Spaces:
Running
on
Zero
Running
on
Zero
Upload 2 files
Browse files
MeshAnything/models/meshanything_v2.py
ADDED
@@ -0,0 +1,162 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as nnf
|
3 |
+
from torch import nn
|
4 |
+
import random
|
5 |
+
from transformers import AutoModelForCausalLM
|
6 |
+
from MeshAnything.miche.encode import load_model
|
7 |
+
from MeshAnything.models.shape_opt import ShapeOPTConfig
|
8 |
+
from einops import repeat, reduce, rearrange, pack, unpack
|
9 |
+
|
10 |
+
class MeshAnythingV2(nn.Module):
|
11 |
+
def __init__(self):
|
12 |
+
super().__init__()
|
13 |
+
self.point_encoder = load_model(ckpt_path=None)
|
14 |
+
self.n_discrete_size = 128
|
15 |
+
self.max_seq_ratio = 0.70
|
16 |
+
self.face_per_token = 9
|
17 |
+
self.cond_length = 257
|
18 |
+
self.cond_dim = 768
|
19 |
+
self.pad_id = -1
|
20 |
+
self.n_max_triangles = 1600
|
21 |
+
self.max_length = int(self.n_max_triangles * self.face_per_token * self.max_seq_ratio + 3 + self.cond_length) # add 1
|
22 |
+
|
23 |
+
self.coor_continuous_range = (-0.5, 0.5)
|
24 |
+
|
25 |
+
self.config = ShapeOPTConfig.from_pretrained(
|
26 |
+
"facebook/opt-350m",
|
27 |
+
n_positions=self.max_length,
|
28 |
+
max_position_embeddings=self.max_length,
|
29 |
+
vocab_size=self.n_discrete_size + 4,
|
30 |
+
_attn_implementation="flash_attention_2"
|
31 |
+
)
|
32 |
+
|
33 |
+
self.bos_token_id = 0
|
34 |
+
self.eos_token_id = 1
|
35 |
+
self.pad_token_id = 2
|
36 |
+
|
37 |
+
self.config.bos_token_id = self.bos_token_id
|
38 |
+
self.config.eos_token_id = self.eos_token_id
|
39 |
+
self.config.pad_token_id = self.pad_token_id
|
40 |
+
self.config._attn_implementation="flash_attention_2"
|
41 |
+
self.config.n_discrete_size = self.n_discrete_size
|
42 |
+
self.config.face_per_token = self.face_per_token
|
43 |
+
self.config.cond_length = self.cond_length
|
44 |
+
|
45 |
+
if self.config.word_embed_proj_dim != self.config.hidden_size:
|
46 |
+
self.config.word_embed_proj_dim = self.config.hidden_size
|
47 |
+
self.transformer = AutoModelForCausalLM.from_config(
|
48 |
+
config=self.config, use_flash_attention_2 = True
|
49 |
+
)
|
50 |
+
self.transformer.to_bettertransformer()
|
51 |
+
|
52 |
+
self.cond_head_proj = nn.Linear(self.cond_dim, self.config.word_embed_proj_dim)
|
53 |
+
self.cond_proj = nn.Linear(self.cond_dim * 2, self.config.word_embed_proj_dim)
|
54 |
+
|
55 |
+
self.eval()
|
56 |
+
|
57 |
+
def adjacent_detokenize(self, input_ids):
|
58 |
+
input_ids = input_ids.reshape(input_ids.shape[0], -1) # B x L
|
59 |
+
batch_size = input_ids.shape[0]
|
60 |
+
continuous_coors = torch.zeros((batch_size, self.n_max_triangles * 3 * 10, 3), device=input_ids.device)
|
61 |
+
continuous_coors[...] = float('nan')
|
62 |
+
|
63 |
+
for i in range(batch_size):
|
64 |
+
cur_ids = input_ids[i]
|
65 |
+
coor_loop_check = 0
|
66 |
+
vertice_count = 0
|
67 |
+
continuous_coors[i, :3, :] = torch.tensor([[-0.1, 0.0, 0.1], [-0.1, 0.1, 0.2], [-0.3, 0.3, 0.2]],
|
68 |
+
device=input_ids.device)
|
69 |
+
for id in cur_ids:
