Create modeling_gpt2vision.py
Browse files- modeling_gpt2vision.py +171 -0
modeling_gpt2vision.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from transformers import PreTrainedModel, AutoTokenizer
|
4 |
+
from .configuration_gpt2vision import GPT2VisionConfig ,GPT2Config
|
5 |
+
from .modeling_gpt2 import GPT2LMHeadModel
|
6 |
+
from .vision_encoder import VisionEncoder
|
7 |
+
|
8 |
+
IMAGE_TOKEN = "<image>"
|
9 |
+
ANSWER_EOS = "<|endoftext|>"
|
10 |
+
|
11 |
+
def resize_token_embeds(model_name="openai-community/gpt2"):
|
12 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
13 |
+
tokenizer.add_special_tokens({"additional_special_tokens": [IMAGE_TOKEN]})
|
14 |
+
return tokenizer
|
15 |
+
|
16 |
+
tokenizer = resize_token_embeds()
|
17 |
+
|
18 |
+
print("tokenizer",tokenizer)
|
19 |
+
def create_labels(input_ids, tokenizer, attention_mask):
|
20 |
+
labels = input_ids.clone()
|
21 |
+
|
22 |
+
labels[attention_mask == 0] = -100
|
23 |
+
|
24 |
+
answer_start_tokens = tokenizer.encode("Answer:", add_special_tokens=False)
|
25 |
+
|
26 |
+
for i, seq in enumerate(input_ids):
|
27 |
+
# Find the start of the answer
|
28 |
+
answer_start = (seq == answer_start_tokens[0]).nonzero(as_tuple=True)[0]
|
29 |
+
if len(answer_start) > 0:
|
30 |
+
answer_start = answer_start[0]
|
31 |
+
if seq[answer_start:answer_start+len(answer_start_tokens)].tolist() == answer_start_tokens:
|
32 |
+
# Mask out everything before the answer
|
33 |
+
labels[i, :answer_start] = -100
|
34 |
+
|
35 |
+
# Find the end of the sequence (last non-padding token)
|
36 |
+
sequence_end = attention_mask[i].nonzero(as_tuple=True)[0][-1]
|
37 |
+
|
38 |
+
# Keep the last token (EOS) as part of the label
|
39 |
+
labels[i, sequence_end+1:] = -100
|
40 |
+
|
41 |
+
return labels
|
42 |
+
|
43 |
+
class MLP(nn.Module):
|
44 |
+
def __init__(self, in_features: int, hidden_features: int = None, out_features: int = None):
|
45 |
+
super().__init__()
|
46 |
+
out_features = out_features or in_features
|
47 |
+
hidden_features = hidden_features or in_features
|
48 |
+
self.fc1 = nn.Linear(in_features, hidden_features)
|
49 |
+
self.act = nn.GELU(approximate="tanh")
|
50 |
+
self.fc2 = nn.Linear(hidden_features, out_features)
|
51 |
+
self.dropout = nn.Dropout(p=0.1)
|
52 |
+
|
53 |
+
# Initialize weights
|
54 |
+
nn.init.xavier_normal_(self.fc1.weight)
|
55 |
+
nn.init.zeros_(self.fc1.bias)
|
56 |
+
nn.init.xavier_normal_(self.fc2.weight)
|
57 |
+
nn.init.zeros_(self.fc2.bias)
|
58 |
+
|
59 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
60 |
+
x = self.fc1(x)
|
61 |
+
x = self.act(x)
|
62 |
+
x = self.dropout(x)
|
63 |
+
x = self.fc2(x)
|
64 |
+
return x
|
65 |
+
|
66 |
+
|
67 |
+
class GPT2Vision(PreTrainedModel):
|
68 |
+
config_class = GPT2VisionConfig
|
69 |
+
|
70 |
+
def __init__(self, config):
|
71 |
+
super().__init__(config)
|
72 |
+
self.vision_encoder = VisionEncoder()
|
73 |
+
self.mlp = MLP(in_features=768, hidden_features=768 * 4, out_features=768)
|
74 |
+
|
75 |
+
self.language_model = GPT2LMHeadModel(config.gpt2_config)
|
76 |
+
|
77 |
+
self.language_model.resize_token_embeddings(len(tokenizer))
|
78 |
+
|
79 |
+
self.tokenizer = tokenizer
|
80 |
+
tokenizer.pad_token = tokenizer.eos_token
|
81 |
+
|
82 |
+
self.image_token_id = self.tokenizer.convert_tokens_to_ids(IMAGE_TOKEN)
|
83 |
+
|
84 |
+
@property
|
85 |
+
def device(self):
|
86 |
+
return next(self.language_model.parameters()).device
|
87 |
+
|
88 |
+
def freeze_model_components(self, freeze_vision=True, freeze_language=True,freeze_mlp=True):
