Idan0405 commited on
Commit
1cc900c
1 Parent(s): 0707df9

Upload ClipMDModel.py

Browse files
Files changed (1) hide show
  1. ClipMDModel.py +138 -0
ClipMDModel.py ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import CLIPModel
2
+ import torch
3
+ from typing import Optional, Tuple
4
+
5
+
6
+ def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
7
+ return torch.nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
8
+
9
+
10
+ def clip_loss(logits_per_text: torch.Tensor) -> torch.Tensor:
11
+ caption_loss = contrastive_loss(logits_per_text)
12
+ image_loss = contrastive_loss(logits_per_text.T)
13
+ return (caption_loss + image_loss) / 2.0
14
+
15
+
16
+ class ClipMDModel(CLIPModel):
17
+
18
+ def embed_text(self,
19
+ input_ids:torch.LongTensor,
20
+ attention_mask:torch.LongTensor,
21
+ output_attentions: Optional[bool] = None,
22
+ output_hidden_states: Optional[bool] = None,
23
+ position_ids: Optional[torch.LongTensor] = None,
24
+ ):
25
+ """
26
+ :param input_ids: tokenized text from CLIPProcessor.
27
+ :param attention_mask: attention mask from CLIPProcessor.
28
+ :return: text embeddings of input_ids (tokens longer then 77 tokens
29
+ is embeded using a sliding window and pooling).
30
+ """
31
+ tokens = []
32
+ masks = []
33
+ pos = []
34
+
35
+ for i in range(input_ids.size()[0]):
36
+ ten = input_ids[i]
37
+ mask = attention_mask[i]
38
+ mask = mask[mask.nonzero().flatten()]
39
+ ten = ten[:mask.size()[0]]
40
+
41
+ if not pos:
42
+ pos.append([0, 0])
43
+ else:
44
+ pos.append([pos[-1][1], pos[-1][1]])
45
+
46
+ #spliting tokenized text into input sized chunks with an overlapping window.
47
+ if ten.size()[0]>77:
48
+ tokens.append(ten.unfold(dimension = 0,size = 77, step = 70))
49
+ masks.append(mask.unfold(dimension = 0,size = 77, step = 70))
50
+
51
+ pos[-1][1]+=tokens[-1].size()[0]
52
+
53
+ ten=ten[tokens[-1].size()[0]*70:]
54
+ mask=mask[tokens[-1].size()[0]*70:]
55
+
56
+ if ten.size()[0] > 0:
57
+ new_mask = torch.zeros((1, 77)).to(self.device)
58
+ new_mask[:, 0:mask.size()[0]] = mask
59
+
60
+ new_ten = torch.full((1, 77), 49407).to(self.device)
61
+ new_ten[:, 0:ten.size()[0]] = ten
62
+
63
+ tokens.append(new_ten)
64
+ masks.append(new_mask)
65
+ pos[-1][1] += 1
66
+ #encoding the tokenized text
67
+ embedded = self.get_text_features(input_ids=torch.cat(tokens, 0),
68
+ attention_mask=torch.cat(masks, 0),
69
+ output_attentions=output_attentions,
70
+ output_hidden_states=output_hidden_states,
71
+ position_ids=position_ids,
72
+ )
73
+
74
+ #pooling the embeddings of segments that came from the same original text
75
+ embeddings = []
76
+ for p in pos:
77
+ if p[1] - p[0] == 1:
78
+ embeddings.append(embedded[p[0]].unsqueeze(0))
79
+ else:
80
+ embeddings.append(torch.mean(embedded[p[0]:p[1]], dim=0).unsqueeze(0))
81
+
82
+ return torch.cat(embeddings, 0)
83
+
84
+ def forward(self,
85
+ input_ids: Optional[torch.LongTensor] = None,
86
+ pixel_values: Optional[torch.FloatTensor] = None,
87
+ attention_mask: Optional[torch.Tensor] = None,
88
+ position_ids: Optional[torch.LongTensor] = None,
89
+ return_loss: Optional[bool] = None,
90
+ output_attentions: Optional[bool] = None,
91
+ output_hidden_states: Optional[bool] = None,
92
+ return_dict: Optional[bool] = None,
93
+ ) -> Tuple:
94
+ """
95
+ :param input_ids: tokenized text from CLIPProcessor.
96
+ :param attention_mask: attention mask from CLIPProcessor.
97
+ :param pixel_values: pixel values from CLIPProcessor.
98
+ :param return_loss: boolean that indicates if loss should be returned
99
+ :return: image-caption cosine similarity as logits per image and per caption (also loss if return_loss is true)
100
+ """
101
+ # Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
102
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
103
+ output_hidden_states = (
104
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
105
+ )
106
+ return_dict = self.config.use_return_dict
107
+
108
+ #encoding the images
109
+ vision_outputs = self.vision_model(
110
+ pixel_values=pixel_values,
111
+ output_attentions=output_attentions,
112
+ output_hidden_states=output_hidden_states,
113
+ )
114
+ image_embeds = vision_outputs[1]
115
+ image_embeds = self.visual_projection(image_embeds)
116
+
117
+ #encoding the text captions
118
+ text_embeds =self.embed_text(input_ids=input_ids,
119
+ attention_mask=attention_mask,
120
+ output_attentions=output_attentions,
121
+ output_hidden_states=output_hidden_states,
122
+ position_ids=position_ids
123
+ )
124
+
125
+
126
+ # normalized features
127
+ image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
128
+ text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
129
+
130
+ # cosine similarity as logits
131
+ logit_scale = self.logit_scale.exp()
132
+ logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
133
+ logits_per_image = logits_per_text.T
134
+
135
+ if return_loss:
136
+ loss = clip_loss(logits_per_text)
137
+ return logits_per_image,logits_per_text,loss
138
+ return logits_per_image,logits_per_text