|
70 |
+
if id == self.pad_id:
|
71 |
+
break
|
72 |
+
elif id == self.n_discrete_size:
|
73 |
+
if coor_loop_check < 9:
|
74 |
+
break
|
75 |
+
if coor_loop_check % 3 !=0:
|
76 |
+
break
|
77 |
+
coor_loop_check = 0
|
78 |
+
else:
|
79 |
+
|
80 |
+
if coor_loop_check % 3 == 0 and coor_loop_check >= 9:
|
81 |
+
continuous_coors[i, vertice_count] = continuous_coors[i, vertice_count-2]
|
82 |
+
continuous_coors[i, vertice_count+1] = continuous_coors[i, vertice_count-1]
|
83 |
+
vertice_count += 2
|
84 |
+
continuous_coors[i, vertice_count, coor_loop_check % 3] = undiscretize(id, self.coor_continuous_range[0], self.coor_continuous_range[1], self.n_discrete_size)
|
85 |
+
if coor_loop_check % 3 == 2:
|
86 |
+
vertice_count += 1
|
87 |
+
coor_loop_check += 1
|
88 |
+
|
89 |
+
continuous_coors = rearrange(continuous_coors, 'b (nf nv) c -> b nf nv c', nv=3, c=3)
|
90 |
+
|
91 |
+
return continuous_coors # b, nf, 3, 3
|
92 |
+
|
93 |
+
|
94 |
+
def forward(self, data_dict: dict, is_eval: bool = False) -> dict:
|
95 |
+
if not is_eval:
|
96 |
+
return self.train_one_step(data_dict)
|
97 |
+
else:
|
98 |
+
return self.generate(data_dict)
|
99 |
+
|
100 |
+
def process_point_feature(self, point_feature):
|
101 |
+
encode_feature = torch.zeros(point_feature.shape[0], self.cond_length, self.config.word_embed_proj_dim,
|
102 |
+
device=self.cond_head_proj.weight.device, dtype=self.cond_head_proj.weight.dtype)
|
103 |
+
encode_feature[:, 0] = self.cond_head_proj(point_feature[:, 0])
|
104 |
+
shape_latents = self.point_encoder.to_shape_latents(point_feature[:, 1:])
|
105 |
+
encode_feature[:, 1:] = self.cond_proj(torch.cat([point_feature[:, 1:], shape_latents], dim=-1))
|
106 |
+
|
107 |
+
return encode_feature
|
108 |
+
|
109 |
+
@torch.no_grad()
|
110 |
+
def forward(self, pc_normal, sampling=False) -> dict:
|
111 |
+
batch_size = pc_normal.shape[0]
|
112 |
+
point_feature = self.point_encoder.encode_latents(pc_normal)
|
113 |
+
processed_point_feature = self.process_point_feature(point_feature)
|
114 |
+
generate_length = self.max_length - self.cond_length
|
115 |
+
net_device = next(self.parameters()).device
|
116 |
+
outputs = torch.ones(batch_size, generate_length).long().to(net_device) * self.eos_token_id
|
117 |
+
# batch x ntokens
|
118 |
+
if not sampling:
|
119 |
+
results = self.transformer.generate(
|
120 |
+
inputs_embeds=processed_point_feature,
|
121 |
+
max_new_tokens=generate_length, # all faces plus two
|
122 |
+
num_beams=1,
|
123 |
+
bos_token_id=self.bos_token_id,
|
124 |
+
eos_token_id=self.eos_token_id,
|
125 |
+
pad_token_id=self.pad_token_id,
|
126 |
+
)
|
127 |
+
else:
|
128 |
+
results = self.transformer.generate(
|
129 |
+
inputs_embeds = processed_point_feature,
|
130 |
+
max_new_tokens = generate_length, # all faces plus two
|
131 |
+
do_sample=True,
|
132 |
+
top_k=50,
|
133 |
+
top_p=0.95,
|
134 |
+
bos_token_id = self.bos_token_id,
|
135 |
+
eos_token_id = self.eos_token_id,
|
136 |
+
pad_token_id = self.pad_token_id,
|
137 |
+
)
|
138 |
+
