File size: 12,734 Bytes
56ef57e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
from transformers import AutoConfig, AutoModel, PretrainedConfig, CLIPTextConfig, CLIPVisionConfig, PreTrainedModel, CLIPTextModelWithProjection, CLIPVisionModelWithProjection
from transformers.utils import ModelOutput
import torch
import open_clip
from dataclasses import dataclass
import safetensors.torch
from peft import get_peft_config, get_peft_model, LoraConfig, TaskType
import os

HF_SAFE_WEIGHTS_NAME = "open_clip_model.safetensors"
HF_SAFE_WEIGHTS_NAME_PRIOR = "prior_model.safetensors"

@dataclass
class PriorTransformerOutput(ModelOutput):
    """
    The output of [`PriorTransformer`].

    Args:
        predicted_image_embedding (`torch.FloatTensor` of shape `(batch_size, embedding_dim)`):
            The predicted CLIP image embedding conditioned on the CLIP text embedding input.
    """

    predicted_image_embedding: torch.FloatTensor

@dataclass
class TextEncoderOutput(ModelOutput):
    """
    Output class for CLIPTextEncoderOnly model to store the outputs in a Hugging Face transformer style.

    Attributes:
        prompt_embeds (torch.Tensor): The embeddings of the input prompts.
        last_hidden_states (torch.Tensor): The last hidden states from the model.
    """
    text_embeds: torch.FloatTensor = None
    last_hidden_state: torch.FloatTensor = None

class CLIPTextEncoderOnlyConfig(CLIPTextConfig):
    model_type = "clip_custom_text_model"

    def __init__(self, model_name: str = None, pretrained: bool = True, frozen: bool = False, lora: dict = None, **kwargs):
        self.model_name = model_name
        self.pretrained = pretrained
        self.frozen = frozen
        self.lora = lora
        super().__init__(**kwargs)

class CLIPTextEncoderOnly(PreTrainedModel):
    config_class = CLIPTextEncoderOnlyConfig

    def __init__(self, config):
        """
        Initializes the Hugging Face text encoder for CLIP model, inheriting from PreTrainedModel.

        :param model_name: The name or path of the pretrained model.
        :param pretrained: Whether to load the pretrained weights.
        """
        super().__init__(config)
        
        if config.pretrained:
            self.model = CLIPTextModelWithProjection.from_pretrained(config.model_name)
        else:
            base_cfg = CLIPTextConfig.from_pretrained(config.model_name)
            self.model = CLIPTextModelWithProjection(base_cfg)

        if config.lora:
            l_config = LoraConfig(
                r=config.lora.lora_r,
                lora_alpha=config.lora.lora_alpha,
                target_modules=[
                    "k_proj", 
                    "v_proj", 
                    "q_proj", 
                    "out_proj", 
                    "fc1",
                    "fc2",
                    "visual_projection", 
                    "text_projection"
                ],
                lora_dropout=config.lora.lora_dropout,
                bias="lora_only",
            )
            self.model = get_peft_model(self.model, l_config)
        

    def forward(self, input_ids, attention_mask=None, position_ids=None):
        """
        Forward pass of the model.

        :param input_ids: Indices of input sequence tokens in the vocabulary.
        :param attention_mask: Mask to avoid performing attention on padding token indices.
        :param token_type_ids: Segment token indices to indicate first and second portions of the inputs.
        :return: Outputs of the model.
        """
        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, output_hidden_states=True)
        return TextEncoderOutput(text_embeds=outputs.text_embeds, last_hidden_state=outputs.last_hidden_state)
    
class CustomTextEncoderOnly(PreTrainedModel):
    def __init__(self, model_name: str, output_hidden_size: int, pretrained: bool = True, frozen: bool = True, last_hidden_state: bool = False, lora: dict = None):
        """
        Initializes the Hugging Face text encoder for CLIP model, inheriting from PreTrainedModel.

