Push model using huggingface_hub.
Browse files- config.json +14 -8
- mini_gpt4_llama_v2.py +758 -0
config.json
CHANGED
@@ -1,8 +1,12 @@
|
|
1 |
{
|
2 |
"arch": "mini_gpt4_llama_v2",
|
3 |
"architectures": [
|
4 |
-
"
|
5 |
],
|
|
|
|
|
|
|
|
|
6 |
"chat_template": true,
|
7 |
"ckpt": "checkpoints/video_llama_checkpoint_last.pth",
|
8 |
"device": "cuda",
|
@@ -14,23 +18,25 @@
|
|
14 |
"length": 50,
|
15 |
"llama_model": "meta-llama/Llama-2-7b-chat-hf",
|
16 |
"lora_alpha": 16,
|
|
|
17 |
"lora_r": 64,
|
|
|
|
|
|
|
|
|
18 |
"low_resource": true,
|
19 |
"max_context_len": 3600,
|
20 |
"max_txt_len": 256,
|
21 |
"model_type": "minigpt4_video",
|
22 |
"num_query_token": 32,
|
23 |
"prompt": "",
|
|
|
|
|
|
|
24 |
"torch_dtype": "float32",
|
25 |
"transformers_version": "4.42.3",
|
26 |
"use_grad_checkpoint": true,
|
27 |
"use_grad_checkpoint_llm": true,
|
28 |
-
"vit_precision": "fp16",
|
29 |
"vit_model": "eva_clip_g",
|
30 |
-
"
|
31 |
-
"lora_target_modules" : ["q_proj","v_proj"],
|
32 |
-
"lora_dropout": 0.05,
|
33 |
-
"remove_template": false,
|
34 |
-
"prompt_path":""
|
35 |
-
|
36 |
}
|
|
|
1 |
{
|
2 |
"arch": "mini_gpt4_llama_v2",
|
3 |
"architectures": [
|
4 |
+
"MiniGPT4_Video"
|
5 |
],
|
6 |
+
"auto_map": {
|
7 |
+
"AutoConfig": "mini_gpt4_llama_v2.minigpt4_video_config",
|
8 |
+
"AutoModel": "mini_gpt4_llama_v2.MiniGPT4_Video"
|
9 |
+
},
|
10 |
"chat_template": true,
|
11 |
"ckpt": "checkpoints/video_llama_checkpoint_last.pth",
|
12 |
"device": "cuda",
|
|
|
18 |
"length": 50,
|
19 |
"llama_model": "meta-llama/Llama-2-7b-chat-hf",
|
20 |
"lora_alpha": 16,
|
21 |
+
"lora_dropout": 0.05,
|
22 |
"lora_r": 64,
|
23 |
+
"lora_target_modules": [
|
24 |
+
"q_proj",
|
25 |
+
"v_proj"
|
26 |
+
],
|
27 |
"low_resource": true,
|
28 |
"max_context_len": 3600,
|
29 |
"max_txt_len": 256,
|
30 |
"model_type": "minigpt4_video",
|
31 |
"num_query_token": 32,
|
32 |
"prompt": "",
|
33 |
+
"prompt_path": "",
|
34 |
+
"remove_template": false,
|
35 |
+
"token_pooling": true,
|
36 |
"torch_dtype": "float32",
|
37 |
"transformers_version": "4.42.3",
|
38 |
"use_grad_checkpoint": true,
|
39 |
"use_grad_checkpoint_llm": true,
|
|
|
40 |
"vit_model": "eva_clip_g",
|
41 |
+
"vit_precision": "fp16"
|
|
|
|
|
|
|
|
|
|
|
42 |
}
|
mini_gpt4_llama_v2.py
ADDED
@@ -0,0 +1,758 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
import random
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from torch.cuda.amp import autocast as autocast
|
6 |
+
import torch.nn as nn
|
7 |
+
|
8 |
+
from minigpt4_video.registry import registry
|
9 |
+
from minigpt4_video.blip2 import Blip2Base, disabled_train
|
10 |
+
# from minigpt4_video.modeling_llama_v2 import LlamaForCausalLM as llm_model
|
11 |
+
# from minigpt4_video.modeling_mistral import MistralForCausalLM as llm_model
|
12 |
+
from minigpt4_video.conversation import Conversation, SeparatorStyle, StoppingCriteriaList, StoppingCriteriaSub
|
13 |
+
|
14 |
+
from transformers import LlamaTokenizer
|
15 |
+
from transformers import BitsAndBytesConfig
|
16 |
+
from transformers import AutoConfig, AutoTokenizer
|
17 |
+
from peft import (
|
18 |
+
LoraConfig,
|
19 |
+
get_peft_model,
|
20 |
+
get_peft_model_state_dict,
|
21 |
+
prepare_model_for_int8_training,
|
22 |
+
set_peft_model_state_dict,
|
23 |
+
)
|
24 |
+
import time
|
25 |
+
import json
|
26 |
+
import numpy as np
|
27 |
+
import os
|
28 |
+
from transformers import PretrainedConfig
|
29 |
+
from transformers import PreTrainedModel
|
30 |
+
from typing import List
|
31 |
+
class minigpt4_video_config(PretrainedConfig):
|
32 |
+
model_type="minigpt4_video"
|
33 |
+
PRETRAINED_MODEL_CONFIG_DICT = {
|
34 |
+
"minigpt4_video": "configs/models/minigpt4.yaml",
|
35 |
+
}
|
36 |
+
def __init__(
|
37 |
+
self,
|
38 |
+
omg_config:dict = {},
|
39 |
+
**kwargs,
|
40 |
+
):
|
41 |
+
for key, value in omg_config.items():
|
42 |
+
setattr(self, key, value)
|
43 |
+
super().__init__(**kwargs)
|
44 |
+
|
45 |
+
# def to_dict(self):
|
46 |
+
# output = super().to_dict()
|
47 |
+
# return output
|
48 |
+
|
49 |
+
@registry.register_model("mini_gpt4_llama_v2")
|
50 |
+
class MiniGPT4_Video(Blip2Base, PreTrainedModel):
|
51 |
+
"""
|
52 |
+
BLIP2 GPT-LLAMA model.