assert results.shape[1] <= generate_length # B x ID bos is not included since it's predicted
|
139 |
+
outputs[:, :results.shape[1]] = results
|
140 |
+
# batch x ntokens ====> batch x ntokens x D
|
141 |
+
outputs = outputs[:, 1: -1]
|
142 |
+
|
143 |
+
outputs[outputs == self.bos_token_id] = self.pad_id
|
144 |
+
outputs[outputs == self.eos_token_id] = self.pad_id
|
145 |
+
outputs[outputs == self.pad_token_id] = self.pad_id
|
146 |
+
|
147 |
+
outputs[outputs != self.pad_id] -= 3
|
148 |
+
gen_mesh = self.adjacent_detokenize(outputs)
|
149 |
+
|
150 |
+
return gen_mesh
|
151 |
+
|
152 |
+
def undiscretize(
|
153 |
+
t,
|
154 |
+
low,#-0.5
|
155 |
+
high,# 0.5
|
156 |
+
num_discrete
|
157 |
+
):
|
158 |
+
t = t.float() #[0, num_discrete-1]
|
159 |
+
|
160 |
+
t /= num_discrete # 0<=t<1
|
161 |
+
t = t * (high - low) + low # -0.5 <= t < 0.5
|
162 |
+
return t
|
MeshAnything/models/shape_opt.py
CHANGED
@@ -8,9 +8,8 @@ from transformers.modeling_outputs import (
|
|
8 |
import torch
|
9 |
from torch import nn
|
10 |
from torch.nn import CrossEntropyLoss
|
11 |
-
from transformers.utils import replace_return_docstrings
|
12 |
from transformers.modeling_outputs import BaseModelOutputWithPast
|
13 |
-
# from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
|
14 |
|
15 |
class ShapeOPTConfig(OPTConfig):
|
16 |
model_type = "shape_opt"
|
@@ -26,23 +25,6 @@ class ShapeOPT(OPTForCausalLM):
|
|
26 |
# Initialize weights and apply final processing
|
27 |
self.post_init()
|
28 |
|
29 |
-
def tie_weights(self):
|
30 |
-
"""
|
31 |
-
Tie the weights between the input embeddings and the output embeddings.
|
32 |
-
|
33 |
-
If the `torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning the
|
34 |
-
weights instead.
|
35 |
-
"""
|
36 |
-
if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
|
37 |
-
if hasattr(self, self.base_model_prefix):
|
38 |
-
self = getattr(self, self.base_model_prefix)
|
39 |
-
self._tie_encoder_decoder_weights(self.encoder, self.decoder, self.base_model_prefix)
|
40 |
-
|
41 |
-
for module in self.modules():
|
42 |
-
if hasattr(module, "_tie_weights"):
|
43 |
-
module._tie_weights()
|
44 |
-
|
45 |
-
|
46 |
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="OPTConfig")
|
47 |
def forward(
|
48 |
self,
|
@@ -140,7 +122,7 @@ class ShapeOPT(OPTForCausalLM):
|
|
140 |
|
141 |
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
142 |
outputs = self.model.decoder(
|
143 |
-
input_ids=input_ids,
|
144 |
face_ids = face_ids,
|
145 |
attention_mask=attention_mask,
|
146 |
head_mask=head_mask,
|
@@ -195,28 +177,18 @@ class ShapeOPTDecoder(OPTDecoder):
|
|
195 |
self.padding_idx = config.pad_token_id
|
196 |
self.max_target_positions = config.max_position_embeddings
|
197 |
self.vocab_size = config.vocab_size
|
198 |
-
|
199 |
-
self.embed_tokens = nn.Embedding(config.vocab_size, config.word_embed_proj_dim, self.padding_idx) # not used
|
200 |
self.hidden_size = config.hidden_size
|
201 |
self.word_embed_proj_dim = config.word_embed_proj_dim
|
202 |
-
self.
