
1061 |
-
vis_attention_mask = torch.ones(
|
1062 |
-
num_vis_tokens, dtype=torch.long
|
1063 |
-
).to(attention_mask.device)
|
1064 |
-
else:
|
1065 |
-
assert (
|
1066 |
-
vision_tokens[i][img_num].shape[0] == self.num_tokens_per_vis
|
1067 |
-
), f"vision token number mismatch: image embedding ({vision_tokens[i][img_num].shape[0]}) \
|
1068 |
-
vs. model.num_tokens_per_vis ({self.num_tokens_per_vis})"
|
1069 |
-
# By default, vision tokens are not padded.
|
1070 |
-
num_vis_tokens = self.num_tokens_per_vis
|
1071 |
-
vis_attention_mask = torch.ones(
|
1072 |
-
num_vis_tokens, dtype=torch.long
|
1073 |
-
).to(attention_mask.device)
|
1074 |
-
|
1075 |
-
new_embed = torch.cat(
|
1076 |
-
(
|
1077 |
-
new_embed[:img_idx],
|
1078 |
-
vision_tokens[i][img_num],
|
1079 |
-
new_embed[img_idx + 1 :],
|
1080 |
-
),
|
1081 |
-
dim=0,
|
1082 |
-
)
|
1083 |
-
new_attention_mask = torch.cat(
|
1084 |
-
(
|
1085 |
-
new_attention_mask[:img_idx],
|
1086 |
-
vis_attention_mask,
|
1087 |
-
new_attention_mask[img_idx + 1 :],
|
1088 |
-
),
|
1089 |
-
dim=0,
|
1090 |
-
)
|
1091 |
-
if has_labels:
|
1092 |
-
new_label = torch.cat(
|
1093 |
-
(
|
1094 |
-
new_label[:img_idx],
|
1095 |
-
torch.ones(num_vis_tokens, dtype=torch.long).to(
|
1096 |
-
labels.device
|
1097 |
-
)
|
1098 |
-
* -100,
|
1099 |
-
new_label[img_idx + 1 :],
|
1100 |
-
),
|
1101 |
-
dim=0,
|
1102 |
-
)
|
1103 |
-
multimodal_embeds.append(new_embed)
|
1104 |
-
multimodal_attention_mask.append(new_attention_mask)
|
1105 |
-
if has_labels:
|
1106 |
-
multimodal_labels.append(new_label)
|
1107 |
-
|
1108 |
-
# stack
|
1109 |
-
multimodal_embeds = stack_with_padding(
|
1110 |
-
multimodal_embeds,
|
1111 |
-
padding_value=self.pad_token_id,
|
1112 |
-
padding_side=padding_side,
|
1113 |
-
)
|
1114 |
-
multimodal_attention_mask = stack_with_padding(
|
1115 |
-
multimodal_attention_mask,
|
1116 |
-
padding_value=0,
|
1117 |
-
padding_side=padding_side,
|
1118 |
-
)
|
1119 |
-
if has_labels:
|
1120 |
-
multimodal_labels = stack_with_padding(
|
1121 |
-
multimodal_labels,
|
1122 |
-
padding_value=-100,
|
1123 |
-
padding_side=padding_side,
|
1124 |
-
)
|
1125 |
-
|
1126 |
-
return {
|
1127 |
-
"inputs_embeds": multimodal_embeds,
|
1128 |
-
"attention_mask": multimodal_attention_mask,
|
1129 |
-
"labels": multimodal_labels,
|
1130 |
-
}
|
1131 |
-
|
1132 |
-
def _postprocess_outputs_from_forward(
|
1133 |
-
self,
|
1134 |
-
output: CausalLMOutputWithPast,
|
1135 |
-
lang_x: torch.Tensor,
|
1136 |
-
vision_tokens: torch.Tensor,
|
1137 |
-
past_vision_tokens: torch.Tensor,
|
1138 |
-
past_media_locations: torch.Tensor,
|
1139 |
-
use_cache: bool = False,
|
1140 |
-
):
|
1141 |
-
# Include the past vision tokens and past media locations in the output
|
1142 |
-
updated_vision_tokens, updated_media_locations = self._concat_vision_cache(
|
1143 |
-
lang_x=lang_x,
|
1144 |
-
vision_tokens=vision_tokens,
|
1145 |
-
past_vision_tokens=past_vision_tokens,
|
1146 |
-
past_media_locations=past_media_locations,
|
1147 |
-
use_cache=use_cache,
|
1148 |
-
)
|
1149 |
-
|
1150 |
-
# return logits that are the same shape as the original input_ids
|
1151 |
-
logits = output.logits
|
1152 |
-
batch_logits = []
|
1153 |
-
B, T_txt = lang_x.shape
|
1154 |
-
for i in range(B):
|
1155 |
-
sequence_logits = []
|
1156 |
-
logits_j = 0
|
1157 |
-
for j in range(T_txt):
|
1158 |
-
if lang_x[i, j] != self.media_token_id:
|
1159 |
-
sequence_logits.append(logits[i, logits_j])
