Skip to content

unitorch.models.blip¤

BlipProcessor¤

Bases: HfTextClassificationProcessor, HfTextGenerationProcessor, HfImageClassificationProcessor

Initializes the BlipProcessor.

Parameters:

Name Type Description Default
vocab_path str

The path to the vocabulary file.

 required
vision_config_path str

The path to the vision configuration file.

 required
max_seq_length Optional[int]

The maximum sequence length for text inputs. Defaults to 128.

 128
max_gen_seq_length Optional[int]

The maximum sequence length for generated outputs. Defaults to 48.

 48
position_start_id Optional[int]

The position start ID. Defaults to 0.

 0
Source code in src/unitorch/models/blip/processing.py 
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
def __init__(
    self,
    vocab_path: str,
    vision_config_path: str,
    max_seq_length: Optional[int] = 128,
    max_gen_seq_length: Optional[int] = 48,
    position_start_id: Optional[int] = 0,
):
    """
    Initializes the BlipProcessor.

    Args:
        vocab_path (str): The path to the vocabulary file.
        vision_config_path (str): The path to the vision configuration file.
        max_seq_length (Optional[int]): The maximum sequence length for text inputs. Defaults to 128.
        max_gen_seq_length (Optional[int]): The maximum sequence length for generated outputs. Defaults to 48.
        position_start_id (Optional[int]): The position start ID. Defaults to 0.
    """
    vision_processor = BlipImageProcessor.from_json_file(vision_config_path)
    HfImageClassificationProcessor.__init__(self, vision_processor=vision_processor)

    tokenizer = BertTokenizer(
        vocab_file=vocab_path,
    )
    tokenizer.bos_token = tokenizer.cls_token
    tokenizer.bos_token_id = tokenizer.cls_token_id
    tokenizer.eos_token = tokenizer.sep_token
    tokenizer.eos_token_id = tokenizer.sep_token_id
    HfTextClassificationProcessor.__init__(
        self,
        tokenizer=tokenizer,
        max_seq_length=max_seq_length,
        source_type_id=0,
        target_type_id=0,
        position_start_id=position_start_id,
    )

    HfTextGenerationProcessor.__init__(
        self,
        tokenizer=tokenizer,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

classification ¤

classification(
    text: str,
    image: Union[Image, str],
    max_seq_length: Optional[int] = None,
) -> GenericOutputs

Performs classification using both text and image inputs.

Parameters:

Name Type Description Default
text str

The input text to classify.

 required
image Image

The input image to classify.

 required
max_seq_length Optional[int]

The maximum sequence length for the text. If None, the default value from initialization is used.

 None

Returns:

Name Type Description
GenericOutputs GenericOutputs

The outputs of the classification.
Source code in src/unitorch/models/blip/processing.py 
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
def classification(
    self,
    text: str,
    image: Union[Image.Image, str],
    max_seq_length: Optional[int] = None,
) -> GenericOutputs:
    """
    Performs classification using both text and image inputs.

    Args:
        text (str): The input text to classify.
        image (PIL.Image.Image): The input image to classify.
        max_seq_length (Optional[int]): The maximum sequence length for the text. If None, the default value from initialization is used.

    Returns:
        GenericOutputs: The outputs of the classification.
    """
    max_seq_length = pop_value(
        max_seq_length,
        self.max_seq_length,
    )

    text_outputs = self.text_classification(text, max_seq_length)
    pixel_outputs = self.image_classification(image)

    return GenericOutputs(
        input_ids=text_outputs.input_ids,
        attention_mask=text_outputs.attention_mask,
        position_ids=text_outputs.position_ids,
        pixel_values=pixel_outputs.pixel_values,
    )

generation ¤

generation(
    text: str,
    image: Union[Image, str],
    max_gen_seq_length: Optional[int] = None,
) -> GenericOutputs

Generate inputs, labels, and tokens for image to text generation.

Parameters:

Name Type Description Default
text str

The input text.

 required
image Image

The input image to caption.

 required
max_gen_seq_length int

Maximum generated sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs GenericOutputs

The generated input tokens, attention masks, label tokens, and attention masks.
Source code in src/unitorch/models/blip/processing.py 
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
def generation(
    self,
    text: str,
    image: Union[Image.Image, str],
    max_gen_seq_length: Optional[int] = None,
) -> GenericOutputs:
    """
    Generate inputs, labels, and tokens for image to text generation.

    Args:
        text (str): The input text.
        image (Image.Image): The input image to caption.
        max_gen_seq_length (int, optional): Maximum generated sequence length. Defaults to None.