|
53 |
+
"""
|
54 |
+
|
55 |
+
PRETRAINED_MODEL_CONFIG_DICT = {
|
56 |
+
"minigpt4_video": "minigpt4/configs/models/minigpt4.yaml",
|
57 |
+
}
|
58 |
+
config_class=minigpt4_video_config
|
59 |
+
|
60 |
+
def __init__(
|
61 |
+
self,
|
62 |
+
cfg={},
|
63 |
+
):
|
64 |
+
## loop through the config minigpt4_video_config object and set the attributes
|
65 |
+
if isinstance(cfg, minigpt4_video_config):
|
66 |
+
cfg = cfg.to_dict()
|
67 |
+
|
68 |
+
for key, value in cfg.items():
|
69 |
+
try:
|
70 |
+
setattr(self, key, value)
|
71 |
+
except:
|
72 |
+
print(f"Error setting attribute {key} with value {value}")
|
73 |
+
PreTrainedModel.__init__(self, minigpt4_video_config(cfg))
|
74 |
+
Blip2Base.__init__(self)
|
75 |
+
if "Mistral" in self.llama_model:
|
76 |
+
from minigpt4_video.modeling_mistral import MistralForCausalLM as llm_model
|
77 |
+
print("Mistral model")
|
78 |
+
self.model_type = "Mistral"
|
79 |
+
else:
|
80 |
+
from minigpt4_video.modeling_llama_v2 import LlamaForCausalLM as llm_model
|
81 |
+
print("Llama model")
|
82 |
+
self.model_type = "Llama"
|
83 |
+
self.tokenizer = self.init_tokenizer()
|
84 |
+
|
85 |
+
print("token pooling", self.token_pooling)
|
86 |
+
if self.freeze_vit:
|
87 |
+
# self.vit_precision="fp32"
|
88 |
+
print("vit precision", self.vit_precision)
|
89 |
+
self.visual_encoder, self.ln_vision = self.init_vision_encoder(
|
90 |
+
self.vit_model, self.img_size, self.drop_path_rate, self.use_grad_checkpoint, self.vit_precision
|
91 |
+
)
|
92 |
+
for name, param in self.visual_encoder.named_parameters():
|
93 |
+
param.requires_grad = False
|
94 |
+
self.visual_encoder = self.visual_encoder.eval()
|
95 |
+
self.visual_encoder.train = disabled_train
|
96 |
+
for name, param in self.ln_vision.named_parameters():
|
97 |
+
param.requires_grad = False
|
98 |
+
self.ln_vision = self.ln_vision.eval()
|
99 |
+
self.ln_vision.train = disabled_train
|
100 |
+
logging.info("freeze vision encoder")
|
101 |
+
print("freeze the vision encoder")
|
102 |
+
|
103 |
+
else:
|
104 |
+
self.vit_precision="fp32"
|
105 |
+
self.visual_encoder, self.ln_vision = self.init_vision_encoder(
|
106 |
+
self.vit_model, self.img_size, self.drop_path_rate, self.use_grad_checkpoint, self.vit_precision
|
107 |
+
)
|
108 |
+
|
109 |
+
print("unfreeze the vision encoder")
|
110 |
+
print('Loading VIT Done')
|
111 |
+
|
112 |
+
print('Loading LLAMA')
|
113 |
+
|
114 |
+
self.B_SYS, self.E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
|
115 |
+
token=os.environ.get("HF_TKN")
|
116 |
+
self.llama_tokenizer = LlamaTokenizer.from_pretrained(self.llama_model,use_fast=False,token=token) #
|
117 |
+
self.llama_tokenizer.pad_token = "$$"
|
118 |
+
# use fastv
|
119 |
+
self.use_fastv = False
|
120 |
+
print("self.low_resource",self.low_resource)
|
121 |
+
if self.low_resource:
|
122 |
+
self.llama_model = llm_model.from_pretrained(
|
123 |
+
self.llama_model,
|
124 |
+
torch_dtype=torch.float16,
|
125 |
+
# torch_dtype = torch.bfloat16,
|
126 |
+
load_in_8bit=True,
|
127 |
+
# device_map = "balanced"
|
128 |
+
# device_map="auto",
|
129 |
+
device_map={'':torch.cuda.current_device()},token=token
|
130 |
+
# device_map={'':0}
|
131 |
+
|
132 |
+
)
|
133 |
+
else:
|
134 |
+
self.llama_model = llm_model.from_pretrained(
|
135 |
+
self.llama_model,
|
136 |
+
torch_dtype=torch.float16,token=token
|
137 |
+
)
|
138 |
+
|
139 |
+
# self.llama_model.resize_token_embeddings(len(self.llama_tokenizer))
|
140 |
+
self.llama_model = prepare_model_for_int8_training(self.llama_model)
|
141 |
+
loraconfig = LoraConfig(
|
142 |
+
r=self.lora_r,
|
143 |
+
lora_alpha=self.lora_alpha,
|
144 |
+
target_modules=self.lora_target_modules,
|
145 |
+
lora_dropout=self.lora_dropout,
|
146 |
+
bias="none",
|
147 |
+
task_type="CAUSAL_LM"
|
148 |
+
)
|
149 |
+
self.llama_model = get_peft_model(self.llama_model, loraconfig)
|
150 |
+
|
151 |
+
self.llama_model.print_trainable_parameters()
|
152 |
+
|
153 |
+
if self.use_grad_checkpoint_llm:
|
154 |
+
self.llama_model.gradient_checkpointing_enable()
|
155 |
+
|
156 |
+
print('Loading LLAMA Done')
|
157 |
+
|
158 |
+
|
159 |
+
if self.token_pooling:
|