        :param model_name: The name or path of the pretrained model.
        :param pretrained: Whether to load the pretrained weights.
        """
        config = AutoModel.from_pretrained(model_name).config
        super().__init__(config)
        self.last_hidden_state = last_hidden_state

        if pretrained:
            self.model = AutoModel.from_pretrained(model_name)
            if frozen:
                for param in self.model.parameters():
                    param.requires_grad = False
        else:
            self.model = AutoModel(config)

        self.fc1 = torch.nn.Linear(self.model.config.hidden_size, output_hidden_size)
        if last_hidden_state:
            self.fc2 = torch.nn.Linear(self.model.config.hidden_size, output_hidden_size)

        if lora:
            l_config = LoraConfig(
                task_type=TaskType.FEATURE_EXTRACTION,
                r=lora.lora_r,
                lora_alpha=lora.lora_alpha,
                lora_dropout=lora.lora_dropout,
                bias="lora_only",
            )
            self.model = get_peft_model(self.model, l_config)

    def forward(self, input_ids, attention_mask=None, token_type_ids=None):
        """
        Forward pass of the model.

        :param input_ids: Indices of input sequence tokens in the vocabulary.
        :param attention_mask: Mask to avoid performing attention on padding token indices.
        :param token_type_ids: Segment token indices to indicate first and second portions of the inputs.
        :return: Outputs of the model.
        """
        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, output_hidden_states=True)
        text_embeds = self.fc1(outputs[1])
        last_hidden_state = None
        if self.last_hidden_state:
            last_hidden_state = self.fc2(outputs[0])
        else:
            last_hidden_state = outputs[0]
        return TextEncoderOutput(text_embeds=text_embeds, last_hidden_state=last_hidden_state)

class CLIPVisionEncoderOnlyConfig(PretrainedConfig):
    model_type = "clip_custom_vision_model"

    def __init__(self, model_name: str = None, pretrained: bool = True, frozen: bool = False, lora: dict = None, **kwargs):
        self.model_name = model_name
        self.pretrained = pretrained
        self.frozen = frozen
        self.lora = lora
        super().__init__(**kwargs)

class CLIPVisionEncoderOnly(PreTrainedModel):
    config_class = CLIPVisionEncoderOnlyConfig

    def __init__(self, config):
        """
        Initializes the Hugging Face text encoder for CLIP model, inheriting from PreTrainedModel.

        :param model_name: The name or path of the pretrained model.
        :param pretrained: Whether to load the pretrained weights.
        """
        super().__init__(config)
        
        if config.pretrained:
            self.model = CLIPVisionModelWithProjection.from_pretrained(config.model_name)
        else:
            base_cfg = CLIPVisionConfig.from_pretrained(config.model_name)
            self.model = CLIPVisionModelWithProjection(base_cfg)

        if config.lora:
            l_config = LoraConfig(
                r=config.lora.lora_r,
                lora_alpha=config.lora.lora_alpha,
                target_modules=[
                    "k_proj", 
                    "v_proj", 
                    "q_proj", 
                    "out_proj", 
                    "fc1",
                    "fc2",
                    "visual_projection", 
                    "text_projection"
                ],
                lora_dropout=config.lora.lora_dropout,
                bias="lora_only",
            )
            self.model = get_peft_model(self.model, l_config)

    def forward(self, data):
        """
        Forward pass of the model.
        """
        return self.model(**data).image_embeds
    
    def parameters(self):
        return self.model.parameters()


class OpenCLIPVisionEncoderOnly(torch.nn.Module):
    def __init__(self, model_name: str, pretrained: bool = True, frozen: bool = False, lora: dict = None):
        """
        Initializes the Hugging Face text encoder for CLIP model, inheriting from PreTrainedModel.