    Returns:
        GenericOutputs: The generated input tokens, attention masks, label tokens, and attention masks.
    """

    max_gen_seq_length = pop_value(max_gen_seq_length, self.max_gen_seq_length)

    tokens = self.generation_inputs(text, max_gen_seq_length)
    pixels = self.image_classification(image)
    labels = self.generation_labels(text, max_gen_seq_length)

    return GenericOutputs(
        input_ids=tokens.input_ids,
        attention_mask=tokens.attention_mask,
        pixel_values=pixels.pixel_values,
        input_ids_label=labels.input_ids,
        attention_mask_label=labels.attention_mask,
    )

generation_inputs ¤

generation_inputs(
    text: str, max_seq_length: Optional[int] = None
) -> GenericOutputs

Generate inputs for text generation.

Parameters:

Name Type Description Default
text str

The input text.

 required
max_seq_length int

Maximum sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs GenericOutputs

The generated input tokens and attention mask.
Source code in src/unitorch/models/blip/processing.py 
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
def generation_inputs(
    self,
    text: str,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs:
    """
    Generate inputs for text generation.

    Args:
        text (str): The input text.
        max_seq_length (int, optional): Maximum sequence length. Defaults to None.

    Returns:
        GenericOutputs: The generated input tokens and attention mask.
    """
    outputs = HfTextGenerationProcessor.generation_inputs(
        self,
        text=text,
        max_seq_length=max_seq_length,
    )
    return GenericOutputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    )

generation_labels ¤

generation_labels(
    text: str, max_gen_seq_length: Optional[int] = None
) -> GenericOutputs

Generates labels for text generation based on the given input text.

Parameters:

Name Type Description Default
text str

The input text for generating labels.

 required
max_gen_seq_length Optional[int]

The maximum sequence length for the generated labels. If None, the default value from initialization is used.

 None

Returns:

Name Type Description
GenericOutputs GenericOutputs

The generated labels.
Source code in src/unitorch/models/blip/processing.py 
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
def generation_labels(
    self,
    text: str,
    max_gen_seq_length: Optional[int] = None,
) -> GenericOutputs:
    """
    Generates labels for text generation based on the given input text.

    Args:
        text (str): The input text for generating labels.
        max_gen_seq_length (Optional[int]): The maximum sequence length for the generated labels. If None, the default value from initialization is used.

    Returns:
        GenericOutputs: The generated labels.
    """
    outputs = HfTextGenerationProcessor.generation_labels(
        self,
        text=text,
        max_gen_seq_length=max_gen_seq_length,
    )
    return GenericOutputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    )

image_classification ¤

image_classification(
    image: Union[Image, str]
) -> GenericOutputs

Performs image classification on the given input image.

Parameters:

Name Type Description Default
image Image

The input image to classify.

 required

Returns:

Name Type Description
GenericOutputs GenericOutputs

The outputs of the image classification.
Source code in src/unitorch/models/blip/processing.py 
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
def image_classification(
    self,
    image: Union[Image.Image, str],
) -> GenericOutputs:
    """
    Performs image classification on the given input image.

    Args:
        image (PIL.Image.Image): The input image to classify.

    Returns:
        GenericOutputs: The outputs of the image classification.
    """
    outputs = HfImageClassificationProcessor.classification(
        self,
        image=image,
    )

    return GenericOutputs(
        pixel_values=outputs.pixel_values,
    )

text_classification ¤

text_classification(
    text: str, max_seq_length: Optional[int] = None
) -> GenericOutputs

Performs text classification on the given input text.

Parameters:

Name Type Description Default
text str

The input text to classify.

 required
max_seq_length Optional[int]

The maximum sequence length for the text. If None, the default value from initialization is used.

 None

Returns:

Name Type Description
GenericOutputs GenericOutputs

The outputs of the text classification.
Source code in src/unitorch/models/blip/processing.py 
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
def text_classification(
    self,
    text: str,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs:
    """
    Performs text classification on the given input text.

    Args:
        text (str): The input text to classify.
        max_seq_length (Optional[int]): The maximum sequence length for the text. If None, the default value from initialization is used.

    Returns:
        GenericOutputs: The outputs of the text classification.
    """
    outputs = HfTextClassificationProcessor.classification(
        self,
        text=text,
        max_seq_length=max_seq_length,
    )
    return GenericOutputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        position_ids=outputs.position_ids,
    )

BlipForPretrain¤

Bases: GenericModel

Initializes the BlipForPretrain model.