|
1160 |
-
logits_j += 1
|
1161 |
-
else:
|
1162 |
-
# append the logit for the first image token, then skip over the rest
|
1163 |
-
# note: the model actually learns to predict <im_patch>, not <image>
|
1164 |
-
sequence_logits.append(logits[i, logits_j])
|
1165 |
-
logits_j += self.num_tokens_per_vis
|
1166 |
-
sequence_logits = torch.stack(sequence_logits, dim=0) # (B, vocab_size)
|
1167 |
-
batch_logits.append(sequence_logits)
|
1168 |
-
|
1169 |
-
batch_logits = torch.stack(batch_logits, dim=0) # (B, T_txt, vocab_size)
|
1170 |
-
# The final logits shape should be the same as the original input_ids shape
|
1171 |
-
assert batch_logits.shape[:2] == (B, T_txt)
|
1172 |
-
|
1173 |
-
# assemble the output
|
1174 |
-
output = VLMOutputWithPast(
|
1175 |
-
loss=output.loss,
|
1176 |
-
logits=batch_logits,
|
1177 |
-
past_key_values=output.past_key_values,
|
1178 |
-
hidden_states=output.hidden_states,
|
1179 |
-
attentions=output.attentions,
|
1180 |
-
past_media_locations=updated_media_locations,
|
1181 |
-
past_vision_tokens=updated_vision_tokens,
|
1182 |
-
)
|
1183 |
-
|
1184 |
-
return output
|
1185 |
-
|
1186 |
-
def _post_forward_hook(self):
|
1187 |
-
pass
|
1188 |
-
|
1189 |
-
|
1190 |
-
@property
|
1191 |
-
def num_params_per_module(self):
|
1192 |
-
"""Print the number of parameters per module in the model"""
|
1193 |
-
return "\n".join(
|
1194 |
-
[
|
1195 |
-
f"Vision encoder: {num_params(self.vision_encoder):,} parameters",
|
1196 |
-
f"Vision tokenizer: {num_params(self.vision_tokenizer):,} parameters",
|
1197 |
-
f"Language model: {num_params(self.lang_model):,} parameters",
|
1198 |
-
]
|
1199 |
-
)
|
1200 |
-
|
1201 |
-
@property
|
1202 |
-
def num_trainable_params_per_module(self):
|
1203 |
-
"""Print the number of trainable parameters per module in the model"""
|
1204 |
-
return "\n".join(
|
1205 |
-
[
|
1206 |
-
f"Vision encoder: {num_params(self.vision_encoder, filter_to_trainable=True):,} trainable parameters",
|
1207 |
-
f"Vision tokenizer: {num_params(self.vision_tokenizer, filter_to_trainable=True):,} trainable parameters",
|
1208 |
-
f"Language model: {num_params(self.lang_model, filter_to_trainable=True):,} trainable parameters",
|
1209 |
-
]
|
1210 |
-
)
|
1211 |
-
|
1212 |
-
|
1213 |
-
class XGenMMPerceiver(VLMWithLanguageStream):
|
1214 |
-
def __init__(
|
1215 |
-
self,
|
1216 |
-
vision_encoder: nn.Module,
|
1217 |
-
vision_tokenizer: nn.Module,
|
1218 |
-
lang_model: nn.Module,
|
1219 |
-
initial_tokenizer_len: int,
|
1220 |
-
pad_token_id: int,
|
1221 |
-
decoder_layers_attr_name: str = None,
|
1222 |
-
gradient_checkpointing: bool = False,
|
1223 |
-
image_aspect_ratio: str = 'anyres',
|
1224 |
-
anyres_patch_sampling: bool = True,
|
1225 |
-
anyres_grids: list[int] = None,
|
1226 |
-
):
|
1227 |
-
"""
|
1228 |
-
Args:
|
1229 |
-
vision_encoder (nn.Module): HF CLIPModel
|
1230 |
-
lang_encoder (nn.Module): HF causal language model
|
1231 |
-
vis_feature_dim (int): final dimension of the visual features outputted by the vision_encoder
|
1232 |
-
initial_tokenizer_len (int): size of the tokenizer vocab
|
1233 |
-
padding_token_id (int): id of the padding token. None if no padding token; then a padding token
|
1234 |
-
will be inserted into self.special_tokens, which factory.py fills after creating new tokens
|
1235 |
-
decoder_layers_attr_name (str, optional): name of the decoder layers attribute. Defaults to None.