|
89 |
+
for param in self.vision_encoder.parameters():
|
90 |
+
param.requires_grad = not freeze_vision
|
91 |
+
for param in self.language_model.parameters():
|
92 |
+
param.requires_grad = not freeze_language
|
93 |
+
for param in self.mlp.parameters():
|
94 |
+
param.requires_grad = not freeze_mlp
|
95 |
+
|
96 |
+
def tokenize_encode(self, batch, device):
|
97 |
+
text = batch['text']
|
98 |
+
images = batch['image']
|
99 |
+
|
100 |
+
if isinstance(text, str):
|
101 |
+
text = [text]
|
102 |
+
|
103 |
+
input_texts = [f"{IMAGE_TOKEN}{self.tokenizer.bos_token}{t}" for t in text]
|
104 |
+
text_inputs = self.tokenizer(
|
105 |
+
input_texts,
|
106 |
+
padding='max_length',
|
107 |
+
truncation=True,
|
108 |
+
max_length=384,
|
109 |
+
return_tensors="pt",
|
110 |
+
pad_to_multiple_of=8,
|
111 |
+
).to(device)
|
112 |
+
|
113 |
+
pixel_values = self.vision_encoder(images,device)
|
114 |
+
|
115 |
+
return {
|
116 |
+
"input_ids": text_inputs.input_ids,
|
117 |
+
"attention_mask": text_inputs.attention_mask,
|
118 |
+
"pixel_values": pixel_values
|
119 |
+
}
|
120 |
+
|
121 |
+
def preprocess_inputs(self, batch):
|
122 |
+
pixel_values = batch['pixel_values'].squeeze(1)
|
123 |
+
input_ids = batch['input_ids'].squeeze(1)
|
124 |
+
attention_mask = batch['attention_mask'].squeeze(1)
|
125 |
+
|
126 |
+
input_ids = input_ids.to(self.device)
|
127 |
+
attention_mask = attention_mask.to(self.device)
|
128 |
+
pixel_values = pixel_values.to(self.device)
|
129 |
+
|
130 |
+
labels = create_labels(input_ids, self.tokenizer, attention_mask)
|
131 |
+
labels = labels.to(self.device)
|
132 |
+
|
133 |
+
img_embs = self.mlp(pixel_values)
|
134 |
+
tok_embs = self.language_model.get_input_embeddings()(input_ids)
|
135 |
+
|
136 |
+
inputs_embeds = torch.cat((tok_embs[:, 0:1, :], img_embs, tok_embs[:, 1:, :]), dim=1)
|
137 |
+
|
138 |
+
img_attention = torch.ones((img_embs.size(0), img_embs.size(1)), dtype=torch.long, device=self.device)
|
139 |
+
attention_mask = torch.cat((attention_mask[:, 0:1], img_attention, attention_mask[:, 1:]), dim=1)
|
140 |
+
|
141 |
+
img_labels = torch.full((labels.size(0), img_embs.size(1)), fill_value=-100, dtype=torch.long, device=self.device)
|
142 |
+
labels = torch.cat((labels[:, 0:1], img_labels, labels[:, 1:]), dim=1)
|
143 |
+
return inputs_embeds, attention_mask, input_ids, labels
|
144 |
+
|
145 |
+
def forward(self, batch, **kwargs):
|
146 |
+
inputs_embeds, attention_mask, input_ids, labels = self.preprocess_inputs(batch)
|
147 |
+
outputs = self.language_model(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels)
|
148 |
+
|
149 |
+
|
150 |
+
return outputs
|
151 |
+
|
152 |
+
def generate(self, question, image, max_new_tokens=30, **kwargs):
|
153 |
+
prompt = prompt = f"Question: {question}\nAnswer:"
|
154 |
+
batch = {"image": [image], "text": prompt}
|
155 |
+
encoded_batch = self.tokenize_encode(batch, self.device)
|
156 |
+
inputs_embeds, attention_mask, input_ids, _ = self.preprocess_inputs(encoded_batch)
|
157 |
+
|
158 |
+
|
159 |
+
|
160 |
+
|
161 |
+
output_sequences = self.language_model.generate(
|
162 |
+
inputs_embeds=inputs_embeds,
|
163 |
+
attention_mask=attention_mask,
|
164 |
+
max_new_tokens=max_new_tokens,
|
165 |
+
pad_token_id=self.tokenizer.eos_token_id,
|
166 |
+
eos_token_id=self.tokenizer.eos_token_id,
|
167 |
+
**kwargs
|
168 |
+
)
|
169 |
+
|
170 |
+
output = self.tokenizer.decode(output_sequences[0], skip_special_tokens=True)
|
171 |
+
return output
|