160 |
+
self.llama_proj = nn.Linear(
|
161 |
+
1408*4, self.llama_model.config.hidden_size
|
162 |
+
)
|
163 |
+
else:
|
164 |
+
self.llama_proj = nn.Linear(
|
165 |
+
1408, self.llama_model.config.hidden_size
|
166 |
+
)
|
167 |
+
if self.prompt_path:
|
168 |
+
with open(self.prompt_path, 'r') as f:
|
169 |
+
raw_prompts = f.read().splitlines()
|
170 |
+
filted_prompts = [raw_prompt for raw_prompt in raw_prompts if "<ImageHere>" in raw_prompt]
|
171 |
+
self.prompt_list = [self.prompt_template.format(p) for p in filted_prompts]
|
172 |
+
print('Load {} training prompts'.format(len(self.prompt_list)))
|
173 |
+
print('Prompt Example \n{}'.format(random.choice(self.prompt_list)))
|
174 |
+
else:
|
175 |
+
self.prompt_list = []
|
176 |
+
|
177 |
+
def encode_img(self, image):
|
178 |
+
device = image.device
|
179 |
+
if len(image.shape) > 4:
|
180 |
+
image = image.reshape(-1, *image.shape[-3:]) # for video input flatten the batch and time dimension (4,50,3,224,224) -> (200,3,224,224)
|
181 |
+
with self.maybe_autocast():
|
182 |
+
image_embeds = self.ln_vision(self.visual_encoder(image)).to(device) # (200,3,224,224) -> (200,257,1408)
|
183 |
+
image_embeds = image_embeds[:,1:,:] # remove the first token (CLS) (200,256,1408)
|
184 |
+
bs, pn, hs = image_embeds.shape
|
185 |
+
if self.token_pooling: # concat the each 4 tokens into one token (200,64,5632)
|
186 |
+
image_embeds = image_embeds.view(bs, int(pn/4), int(hs*4)) # (200,64,5632)
|
187 |
+
|
188 |
+
inputs_llama = self.llama_proj(image_embeds) # project to llama input size (200,64,5632) -> (200,64,4096)
|
189 |
+
atts_llama = torch.ones(inputs_llama.size()[:-1], dtype=torch.long).to(image.device)
|
190 |
+
return inputs_llama, atts_llama
|
191 |
+
|
192 |
+
def get_context_emb(self, prompt, img_list):
|
193 |
+
img_device = img_list[0].device
|
194 |
+
prompt_segs = prompt.split('<ImageHere>')
|
195 |
+
assert len(prompt_segs) == len(img_list) + 1, "Unmatched numbers of image placeholders and images."
|
196 |
+
seg_tokens = [
|
197 |
+
self.llama_tokenizer(
|
198 |
+
seg, return_tensors="pt", add_special_tokens=i==0).to(img_device).input_ids # only add bos to the first seg
|
199 |
+
for i, seg in enumerate(prompt_segs)
|
200 |
+
]
|
201 |
+
|
202 |
+
seg_embs = [self.embed_tokens(seg_t) for seg_t in seg_tokens]
|
203 |
+
|
204 |
+
mixed_embs = [emb for pair in zip(seg_embs[:-1], img_list) for emb in pair] + [seg_embs[-1]]
|
205 |
+
|
206 |
+
mixed_embs = torch.cat(mixed_embs, dim=1)
|
207 |
+
|
208 |
+
return mixed_embs
|
209 |
+
|
210 |
+
def prompt_wrap(self, img_embeds, atts_img, prompts, lengths=None):
|
211 |
+
if prompts is None or len(prompts) == 0:
|
212 |
+
# prompts is not provided, just return the original image embedding
|
213 |
+
return img_embeds, atts_img
|
214 |
+
elif img_embeds is None:
|
215 |
+
# prompt is provided but there is no image embedding. return the prompt embedding in right padding
|
216 |
+
self.llama_tokenizer.padding_side = "right"
|
217 |
+
prompt_tokens = self.llama_tokenizer(
|
218 |
+
prompts,
|
219 |
+
return_tensors="pt",
|
220 |
+
padding="max_length",
|
221 |
+
add_special_tokens=False
|
222 |
+
).to(self.device)
|
223 |
+
prompt_embeds = self.embed_tokens(prompt_tokens.input_ids)
|
224 |
+
atts_prompt = prompt_tokens.attention_mask
|
225 |
+
return prompt_embeds, atts_prompt
|
226 |
+
|
227 |
+
else:
|
228 |
+
# return the multi-modal embedding in right padding
|
229 |
+
emb_lists = []
|
230 |
+
if type(prompts) == str:
|
231 |
+
prompts = [prompts] * len(img_embeds)
|
232 |
+
for idx, (each_img_embed, each_prompt) in enumerate(zip(img_embeds, prompts)):
|
233 |
+
pn = each_img_embed.shape[-2]
|
234 |
+
if lengths is not None:
|
235 |
+
each_img_embed = each_img_embed.reshape(-1, each_img_embed.shape[-1])
|
236 |
+
each_img_embed = each_img_embed[:lengths[idx] * pn]
|
237 |
+
|
238 |
+
p_segs = each_prompt.split('<ImageHere>')
|
239 |
+
interleave_emb = []
|
240 |
+
for idx, seg in enumerate(p_segs[:-1]):
|
241 |
+
p_tokens = self.llama_tokenizer(seg, return_tensors="pt", add_special_tokens=False).to(img_embeds.device)
|
242 |
+
p_embed = self.embed_tokens(p_tokens.input_ids)