Parameters:

Name Type Description Default
config_path str

Path to the configuration file.

 required
projection_dim Optional[int]

Dimension of the projection. Defaults to 512.

 512
freeze_base_model Optional[bool]

Whether to freeze the base model. Defaults to True.

 True
gradient_checkpointing Optional[bool]

Whether to use gradient checkpointing. Defaults to False.

 False
use_all_gather Optional[bool]

Whether to use all_gather operation for distributed training. Defaults to True.

 True
Source code in src/unitorch/models/blip/modeling.py 
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
def __init__(
    self,
    config_path: str,
    projection_dim: Optional[int] = 512,
    freeze_base_model: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
    use_all_gather: Optional[bool] = True,
):
    """
    Initializes the BlipForPretrain model.

    Args:
        config_path (str): Path to the configuration file.
        projection_dim (Optional[int], optional): Dimension of the projection. Defaults to 512.
        freeze_base_model (Optional[bool], optional): Whether to freeze the base model. Defaults to True.
        gradient_checkpointing (Optional[bool], optional): Whether to use gradient checkpointing. Defaults to False.
        use_all_gather (Optional[bool], optional): Whether to use all_gather operation for distributed training. Defaults to True.
    """
    super().__init__()

    config = BlipConfig.from_json_file(config_path)
    text_config = config.text_config
    vision_config = config.vision_config
    text_config.gradient_checkpointing = gradient_checkpointing
    vision_config.gradient_checkpointing = gradient_checkpointing

    self.projection_dim = projection_dim
    self.use_all_gather = use_all_gather

    self.text_embed_dim = text_config.hidden_size
    self.vision_embed_dim = vision_config.hidden_size

    self.text_model = BlipTextModel(text_config)
    self.vision_model = BlipVisionModel(vision_config)

    self.visual_projection = nn.Linear(
        self.vision_embed_dim,
        self.projection_dim,
        bias=False,
    )
    self.text_projection = nn.Linear(
        self.text_embed_dim,
        self.projection_dim,
        bias=False,
    )
    self.logit_scale = nn.Parameter(torch.ones([]) * config.logit_scale_init_value)

    self.init_weights()

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

        for p in self.vision_model.parameters():
            p.requires_grad = False

    self.text_model.encoder.gradient_checkpointing = gradient_checkpointing
    self.vision_model.encoder.gradient_checkpointing = gradient_checkpointing

forward ¤

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Tensor,
    position_ids: Tensor,
)

Forward pass of the BlipForPretrain model.

Parameters:

Name Type Description Default
input_ids Tensor

Input token IDs.

 required
pixel_values Tensor

Pixel values of the images.

 required
attention_mask Tensor

Attention mask for the input.

 required
position_ids Tensor

Position IDs for the input.

 required

Returns:

Type Description
Tensor

Output loss for the pretraining task.
Source code in src/unitorch/models/blip/modeling.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: torch.Tensor,
    position_ids: torch.Tensor,
):
    """
    Forward pass of the BlipForPretrain model.

    Args:
        input_ids (torch.Tensor): Input token IDs.
        pixel_values (torch.Tensor): Pixel values of the images.
        attention_mask (torch.Tensor): Attention mask for the input.
        position_ids (torch.Tensor): Position IDs for the input.

    Returns:
        (torch.Tensor):Output loss for the pretraining task.
    """
    vision_outputs = self.vision_model(
        pixel_values=pixel_values,
    )

    text_outputs = self.text_model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )

    image_embeds = vision_outputs[1]
    image_embeds = self.visual_projection(image_embeds)

    text_embeds = text_outputs[1]
    text_embeds = self.text_projection(text_embeds)

    # normalized features
    image_embeds = image_embeds / image_embeds.norm(dim=-1, keepdim=True)
    text_embeds = text_embeds / text_embeds.norm(dim=-1, keepdim=True)

    logit_scale = self.logit_scale.exp()
    if self.use_all_gather and dist.is_initialized():
        text_embeds = self._all_gather(text_embeds)
        image_embeds = self._all_gather(image_embeds)
    logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
    return _blip_loss(logits_per_text)

BlipForClassification¤

Bases: Module

Initializes the BlipForClassification model.

Parameters:

Name Type Description Default
config_path str

Path to the configuration file.

 required
projection_dim Optional[int]

Dimension of the projection. Defaults to 512.

 512
num_classes Optional[int]

Number of classes for classification. Defaults to 1.

 1
freeze_base_model Optional[bool]

Whether to freeze the base model. Defaults to True.