|
1236 |
-
gradient_checkpointing (bool, optional): whether to use gradient checkpointing. Defaults to False.
|
1237 |
-
"""
|
1238 |
-
self._special_tokens = {
|
1239 |
-
"media_token": "<image>",
|
1240 |
-
"image_placeholder_token": "<image placeholder>",
|
1241 |
-
"end_of_trunk_token": "<|endofchunk|>",
|
1242 |
-
}
|
1243 |
-
lang_embedding_dim = lang_model.get_input_embeddings().weight.shape[1]
|
1244 |
-
super().__init__(
|
1245 |
-
vision_encoder=vision_encoder,
|
1246 |
-
vision_tokenizer=vision_tokenizer,
|
1247 |
-
lang_model=lang_model,
|
1248 |
-
initial_tokenizer_len=initial_tokenizer_len,
|
1249 |
-
gradient_checkpointing=gradient_checkpointing,
|
1250 |
-
decoder_layers_attr_name=decoder_layers_attr_name,
|
1251 |
-
pad_token_id=pad_token_id,
|
1252 |
-
)
|
1253 |
-
self.image_aspect_ratio = image_aspect_ratio
|
1254 |
-
self.anyres_patch_sampling = anyres_patch_sampling
|
1255 |
-
self.anyres_grids = anyres_grids
|
1256 |
-
|
1257 |
-
def set_trainable(self):
|
1258 |
-
"""
|
1259 |
-
Unfreeze everything except the vision_encoder
|
1260 |
-
"""
|
1261 |
-
self.requires_grad_(True)
|
1262 |
-
self.vision_encoder.requires_grad_(False)
|
1263 |
-
|
1264 |
-
def _should_apply_weight_decay(self, parameter_name):
|
1265 |
-
"""
|
1266 |
-
Kosmos applies 0.01 weight deacy to everything
|
1267 |
-
"""
|
1268 |
-
return True
|
1269 |
-
|
1270 |
-
def generate(
|
1271 |
-
self,
|
1272 |
-
vision_x: torch.Tensor,
|
1273 |
-
lang_x: torch.Tensor,
|
1274 |
-
image_size: Optional[Tuple] = None,
|
1275 |
-
attention_mask: torch.Tensor = None,
|
1276 |
-
past_key_values: Optional[
|
1277 |
-
List[Union[torch.Tensor, Tuple[torch.Tensor]]]
|
1278 |
-
] = None,
|
1279 |
-
past_media_locations: Optional[torch.Tensor] = None,
|
1280 |
-
past_vision_tokens: Optional[torch.Tensor] = None,
|
1281 |
-
**kwargs,
|
1282 |
-
):
|
1283 |
-
"""
|
1284 |
-
Generate text conditioned on vision and language inputs.
|
1285 |
-
Args:
|
1286 |
-
vision_x (torch.Tensor): Vision input
|
1287 |
-
shape (B, T_img, F, C, H, W)
|
1288 |
-
see documentation for forward
|
1289 |
-
lang_x (torch.Tensor): Language input
|
1290 |
-
shape (B, T_txt)
|
1291 |
-
attention_mask (torch.Tensor, optional): Attention mask. Defaults to None.
|
1292 |
-
**kwargs: see generate documentation in Hugging Face CausalLM models.
|
1293 |
-
Returns:
|
1294 |
-
torch.Tensor: lang_x with generated tokens appended to it
|
1295 |
-
"""
|
1296 |
-
num_beams = kwargs.pop("num_beams", 1)
|
1297 |
-
|
1298 |
-
# convert pixels to vision tokens
|
1299 |
-
vision_attention_mask = None
|
1300 |
-
if vision_x is not None:
|
1301 |
-
if self.image_aspect_ratio == 'anyres':
|
1302 |
-
input_dict = dict(image=vision_x, image_size=image_size)
|
1303 |
-
vision_features, vision_attn_masks = self._encode_vision_x_anyres(input_dict, lang_x.device)
|
1304 |
-
else:
|
1305 |
-
vision_features = self._encode_vision_x(vision_x=vision_x)
|
1306 |
-
vision_attn_masks = None
|
1307 |
-
# If doing patch sampling, then flatten patches of shape [b, Np_i, v, d] -> [b*Np, v, d]
|
1308 |
-
# Same for attention masks: [b, Np, v] -> [b*Np, v]
|
1309 |
-
if self.anyres_patch_sampling:
|
1310 |
-
split_sizes = [feature.shape[0] for feature in vision_features]