|
203 |
-
self.input_layer = nn.Linear(config.quantize_codebook_dim, config.word_embed_proj_dim)
|
204 |
|
205 |
self.embed_positions = OPTLearnedPositionalEmbedding(config.max_position_embeddings, config.hidden_size)
|
206 |
-
self.token_embed_positions =
|
|
|
207 |
self.face_per_token = config.face_per_token
|
208 |
self.cond_length = config.cond_length
|
209 |
self.cond_embed = nn.Embedding(2, config.word_embed_proj_dim)
|
210 |
|
211 |
-
if config.word_embed_proj_dim != config.hidden_size:
|
212 |
-
self.project_out = nn.Linear(config.hidden_size, config.word_embed_proj_dim, bias=False)
|
213 |
-
else:
|
214 |
-
self.project_out = None
|
215 |
-
|
216 |
-
if config.word_embed_proj_dim != config.hidden_size:
|
217 |
-
self.project_in = nn.Linear(config.word_embed_proj_dim, config.hidden_size, bias=False)
|
218 |
-
else:
|
219 |
-
self.project_in = None
|
220 |
# Note that the only purpose of `config._remove_final_layer_norm` is to keep backward compatibility
|
221 |
# with checkpoints that have been fine-tuned before transformers v4.20.1
|
222 |
# see https://github.com/facebookresearch/metaseq/pull/164
|
@@ -234,17 +206,6 @@ class ShapeOPTDecoder(OPTDecoder):
|
|
234 |
# Initialize weights and apply final processing
|
235 |
self.post_init()
|
236 |
|
237 |
-
def embed_with_vae(self, input_ids):
|
238 |
-
inputs_embeds = repeat(torch.zeros(input_ids.shape, device=input_ids.device), 'b n -> b n d',
|
239 |
-
d=self.word_embed_proj_dim).clone().detach()
|
240 |
-
idx_in_extra = torch.isin(input_ids, torch.LongTensor([0, 1, 2]).to(input_ids.device))
|
241 |
-
inputs_embeds[idx_in_extra] += self.extra_embeds(input_ids[idx_in_extra])
|
242 |
-
self.quantize_codebooks = self.quantize_codebooks.to(input_ids.device)
|
243 |
-
inputs_embeds[~idx_in_extra] += self.input_layer(self.quantize_codebooks[0][input_ids[~idx_in_extra] - 3])
|
244 |
-
|
245 |
-
return inputs_embeds
|
246 |
-
|
247 |
-
|
248 |
def forward(
|
249 |
self,
|
250 |
input_ids: torch.LongTensor = None,
|
@@ -315,11 +276,13 @@ class ShapeOPTDecoder(OPTDecoder):
|
|
315 |
|
316 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
317 |
# Transformer Decoder
|
318 |
-
if input_ids is not None:
|
|
|
|
|
|
|
319 |
input_shape = input_ids.size()
|
320 |
input_ids = input_ids.view(-1, input_shape[-1])
|
321 |
-
inputs_embeds = self.
|
322 |
-
|
323 |
face_embeds = self.token_embed_positions(attention_mask[:, self.cond_length:], face_ids, input_ids,
|
324 |
self.face_per_token)
|
325 |
inputs_embeds += face_embeds
|
@@ -329,7 +292,8 @@ class ShapeOPTDecoder(OPTDecoder):
|
|
329 |
|
330 |
elif inputs_embeds is not None:
|
331 |
# assert self.cond and not self.training
|
332 |
-
|
|
|
333 |
total_length = inputs_embeds.shape[1] # B x length x embeding
|
334 |
cond_embed_query = torch.zeros((inputs_embeds.shape[0], total_length), device=inputs_embeds.device,
|
335 |
dtype=inputs_embeds.dtype).long()
|
@@ -357,9 +321,6 @@ class ShapeOPTDecoder(OPTDecoder):
|
|
357 |
|
358 |
pos_embeds = self.embed_positions(attention_mask, past_key_values_length)
|
359 |
|
360 |
-
if self.project_in is not None:
|
361 |
-
inputs_embeds = self.project_in(inputs_embeds)
|
362 |
-
|
363 |
hidden_states = inputs_embeds + pos_embeds
|
364 |
|
365 |
# decoder layers
|
@@ -419,9 +380,6 @@ class ShapeOPTDecoder(OPTDecoder):
|
|
419 |
if self.final_layer_norm is not None:
|
420 |
hidden_states = self.final_layer_norm(hidden_states)
|
421 |
|
422 |
-
if self.project_out is not None:
|
423 |
-
hidden_states = self.project_out(hidden_states)
|
424 |
-
|
425 |
# add hidden states from the last decoder layer
|
426 |
if output_hidden_states:
|
427 |
all_hidden_states += (hidden_states,)
|
@@ -436,6 +394,56 @@ class ShapeOPTDecoder(OPTDecoder):
|
|
436 |
attentions=all_self_attns,
|
437 |
)
|
438 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
439 |
class OPTFacePositionalEmbedding(nn.Embedding):
|
440 |
"""
|
441 |
This module learns positional embeddings up to a fixed maximum size.