        :param model_name: The name or path of the pretrained model.
        :param pretrained: Whether to load the pretrained weights.
        """
        super().__init__()
        if pretrained:
            model, _ = open_clip.create_model_from_pretrained(f"hf-hub:{model_name}")
            model = model.visual
        else:
            raise NotImplemented
        self.model = model

        if lora:
            l_config = LoraConfig(
                r=lora.lora_r,
                lora_alpha=lora.lora_alpha,
                target_modules=[
                    "k_proj", 
                    "v_proj", 
                    "q_proj", 
                    "out_proj", 
                    "fc1",
                    "fc2",
                    "visual_projection", 
                    "text_projection"
                ],
                lora_dropout=lora.lora_dropout,
                bias="lora_only",
            )
            self.model = get_peft_model(self.model, l_config)

    def forward(self, image):
        """
        Forward pass of the model.
        """
        return self.model(image)
    
    def save_pretrained(self, save_dir):
        tensors = self.model.state_dict()
        safetensors.torch.save_file(tensors, save_dir / HF_SAFE_WEIGHTS_NAME)

class CustomPriorModel(torch.nn.Module):
    def __init__(self, in_hidden_state, out_hidden_state):
        """
        Initializes the Hugging Face text encoder for CLIP model, inheriting from PreTrainedModel.

        :param model_name: The name or path of the pretrained model.
        :param pretrained: Whether to load the pretrained weights.
        """
        super().__init__()
        mid_hidden_state = max(in_hidden_state, out_hidden_state)

        self.fc1 = torch.nn.Linear(in_hidden_state*2, mid_hidden_state)
        self.relu = torch.nn.ReLU()
        self.fc2 = torch.nn.Linear(mid_hidden_state, out_hidden_state)
    
    def reinitialize_model(self):
        for name, param in self.named_parameters():
            if param.requires_grad:
                if len(param.shape) > 1:
                    torch.nn.init.xavier_uniform_(param)
                else:
                    if 'weight' in name:
                        torch.nn.init.normal_(param)
                    else:
                        torch.nn.init.zeros_(param)

    def forward(self, feats):
        """
        Forward pass of the model.
        """
        return PriorTransformerOutput(predicted_image_embedding=self.fc2(self.relu(self.fc1(feats))))
    
    def save_pretrained(self, save_dir):
        pass
        # tensors = self.state_dict()
        # safetensors.torch.save_file(tensors, os.path.join(save_dir, HF_SAFE_WEIGHTS_NAME_PRIOR))


def test_text_model(register=False, upload=False):
    # register the classes
    if register:
        AutoConfig.register("clip_custom_text_model", CLIPTextEncoderOnlyConfig)
        AutoModel.register(CLIPTextEncoderOnlyConfig, CLIPTextEncoderOnly)
        CLIPTextEncoderOnlyConfig.register_for_auto_class()
        CLIPTextEncoderOnly.register_for_auto_class("AutoModel")

    if upload:
        # Initialize the model
        model_name = "openai/clip-vit-base-patch32"
        pretrained=True 
        lora=None

        cfg = CLIPTextEncoderOnlyConfig(model_name=model_name, pretrained=pretrained, lora=lora)
        model = CLIPTextEncoderOnly(cfg)
        model.push_to_hub("test-text-hf-upload")

        model = CLIPTextEncoderOnly.from_pretrained("mpatel57/test-text-hf-upload", force_download=True)

def test_vision_model(register=False, upload=False):
    # register the classes
    if register:
        AutoConfig.register("clip_custom_vision_model", CLIPVisionEncoderOnlyConfig)
        AutoModel.register(CLIPVisionEncoderOnlyConfig, CLIPVisionEncoderOnly)
        CLIPVisionEncoderOnlyConfig.register_for_auto_class()
        CLIPVisionEncoderOnly.register_for_auto_class("AutoModel")

    if upload:
        # Initialize the model
        model_name = "openai/clip-vit-base-patch32"
        pretrained=True 
        lora=None

        cfg = CLIPVisionEncoderOnlyConfig(model_name=model_name, pretrained=pretrained, lora=lora)
        model = CLIPVisionEncoderOnly(cfg)
        model.push_to_hub("test-vision-hf-upload")

        model = CLIPVisionEncoderOnly.from_pretrained("mpatel57/test-vision-hf-upload", force_download=True)


if __name__ == "__main__":
    test_text_model(register=False, upload=True)
    test_vision_model(register=False, upload=True)