|
243 |
+
|
244 |
+
interleave_emb.append(torch.cat([p_embed, each_img_embed[None][:, idx*pn:(idx+1)*pn]], dim=1))
|
245 |
+
|
246 |
+
wrapped_emb = torch.cat(interleave_emb, dim=1)
|
247 |
+
p_tokens = self.llama_tokenizer(p_segs[-1], return_tensors="pt", add_special_tokens=False).to(img_embeds.device)
|
248 |
+
p_embed = self.embed_tokens(p_tokens.input_ids)
|
249 |
+
wrapped_emb = torch.cat([wrapped_emb,p_embed], dim=1)
|
250 |
+
emb_lists.append(wrapped_emb)
|
251 |
+
|
252 |
+
emb_lens = [emb.shape[1] for emb in emb_lists]
|
253 |
+
pad_emb = self.embed_tokens(torch.tensor(self.llama_tokenizer.pad_token_id, device=img_embeds.device))
|
254 |
+
|
255 |
+
# max_length = max(emb_lens) if max(emb_lens) < self.max_context_len else self.max_context_len
|
256 |
+
max_length = self.max_context_len
|
257 |
+
wrapped_embs = pad_emb.expand(len(emb_lens), max_length, -1).clone()
|
258 |
+
wrapped_atts = torch.zeros([len(emb_lens), max_length], dtype=torch.int, device=img_embeds.device)
|
259 |
+
|
260 |
+
for i, emb in enumerate(emb_lists):
|
261 |
+
length = emb_lens[i] if emb_lens[i] < self.max_context_len else self.max_context_len
|
262 |
+
wrapped_embs[i, :length] = emb[:, :length]
|
263 |
+
wrapped_atts[i, :length] = 1
|
264 |
+
|
265 |
+
return wrapped_embs, wrapped_atts
|
266 |
+
|
267 |
+
def concat_emb_input_output(self, input_embs, input_atts, output_embs, output_atts):
|
268 |
+
"""
|
269 |
+
Concatenate the batched input embedding and batched output embedding together.
|
270 |
+
Both the input and the output embedding should be right padded.
|
271 |
+
"""
|
272 |
+
|
273 |
+
input_lens = []
|
274 |
+
cat_embs = []
|
275 |
+
cat_atts = []
|
276 |
+
|
277 |
+
for i in range(input_embs.size(0)):
|
278 |
+
input_len = input_atts[i].sum()
|
279 |
+
input_lens.append(input_len)
|
280 |
+
|
281 |
+
cat_embs.append(
|
282 |
+
torch.cat([
|
283 |
+
input_embs[i][:input_len],
|
284 |
+
output_embs[i],
|
285 |
+
input_embs[i][input_len:]
|
286 |
+
])
|
287 |
+
)
|
288 |
+
cat_atts.append(
|
289 |
+
torch.cat([
|
290 |
+
input_atts[i][:input_len],
|
291 |
+
output_atts[i],
|
292 |
+
input_atts[i][input_len:]
|
293 |
+
])
|
294 |
+
)
|
295 |
+
|
296 |
+
cat_embs = torch.stack(cat_embs)
|
297 |
+
cat_atts = torch.stack(cat_atts)
|
298 |
+
return cat_embs, cat_atts, input_lens
|
299 |
+
|
300 |
+
def get_conv_emb(self, conv_q, conv_a, conv_img):
|
301 |
+
"""concatenate conversation and make sure the model is only trained to regress the answer"""
|
302 |
+
|
303 |
+
regress_embs_list = []
|
304 |
+
targets_list = []
|
305 |
+
|
306 |
+
batch_size = len(conv_q)
|
307 |
+
for batch_idx in range(batch_size):
|
308 |
+
questions, answers = conv_q[batch_idx], conv_a[batch_idx]
|
309 |
+
assigned_imgs = conv_img[batch_idx]
|
310 |
+
questions = [self.prompt_wrap(
|
311 |
+
img_embeds=img,
|
312 |
+
atts_img=None,
|
313 |
+
prompts=[q],
|
314 |
+
lengths=[img.shape[1]] if img is not None else None) for q, img in zip(questions, assigned_imgs)]
|
315 |
+
q_embs = [emb for emb, _ in questions]
|
316 |
+
|
317 |
+
answers = [self.llama_tokenizer(a, return_tensors="pt", add_special_tokens=False).to(self.device) for a in answers]
|
318 |
+
cur_emb = []
|
319 |
+
cur_target = []
|
320 |
+
for i in range(len(questions)):
|
321 |
+
cur_emb.append(q_embs[i])
|
322 |
+
cur_target.append(torch.ones_like(q_embs[i][..., 0], dtype=torch.int) * -100)
|
323 |
+
|
324 |
+
cur_emb.append(self.embed_tokens(answers[i].input_ids))
|
325 |
+
cur_target.append(answers[i].input_ids)
|
326 |
+
|
327 |
+
cur_emb = torch.cat(cur_emb, dim=1)
|
328 |
+
cur_target = torch.cat(cur_target, dim=1)
|
329 |
+
|
330 |
+
regress_embs_list.append(cur_emb)
|
331 |
+
targets_list.append(cur_target)
|
332 |
+
|
333 |
+
max_len = min(max([target.shape[1] for target in targets_list]), self.max_txt_len)
|
334 |
+
|
335 |
+
regress_embeds = torch.zeros([batch_size, max_len, cur_emb.shape[-1]], device=self.device)
|
336 |
+
regress_attn = torch.zeros([batch_size, max_len], dtype=torch.int, device=self.device)
|
337 |
+
targets = torch.ones([batch_size, max_len], dtype=torch.long, device=self.device) * -100
|
338 |
+
|