|
|
|
8 |
import torch
|
9 |
from torch import nn
|
10 |
from torch.nn import CrossEntropyLoss
|
11 |
+
from transformers.utils import replace_return_docstrings
|
12 |
from transformers.modeling_outputs import BaseModelOutputWithPast
|
|
|
13 |
|
14 |
class ShapeOPTConfig(OPTConfig):
|
15 |
model_type = "shape_opt"
|
|
|
25 |
# Initialize weights and apply final processing
|
26 |
self.post_init()
|
27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="OPTConfig")
|
29 |
def forward(
|
30 |
self,
|
|
|
122 |
|
123 |
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
124 |
outputs = self.model.decoder(
|
125 |
+
input_ids = input_ids,
|
126 |
face_ids = face_ids,
|
127 |
attention_mask=attention_mask,
|
128 |
head_mask=head_mask,
|
|
|
177 |
self.padding_idx = config.pad_token_id
|
178 |
self.max_target_positions = config.max_position_embeddings
|
179 |
self.vocab_size = config.vocab_size
|
180 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.word_embed_proj_dim, self.padding_idx)
|
|
|
181 |
self.hidden_size = config.hidden_size
|
182 |
self.word_embed_proj_dim = config.word_embed_proj_dim
|
183 |
+
self.n_discrete_size = config.n_discrete_size
|
|
|
184 |
|
185 |
self.embed_positions = OPTLearnedPositionalEmbedding(config.max_position_embeddings, config.hidden_size)
|
186 |
+
self.token_embed_positions = OPTLoopEmbedding(10, config.word_embed_proj_dim, self.n_discrete_size) #padding_idx=self.padding_idx)
|
187 |
+
|
188 |
self.face_per_token = config.face_per_token
|
189 |
self.cond_length = config.cond_length
|
190 |
self.cond_embed = nn.Embedding(2, config.word_embed_proj_dim)
|
191 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
# Note that the only purpose of `config._remove_final_layer_norm` is to keep backward compatibility
|
193 |
# with checkpoints that have been fine-tuned before transformers v4.20.1
|
194 |
# see https://github.com/facebookresearch/metaseq/pull/164
|
|
|
206 |
# Initialize weights and apply final processing
|
207 |
self.post_init()
|
208 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
209 |
def forward(
|
210 |
self,
|
211 |
input_ids: torch.LongTensor = None,
|
|
|
276 |
|
277 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
278 |
# Transformer Decoder
|
279 |
+
if input_ids is not None and inputs_embeds is not None: # when train and first generate
|
280 |
+
assert False
|
281 |
+
elif input_ids is not None:
|
282 |
+
assert not self.training
|
283 |
input_shape = input_ids.size()
|
284 |
input_ids = input_ids.view(-1, input_shape[-1])
|
285 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
|
|
286 |
face_embeds = self.token_embed_positions(attention_mask[:, self.cond_length:], face_ids, input_ids,
|
287 |
self.face_per_token)
|
288 |
inputs_embeds += face_embeds
|
|
|
292 |
|
293 |
elif inputs_embeds is not None:
|
294 |
# assert self.cond and not self.training
|
295 |
+
assert not self.training
|
296 |
+
self.token_embed_positions.init_state(inputs_embeds)
|
297 |
total_length = inputs_embeds.shape[1] # B x length x embeding
|
298 |
cond_embed_query = torch.zeros((inputs_embeds.shape[0], total_length), device=inputs_embeds.device,
|
299 |
dtype=inputs_embeds.dtype).long()
|
|
|
321 |
|
322 |
pos_embeds = self.embed_positions(attention_mask, past_key_values_length)
|
323 |
|
|
|
|
|
|
|
324 |
hidden_states = inputs_embeds + pos_embeds
|
325 |
|
326 |
# decoder layers
|
|
|
380 |
if self.final_layer_norm is not None:
|
381 |
hidden_states = self.final_layer_norm(hidden_states)
|
382 |
|
|
|
|
|
|
|
383 |
# add hidden states from the last decoder layer
|
384 |
if output_hidden_states:
|
385 |
all_hidden_states += (hidden_states,)
|
|
|
394 |
attentions=all_self_attns,
|
395 |
)
|
396 |
|
397 |
+
class OPTLoopEmbedding(nn.Embedding):
|
398 |
+
"""
|
399 |
+
This module learns positional embeddings up to a fixed maximum size.