|
1311 |
-
# Nested splits for multi-image samples.
|
1312 |
-
if isinstance(vision_x[0], list):
|
1313 |
-
nt_images = [len(images) for images in vision_x]
|
1314 |
-
split_split_sizes = []
|
1315 |
-
img_id = 0
|
1316 |
-
for nt in nt_images:
|
1317 |
-
split_split_sizes.append(split_sizes[img_id:img_id+nt])
|
1318 |
-
img_id += nt
|
1319 |
-
else:
|
1320 |
-
nt_images = [1] * len(vision_x)
|
1321 |
-
split_split_sizes = split_sizes
|
1322 |
-
vision_features = torch.cat(vision_features, dim=0)
|
1323 |
-
vision_features = vision_features[:, None, None, :, :] # Expand dimensions.
|
1324 |
-
vision_attn_masks = torch.cat(vision_attn_masks, dim=0)
|
1325 |
-
vision_tokens = self.vision_tokenizer(vision_features, vision_attn_masks)
|
1326 |
-
|
1327 |
-
# Post-processing: Split the batches into groups of patches and concatenate them together.
|
1328 |
-
if self.anyres_patch_sampling:
|
1329 |
-
assert isinstance(vision_x, list)
|
1330 |
-
if isinstance(vision_x[0], list):
|
1331 |
-
vision_token_groups = torch.split(vision_tokens, list(sum(nt_img) for nt_img in split_split_sizes), dim=0)
|
1332 |
-
vision_tokens = []
|
1333 |
-
|
1334 |
-
for sample_id, patch_vis_tokens in enumerate(vision_token_groups):
|
1335 |
-
patch_vis_token_groups = torch.split(patch_vis_tokens, split_split_sizes[sample_id], dim=0) # [Np*nt, 1, v, d] -> [[Np_t, 1, v, d], ...]
|
1336 |
-
flatten_vision_tokens = []
|
1337 |
-
for image_vis_token in patch_vis_token_groups:
|
1338 |
-
image_vis_token = image_vis_token.flatten(0, 2) # [Np, 1, v, d] -> [Np*v, d]
|
1339 |
-
flatten_vision_tokens.append(image_vis_token)
|
1340 |
-
vision_tokens_i = flatten_vision_tokens
|
1341 |
-
vision_tokens.append(vision_tokens_i)
|
1342 |
-
else:
|
1343 |
-
vision_token_groups = torch.split(vision_tokens, split_sizes, dim=0)
|
1344 |
-
vision_tokens = []
|
1345 |
-
for patch_vis_tokens in vision_token_groups:
|
1346 |
-
patch_vis_tokens = patch_vis_tokens.flatten(0, 2) # [Np, 1, v, d] -> [Np*v, d]
|
1347 |
-
vision_tokens.append(patch_vis_tokens.unsqueeze(0)) # Add the nt dimension.
|
1348 |
-
else:
|
1349 |
-
vision_tokens = None
|
1350 |
-
|
1351 |
-
# fuse the vision and language tokens
|
1352 |
-
# for xattn, vision_x and media_location are repeat_interleaved s.t.
|
1353 |
-
# the total batch size is B * num_beams
|
1354 |
-
new_inputs = self._prepare_inputs_for_forward(
|
1355 |
-
vision_tokens=vision_tokens,
|
1356 |
-
lang_x=lang_x,
|
1357 |
-
attention_mask=attention_mask,
|
1358 |
-
vision_attention_mask=vision_attention_mask,
|
1359 |
-
past_key_values=past_key_values,
|
1360 |
-
past_media_locations=past_media_locations,
|
1361 |
-
past_vision_tokens=past_vision_tokens,
|
1362 |
-
padding_side="left",
|
1363 |
-
num_beams=num_beams,
|
1364 |
-
)
|
1365 |
-
if past_key_values is not None:
|
1366 |
-
output = self.lang_model.generate(
|
1367 |
-
**new_inputs,
|
1368 |
-
past_key_values=past_key_values,
|
1369 |
-
num_beams=num_beams,
|
1370 |
-
use_cache=True,
|
1371 |
-
**kwargs,
|
1372 |
-
)
|
1373 |
-
else:
|
1374 |
-
output = self.lang_model.generate(
|
1375 |
-
**new_inputs,
|
1376 |
-
num_beams=num_beams,
|
1377 |
-
use_cache=True,
|
1378 |
-
**kwargs,
|
1379 |
-
)
|
1380 |
-
self._post_forward_hook()
|
1381 |
-
return output
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|