339 |
+
for batch_idx in range(batch_size):
|
340 |
+
cur_len = regress_embs_list[batch_idx].shape[1]
|
341 |
+
regress_embeds[batch_idx, :cur_len] = regress_embs_list[batch_idx][0, :max_len]
|
342 |
+
regress_attn[batch_idx, :cur_len] = 1
|
343 |
+
targets[batch_idx, :cur_len] = targets_list[batch_idx][0, :max_len]
|
344 |
+
|
345 |
+
return regress_embeds, regress_attn, targets
|
346 |
+
|
347 |
+
def preparing_embedding(self, samples):
|
348 |
+
def remove_special_tokens(data):
|
349 |
+
|
350 |
+
# if "instruction_input" in data:
|
351 |
+
data = [instruct.replace(" [caption]","") for instruct in data]
|
352 |
+
data = [instruct.replace(" [vqa]","") for instruct in data]
|
353 |
+
data = [instruct.replace(" [grounding]","") for instruct in data]
|
354 |
+
data = [instruct.replace(" [identify]","") for instruct in data]
|
355 |
+
data = [instruct.replace(" [refer]","") for instruct in data]
|
356 |
+
return data
|
357 |
+
|
358 |
+
### prepare input tokens
|
359 |
+
if 'image' in samples:
|
360 |
+
img_embeds, img_atts = self.encode_img(samples["image"])
|
361 |
+
else:
|
362 |
+
img_embeds = img_atts = None
|
363 |
+
|
364 |
+
if 'conv_q' in samples:
|
365 |
+
# handeling conversation datasets
|
366 |
+
conv_q, conv_a = samples['conv_q'], samples['conv_a']
|
367 |
+
|
368 |
+
connect_sym = samples['connect_sym'][0]
|
369 |
+
conv_q = [q.split(connect_sym)for q in conv_q]
|
370 |
+
conv_a = [a.split(connect_sym) for a in conv_a]
|
371 |
+
conv_img = assign_imgs(conv_q, img_embeds)
|
372 |
+
|
373 |
+
if self.chat_template:
|
374 |
+
conv_q = [["[INST] " + item + "[/INST]" for item in items] for items in conv_q]
|
375 |
+
|
376 |
+
regress_embeds, regress_atts, part_targets = self.get_conv_emb(conv_q, conv_a, conv_img)
|
377 |
+
cond_embeds, cond_atts = regress_embeds[:, :0], regress_atts[:, :0]
|
378 |
+
|
379 |
+
else:
|
380 |
+
if "instruction_input" in samples:
|
381 |
+
instruction = samples["instruction_input"]
|
382 |
+
elif len(self.prompt_list) > 1:
|
383 |
+
instruction = random.choice(self.prompt_list)
|
384 |
+
else:
|
385 |
+
instruction = None
|
386 |
+
|
387 |
+
if self.remove_template:
|
388 |
+
instruction = remove_special_tokens(instruction)
|
389 |
+
|
390 |
+
if self.chat_template:
|
391 |
+
instruction = ["[INST] " + instruct + "[/INST]" for instruct in instruction]
|
392 |
+
|
393 |
+
if 'length' in samples:
|
394 |
+
# the input is a image train (like videos)
|
395 |
+
bsz, pn, hs = img_embeds.shape
|
396 |
+
img_embeds = img_embeds.reshape(len(samples['image']), -1, pn, hs) # (200,64,4096) -> (4,50,64,4096)
|
397 |
+
cond_embeds, cond_atts = self.prompt_wrap(img_embeds, img_atts, instruction, samples['length'])
|
398 |
+
else:
|
399 |
+
cond_embeds, cond_atts = self.prompt_wrap(img_embeds, img_atts, instruction)
|
400 |
+
|
401 |
+
### prepare target tokens
|
402 |
+
self.llama_tokenizer.padding_side = "right"
|
403 |
+
text = [t + self.end_sym for t in samples["answer"]]
|
404 |
+
|
405 |
+
regress_tokens = self.llama_tokenizer(
|
406 |
+
text,
|
407 |
+
return_tensors="pt",
|
408 |
+
padding="max_length",
|
409 |
+
truncation=True,
|
410 |
+
max_length=self.max_txt_len,
|
411 |
+
add_special_tokens=False
|
412 |
+
).to(self.device)
|
413 |
+
|
414 |
+
regress_token_ids = regress_tokens.input_ids
|
415 |
+
regress_atts = regress_tokens.attention_mask
|
416 |
+
part_targets = regress_token_ids.masked_fill(
|
417 |
+
regress_token_ids == self.llama_tokenizer.pad_token_id, -100
|
418 |
+
)
|
419 |
+
|
420 |
+
regress_embeds = self.embed_tokens(regress_token_ids)
|
421 |
+
|
422 |
+
return cond_embeds, cond_atts, regress_embeds, regress_atts, part_targets
|
423 |
+
|
424 |
+
def forward(self, samples, reduction="mean"):
|
425 |
+
# prepare the embedding to condition and the embedding to regress
|
426 |
+
cond_embeds, cond_atts, regress_embeds, regress_atts, part_targets = \
|
427 |
+
self.preparing_embedding(samples)
|
428 |
+
|
429 |
+
# concat the embedding to condition and the embedding to regress
|
430 |
+
inputs_embeds, attention_mask, input_lens = \
|
431 |
+
self.concat_emb_input_output(cond_embeds, cond_atts, regress_embeds, regress_atts)
