File size: 4,703 Bytes
a04730e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
from __future__ import annotations

from dataclasses import dataclass

import torch
from torch import nn
from transformers import CLIPTextModel, CLIPTokenizer


@dataclass
class TextConditioningOutput:
    """
    Output of the CLIP text encoder.

    hidden_states:
        Token-level CLIP embeddings.
        Shape: [B, seq_len, hidden_dim]

    attention_mask:
        Token attention mask.
        Shape: [B, seq_len]

    pooled:
        Optional pooled text embedding.
        Shape: [B, hidden_dim]
    """

    hidden_states: torch.Tensor
    attention_mask: torch.Tensor
    pooled: torch.Tensor | None = None


class FrozenCLIPTextEncoder(nn.Module):
    """
    Frozen CLIP text encoder for latent diffusion conditioning
    model:
        openai/clip-vit-large-patch14

    This gives:
        context_dim = 768
        max_length = 77

    """

    def __init__(
        self,
        model_name: str = "openai/clip-vit-large-patch14",
        max_length: int = 77,
        freeze: bool = True,
        use_last_hidden_state: bool = True,
        cache_dir: str | None = None,
        local_files_only: bool = False,
    ):
        super().__init__()

        self.model_name = model_name
        self.max_length = max_length
        self.freeze = freeze
        self.use_last_hidden_state = use_last_hidden_state
        self.cache_dir = cache_dir
        self.local_files_only = local_files_only

        self.tokenizer = CLIPTokenizer.from_pretrained(
            model_name,
            cache_dir=cache_dir,
            local_files_only=local_files_only,
        )

        self.text_model = CLIPTextModel.from_pretrained(
            model_name,
            cache_dir=cache_dir,
            local_files_only=local_files_only,
        )

        if self.freeze:
            self.text_model.eval()
            for p in self.text_model.parameters():
                p.requires_grad = False

    @property
    def context_dim(self) -> int:
        return int(self.text_model.config.hidden_size)

    @property
    def vocab_size(self) -> int:
        return int(self.tokenizer.vocab_size)

    @property
    def pad_token_id(self) -> int:
        return int(self.tokenizer.pad_token_id)

    def train(self, mode: bool = True):
        """
        Keep CLIP frozen/eval even if parent model calls .train().
        """
        super().train(mode)

        if self.freeze:
            self.text_model.eval()

        return self

    def tokenize(
        self,
        captions: list[str] | tuple[str, ...],
        device: torch.device | str | None = None,
    ) -> dict[str, torch.Tensor]:
        """
        Tokenize captions into CLIP input tensors.
        """
        tokens = self.tokenizer(
            list(captions),
            padding="max_length",
            truncation=True,
            max_length=self.max_length,
            return_tensors="pt",
        )

        if device is not None:
            tokens = {
                key: value.to(device)
                for key, value in tokens.items()
            }

        return tokens

    def forward(
        self,
        captions: list[str] | tuple[str, ...],
        device: torch.device | str | None = None,
    ) -> TextConditioningOutput:
        """
        Produces CLIP textual embeddings as diffusion condition.
        """
        if device is None:
            device = next(self.text_model.parameters()).device

        tokens = self.tokenize(
            captions=captions,
            device=device,
        )

        with torch.no_grad() if self.freeze else torch.enable_grad():
            outputs = self.text_model(
                input_ids=tokens["input_ids"],
                attention_mask=tokens["attention_mask"],
                output_hidden_states=not self.use_last_hidden_state,
                return_dict=True,
            )

        if self.use_last_hidden_state:
            hidden_states = outputs.last_hidden_state
        else:
            # Penultimate layer is sometimes used in diffusion models.
            hidden_states = outputs.hidden_states[-2]

        pooled = outputs.pooler_output

        return TextConditioningOutput(
            hidden_states=hidden_states,
            attention_mask=tokens["attention_mask"],
            pooled=pooled,
        )

    @torch.no_grad()
    def encode(
        self,
        captions: list[str] | tuple[str, ...],
        device: torch.device | str | None = None,
    ) -> torch.Tensor:
        """
        Convenience function.

        Returns only token-level context:

            [B, seq_len, context_dim]
        """
        return self.forward(
            captions=captions,
            device=device,
        ).hidden_states