|
400 |
+
"""
|
401 |
+
|
402 |
+
def __init__(self, num_embeddings: int, embedding_dim: int, n_discrete_size: int):
|
403 |
+
super().__init__(num_embeddings, embedding_dim)
|
404 |
+
self.state = None
|
405 |
+
self.loop_state = None
|
406 |
+
self.n_discrete_size = n_discrete_size + 3 # for padding
|
407 |
+
|
408 |
+
def forward(self, attention_mask=None, face_ids = None, input_ids = None, face_per_token = None):
|
409 |
+
"""`input_ids_shape` is expected to be [bsz x seqlen]."""
|
410 |
+
if face_ids is not None:
|
411 |
+
return super().forward(face_ids)
|
412 |
+
|
413 |
+
assert input_ids.shape[1] == 1, "Only one token is allowed for loop embedding"
|
414 |
+
assert self.state is not None, "State is not initialized"
|
415 |
+
# zero as beginning
|
416 |
+
batch_size = input_ids.shape[0]
|
417 |
+
face_ids = input_ids.clone().detach()
|
418 |
+
|
419 |
+
for cur_batch_index in range(batch_size):
|
420 |
+
cur_ids = input_ids[cur_batch_index]
|
421 |
+
|
422 |
+
idx_in_extra = torch.isin(cur_ids, torch.LongTensor([0, 1, 2]).to(input_ids.device))
|
423 |
+
if idx_in_extra:
|
424 |
+
self.state[cur_batch_index] = 9 # init
|
425 |
+
self.loop_state[cur_batch_index] = 0
|
426 |
+
else:
|
427 |
+
if cur_ids == self.n_discrete_size:
|
428 |
+
face_ids[cur_batch_index] = 3
|
429 |
+
self.state[cur_batch_index] = 9 # init
|
430 |
+
self.loop_state[cur_batch_index] = 0
|
431 |
+
else:
|
432 |
+
if self.state[cur_batch_index] == 0:
|
433 |
+
face_ids[cur_batch_index] = 7 + self.loop_state[cur_batch_index] % 3
|
434 |
+
else:
|
435 |
+
self.state[cur_batch_index] -= 1
|
436 |
+
face_ids[cur_batch_index] = 4 + self.loop_state[cur_batch_index] % 3
|
437 |
+
self.loop_state[cur_batch_index] += 1
|
438 |
+
|
439 |
+
return super().forward(face_ids)
|
440 |
+
|
441 |
+
def init_state(self, template_tensor):
|
442 |
+
batch_size = template_tensor.shape[0]
|
443 |
+
self.state = torch.zeros((batch_size, 1), dtype=torch.long, device=template_tensor.device)
|
444 |
+
self.state[...] = 9
|
445 |
+
self.loop_state = torch.zeros((batch_size, 1), dtype=torch.long, device=template_tensor.device)
|
446 |
+
|
447 |
class OPTFacePositionalEmbedding(nn.Embedding):
|
448 |
"""
|
449 |
This module learns positional embeddings up to a fixed maximum size.
|