|
432 |
+
# get bos token embedding
|
433 |
+
bos = torch.ones_like(part_targets[:, :1]) * self.llama_tokenizer.bos_token_id
|
434 |
+
bos_embeds = self.embed_tokens(bos)
|
435 |
+
bos_atts = attention_mask[:, :1]
|
436 |
+
|
437 |
+
# add bos token at the begining
|
438 |
+
inputs_embeds = torch.cat([bos_embeds, inputs_embeds], dim=1)
|
439 |
+
attention_mask = torch.cat([bos_atts, attention_mask], dim=1)
|
440 |
+
|
441 |
+
targets = torch.ones([inputs_embeds.shape[0], inputs_embeds.shape[1]],
|
442 |
+
dtype=torch.long).to(self.device).fill_(-100)
|
443 |
+
for i, target in enumerate(part_targets):
|
444 |
+
targets[i, input_lens[i]+1:input_lens[i]+len(target)+1] = target # plus 1 for bos
|
445 |
+
|
446 |
+
with self.maybe_autocast():
|
447 |
+
outputs = self.llama_model(
|
448 |
+
inputs_embeds=inputs_embeds,
|
449 |
+
attention_mask=attention_mask,
|
450 |
+
return_dict=True,
|
451 |
+
labels=targets,
|
452 |
+
reduction=reduction,
|
453 |
+
use_fastv=self.use_fastv
|
454 |
+
)
|
455 |
+
loss = outputs.loss
|
456 |
+
|
457 |
+
return {"loss": loss}
|
458 |
+
|
459 |
+
@torch.no_grad()
|
460 |
+
def generate(
|
461 |
+
self,
|
462 |
+
images,
|
463 |
+
texts,
|
464 |
+
use_nucleus_sampling=False,
|
465 |
+
num_beams=1,
|
466 |
+
max_new_tokens=20,
|
467 |
+
min_length=1,
|
468 |
+
top_p=0.9,
|
469 |
+
repetition_penalty=1.5,
|
470 |
+
length_penalty=1,
|
471 |
+
temperature=1,
|
472 |
+
do_sample=False,
|
473 |
+
stop_words_ids=[2],
|
474 |
+
lengths=None,
|
475 |
+
return_video_temporal_features=False,
|
476 |
+
img_embeds=None,
|
477 |
+
):
|
478 |
+
'''
|
479 |
+
function for generate test use
|
480 |
+
'''
|
481 |
+
|
482 |
+
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(
|
483 |
+
stops=[torch.tensor([i]).to(self.device) for i in stop_words_ids])])
|
484 |
+
if img_embeds is None:
|
485 |
+
img_embeds, atts_img = self.encode_img(images.to(self.device))
|
486 |
+
else:
|
487 |
+
# Use images features from the input(4,45,64,5632)
|
488 |
+
img_embeds = img_embeds.reshape(-1, *img_embeds.shape[-2:])
|
489 |
+
img_embeds= img_embeds.to(self.device)
|
490 |
+
img_embeds = self.llama_proj(img_embeds) # project to llama input size (200,64,5632) -> (200,64,4096)
|
491 |
+
atts_img = torch.ones(img_embeds.size()[:-1], dtype=torch.long).to(self.device)
|
492 |
+
|
493 |
+
if lengths is not None:
|
494 |
+
image_lists = []
|
495 |
+
img_embeds = img_embeds.reshape(len(lengths), -1, img_embeds.shape[-2], img_embeds.shape[-1])
|
496 |
+
for idx, img_embed in enumerate(img_embeds):
|
497 |
+
image_lists.append([img_embed[i][None] for i in range(lengths[idx])])
|
498 |
+
else:
|
499 |
+
image_lists = [[image_emb[None]] for image_emb in img_embeds]
|
500 |
+
assert len(texts) == len(image_lists)
|
501 |
+
batch_embs = [self.get_context_emb(text, img_list) for text, img_list in zip(texts, image_lists)]
|
502 |
+
|
503 |
+
batch_size = len(batch_embs)
|
504 |
+
max_len = max([emb.shape[1] for emb in batch_embs])
|
505 |
+
emb_dim = batch_embs[0].shape[2]
|
506 |
+
dtype = batch_embs[0].dtype
|
507 |
+
device = batch_embs[0].device
|
508 |
+
|
509 |
+
embs = torch.zeros([batch_size, max_len, emb_dim], dtype=dtype, device=device)
|
510 |
+
attn_mask = torch.zeros([batch_size, max_len], dtype=torch.int, device=device)
|
511 |
+
for i, emb in enumerate(batch_embs):
|
512 |
+
emb_len = emb.shape[1]
|
513 |
+
embs[i, -emb_len:] = emb[0]
|
514 |
+
attn_mask[i, -emb_len:] = 1
|
515 |
+
# check if the input embedding tokens are in the range of the model cotext window (4096) and if it is not, then truncate it to the max context window
|
516 |
+
if self.model_type == "Llama":
|
517 |
+
context_window = 3700
|
518 |
+
else:
|
519 |
+
context_window = 7500
|
520 |
+
if embs.shape[1] > context_window:
|
521 |
+
embs = embs[:, -context_window:]
|
522 |
+
attn_mask = attn_mask[:, -context_window:]
|
523 |
+
with self.maybe_autocast():
|
524 |
+
if return_video_temporal_features:
|
525 |
+
last_hidden_state = self.llama_model(
|
526 |
+
inputs_embeds=embs,
|
527 |
+
attention_mask=attn_mask,
|
528 |
+
output_hidden_states=True,
|
529 |
+
).hidden_states[-1]
|
530 |
+
video_temporal_features = last_hidden_state.mean(dim=1)
|
531 |
+
# normalize the temporal features using L2 norm
|
532 |
+
# video_temporal_features = video_temporal_features / video_temporal_features.norm(dim=-1, keepdim=True)
|
533 |
+
outputs = self.llama_model.generate(
|
534 |
+
inputs_embeds=embs,
|
535 |
+
attention_mask=attn_mask,
|
536 |
+
max_new_tokens=max_new_tokens,
|
537 |
+
num_beams=num_beams,
|
538 |
+
do_sample=do_sample,
|
539 |
+
temperature=temperature,
|
540 |
+
repetition_penalty=repetition_penalty,
|
541 |
+
# stopping_criteria=stopping_criteria,
|
542 |
+
use_fastv=False,
|
543 |
+
)
|
544 |
+
|
545 |
+
answers = []
|
546 |
+
for output_token in outputs:
|
547 |
+
if output_token[0] == 0:
|
548 |
+
output_token = output_token[1:]
|
549 |
+
output_texts = self.llama_tokenizer.decode(output_token, skip_special_tokens=True)
|
550 |
+
output_texts = output_texts.split('</s>')[0] # remove the stop sign </s>
|
551 |
+
output_texts = output_texts.replace("<s>", "")
|
552 |
+
output_texts = output_texts.split(r'[/INST]')[-1].strip()
|
553 |
+
answers.append(output_texts)
|
554 |
+
if return_video_temporal_features:
|
555 |
+
return answers, video_temporal_features
|
556 |
+
else:
|
557 |
+
return answers
|
558 |
+
|
559 |
+
@torch.no_grad()
|
560 |
+
def generate_text_only(
|
561 |
+
self,
|
562 |
+
images,
|
563 |
+
seg_tokens,
|
564 |
+
use_nucleus_sampling=False,
|
565 |
+
num_beams=1,
|
566 |
+
max_new_tokens=20,
|
567 |
+
min_length=1,
|
568 |
+
top_p=0.9,
|
569 |
+
repetition_penalty=1.5,
|
570 |
+
length_penalty=1,
|
571 |
+
temperature=1,
|
572 |
+
do_sample=False,
|
573 |
+
stop_words_ids=[2],
|
574 |
+
lengths=None,
|
575 |
+
return_video_temporal_features=False,
|
576 |
+
img_embeds=None,
|
577 |
+
):
|
578 |
+
'''
|
579 |
+
function for generate test use
|
580 |
+
'''
|
581 |
+
|
582 |
+
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(
|
583 |
+
stops=[torch.tensor([i]).to(self.device) for i in stop_words_ids])])
|
584 |
+
|
585 |
+
batch_embs = [torch.cat([self.embed_tokens(seg_t)]) for seg_t in seg_tokens]
|
586 |
+
|
587 |
+
batch_size = len(batch_embs)
|
588 |
+
max_len = max([emb.shape[1] for emb in batch_embs])
|
589 |
+
emb_dim = batch_embs[0].shape[2]
|
590 |
+
dtype = batch_embs[0].dtype
|
591 |
+
device = batch_embs[0].device
|
592 |
+
|
593 |
+
embs = torch.zeros([batch_size, max_len, emb_dim], dtype=dtype, device=device)
|
594 |
+
attn_mask = torch.zeros([batch_size, max_len], dtype=torch.int, device=device)
|
595 |
+
for i, emb in enumerate(batch_embs):
|
596 |
+
emb_len = emb.shape[1]
|
597 |
+
embs[i, -emb_len:] = emb[0]
|
598 |
+
attn_mask[i, -emb_len:] = 1
|
599 |
+
|
600 |
+
with self.maybe_autocast():
|
601 |
+
outputs = self.llama_model.generate(
|
602 |
+
inputs_embeds=embs,
|
603 |
+
attention_mask=attn_mask,
|
604 |
+
max_new_tokens=max_new_tokens,
|
605 |
+
num_beams=num_beams,
|
606 |
+
do_sample=do_sample,
|
607 |
+
temperature=temperature,
|
608 |
+
repetition_penalty=repetition_penalty,
|
609 |
+
# stopping_criteria=stopping_criteria,
|
610 |
+
)
|
611 |
+
|
612 |
+
answers = []
|
613 |
+
for output_token in outputs:
|
614 |
+
if output_token[0] == 0:
|
615 |
+
output_token = output_token[1:]
|
616 |
+
output_texts = self.llama_tokenizer.decode(output_token, skip_special_tokens=True)
|
617 |
+
output_texts = output_texts.split('</s>')[0] # remove the stop sign </s>
|
618 |
+
output_texts = output_texts.replace("<s>", "")
|
619 |
+
output_texts = output_texts.split(r'[/INST]')[-1].strip()
|
620 |
+
answers.append(output_texts)
|
621 |
+
return answers
|
622 |
+
|
623 |
+
|
624 |
+
|
625 |
+
@torch.no_grad()
|
626 |
+
def multi_select(self, images, texts, answers, num_cand=None):
|
627 |
+
all_losses = []
|
628 |
+
for answer in answers:
|
629 |
+
choice_samples = {
|
630 |
+
'image': images,
|
631 |
+
'instruction_input': texts,
|
632 |
+
'answer': answer
|
633 |
+
}
|
634 |
+
loss = self.forward(choice_samples, reduction='none')['loss'].reshape(-1, 1)
|
635 |
+
all_losses.append(loss)
|
636 |
+
torch.cuda.empty_cache()
|
637 |
+
all_losses = torch.cat(all_losses, dim=-1)
|
638 |
+
if num_cand is not None:
|
639 |
+
for i in range(all_losses.shape[0]):
|
640 |
+
all_losses[i, num_cand[i]:] = 9999
|
641 |
+
output_class_ranks = torch.argsort(all_losses, dim=-1)
|
642 |
+
return output_class_ranks.tolist()
|
643 |
+
|
644 |
+
def predict_answers(
|
645 |
+
self,
|
646 |
+
samples,
|
647 |
+
num_beams=5,
|
648 |
+
inference_method="generate",
|
649 |
+
max_len=10,
|
650 |
+
min_len=1,
|
651 |
+
num_ans_candidates=128,
|
652 |
+
answer_list=None,
|
653 |
+
prompt="",
|
654 |
+
length_penalty=0,
|
655 |
+
**kwargs
|
656 |
+
):
|
657 |
+
'''
|
658 |
+
function for open-ended VQA
|
659 |
+
'''
|
660 |
+
images = samples["image"].cuda()
|
661 |
+
texts = samples["instruction_input"]
|
662 |
+
|
663 |
+
output_text = self.generate(
|
664 |
+
images=images,
|
665 |
+
texts=texts,
|
666 |
+
num_beams=num_beams,
|
667 |
+
max_new_tokens=max_len,
|
668 |
+
min_length=min_len,
|
669 |
+
length_penalty=length_penalty
|
670 |
+
)
|
671 |
+
|
672 |
+
if "apply_lemmatizer" in samples.keys() and samples["apply_lemmatizer"]:
|
673 |
+
output_text = self._lemmatize(output_text)
|
674 |
+
|
675 |
+
return output_text
|
676 |
+
|
677 |
+
def predict_class(
|
678 |
+
self,
|
679 |
+
samples,
|
680 |
+
num_beams=5,
|
681 |
+
inference_method="generate",
|
682 |
+
max_len=10,
|
683 |
+
min_len=1,
|
684 |
+
num_ans_candidates=5,
|
685 |
+
answer_list=None,
|
686 |
+
prompt="",
|
687 |
+
length_penalty=0,
|
688 |
+
**kwargs
|
689 |
+
):
|
690 |
+
'''
|
691 |
+
function for multi-choice VQA
|
692 |
+
'''
|
693 |
+
|
694 |
+
image = samples["image"].cuda()
|
695 |
+
instruction = samples['instruction_input']
|
696 |
+
answers = samples["choices"]
|
697 |
+
num_cand = samples["num_choices"]
|
698 |
+
|
699 |
+
ranks = self.multi_select(image, instruction, answers, num_cand)
|
700 |
+
|
701 |
+
pred_ans = []
|
702 |
+
for i, rank in enumerate(ranks):
|
703 |
+
pred = answers[rank[0]][i]
|
704 |
+
pred_ans.append(pred)
|
705 |
+
return pred_ans
|
706 |
+
|
707 |
+
def embed_tokens(self, token_ids):
|
708 |
+
try:
|
709 |
+
embeds = self.llama_model.base_model.model.model.embed_tokens(token_ids)
|
710 |
+
except AttributeError:
|
711 |
+
embeds = self.llama_model.model.embed_tokens(token_ids)
|
712 |
+
|
713 |
+
return embeds
|
714 |
+
|
715 |
+
@classmethod
|
716 |
+
def from_config(cls, cfg):
|
717 |
+
model = cls(
|
718 |
+
cfg=cfg,
|
719 |
+
)
|
720 |
+
ckpt_path = cfg.get("ckpt", "") # load weights of MiniGPT-4
|
721 |
+
if ckpt_path:
|
722 |
+
print("Load Minigpt-4-LLM Checkpoint: {}".format(ckpt_path))
|
723 |
+
ckpt = torch.load(ckpt_path, map_location="cpu")
|
724 |
+
msg = model.load_state_dict(ckpt['model'], strict=False)
|
725 |
+
# push the model to the hub with its metadata and config file
|
726 |
+
# model.push_to_hub("MiniGPT4-video-v2")
|
727 |
+
# video_config = minigpt4_video_config(cfg)
|
728 |
+
# video_config.save_pretrained("minigpt4_video_config")
|
729 |
+
# print("Save Minigpt-4-LLM Config: minigpt4_video_config")
|
730 |
+
# video_config.push_to_hub("MiniGPT4-video")
|
731 |
+
return model
|
732 |
+
|
733 |
+
|
734 |
+
def assign_imgs(batched_instruct_list, batched_img_embeds):
|
735 |
+
'''this function is used when the data is interleaved.
|
736 |
+
the interlevaed data is separated, and this function assign
|
737 |
+
corresponding image embeddings to each segment'''
|
738 |
+
if len(batched_img_embeds.shape) == 3:
|
739 |
+
batched_img_embeds = batched_img_embeds[:, None]
|
740 |
+
|
741 |
+
batched_assigned = []
|
742 |
+
|
743 |
+
for instruct_list, img_embeds in zip(batched_instruct_list, batched_img_embeds):
|
744 |
+
img_idx = 0
|
745 |
+
assigned_img = []
|
746 |
+
n_assigned = []
|
747 |
+
for instruct in instruct_list:
|
748 |
+
n_img = instruct.count('<ImageHere>')
|
749 |
+
if n_img > 0: # this instruction include images.
|
750 |
+
assigned_img.append(img_embeds[None, img_idx:img_idx+n_img])
|
751 |
+
img_idx += n_img
|
752 |
+
n_assigned.append(n_img)
|
753 |
+
else: # this instruction doesn't include images
|
754 |
+
assigned_img.append(None)
|
755 |
+
n_assigned.append(None)
|
756 |
+
batched_assigned.append(assigned_img)
|
757 |
+
|
758 |
+
return batched_assigned
|