 True
gradient_checkpointing Optional[bool]

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/blip/modeling.py 
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
def __init__(
    self,
    config_path: str,
    projection_dim: Optional[int] = 512,
    num_classes: Optional[int] = 1,
    freeze_base_model: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initializes the BlipForClassification model.

    Args:
        config_path (str): Path to the configuration file.
        projection_dim (Optional[int], optional): Dimension of the projection. Defaults to 512.
        num_classes (Optional[int], optional): Number of classes for classification. Defaults to 1.
        freeze_base_model (Optional[bool], optional): Whether to freeze the base model. Defaults to True.
        gradient_checkpointing (Optional[bool], optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()

    # Load the BLIP model configuration
    config = BlipConfig.from_json_file(config_path)
    text_config = config.text_config
    vision_config = config.vision_config

    # Set gradient checkpointing option
    text_config.gradient_checkpointing = gradient_checkpointing
    vision_config.gradient_checkpointing = gradient_checkpointing

    self.projection_dim = projection_dim

    self.text_embed_dim = text_config.hidden_size
    self.vision_embed_dim = vision_config.hidden_size

    # Initialize the text and vision models
    self.text_model = BlipTextModel(text_config)
    self.vision_model = BlipVisionModel(vision_config)

    # Projection layers for text and vision embeddings
    self.visual_projection = nn.Linear(
        self.vision_embed_dim,
        self.projection_dim,
        bias=False,
    )
    self.text_projection = nn.Linear(
        self.text_embed_dim,
        self.projection_dim,
        bias=False,
    )

    # Classifier layer
    self.classifier = nn.Linear(self.projection_dim * 2, num_classes)

    # Initialize the weights of the model
    self.init_weights()

    if freeze_base_model:
        # Freeze the parameters of the base models if specified
        for p in self.text_model.parameters():
            p.requires_grad = False

        for p in self.vision_model.parameters():
            p.requires_grad = False

    # Set gradient checkpointing option for encoders
    self.text_model.encoder.gradient_checkpointing = gradient_checkpointing
    self.vision_model.encoder.gradient_checkpointing = gradient_checkpointing

forward ¤

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Tensor,
    position_ids: Tensor,
)

Forward pass of the BlipForClassification model.

Parameters:

Name Type Description Default
input_ids Tensor

Input token IDs.

 required
pixel_values Tensor

Pixel values of the images.

 required
attention_mask Tensor

Attention mask for the input.

 required
position_ids Tensor

Position IDs for the input.

 required

Returns:

Type Description
Tensor

Output logits for classification.
Source code in src/unitorch/models/blip/modeling.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: torch.Tensor,
    position_ids: torch.Tensor,
):
    """
    Forward pass of the BlipForClassification model.

    Args:
        input_ids (torch.Tensor): Input token IDs.
        pixel_values (torch.Tensor): Pixel values of the images.
        attention_mask (torch.Tensor): Attention mask for the input.
        position_ids (torch.Tensor): Position IDs for the input.

    Returns:
        (torch.Tensor):Output logits for classification.
    """
    # Process the vision modality
    vision_outputs = self.vision_model(
        pixel_values=pixel_values,
    )

    # Process the text modality
    text_outputs = self.text_model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )

    # Project vision embeddings to the specified dimensionality
    image_embeds = vision_outputs[1]
    image_embeds = self.visual_projection(image_embeds)

    # Project text embeddings to the specified dimensionality
    text_embeds = text_outputs[1]
    text_embeds = self.text_projection(text_embeds)

    # Concatenate and classify the projected embeddings
    return self.classifier(F.relu(torch.cat([image_embeds, text_embeds], axis=1)))

BlipForTextClassification¤

Bases: GenericModel

Initializes the BlipForTextClassification model.

Parameters:

Name Type Description Default
config_path str

Path to the configuration file.

 required
projection_dim Optional[int]

Dimension of the projection. Defaults to 512.

 512
num_classes Optional[int]

Number of classes for classification. Defaults to 1.

 1
freeze_base_model Optional[bool]

Whether to freeze the base model. Defaults to True.

 True
gradient_checkpointing Optional[bool]

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/blip/modeling.py 
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
def __init__(
    self,
    config_path: str,
    projection_dim: Optional[int] = 512,
    num_classes: Optional[int] = 1,
    freeze_base_model: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initializes the BlipForTextClassification model.

    Args:
        config_path (str): Path to the configuration file.
        projection_dim (Optional[int], optional): Dimension of the projection. Defaults to 512.
        num_classes (Optional[int], optional): Number of classes for classification. Defaults to 1.
        freeze_base_model (Optional[bool], optional): Whether to freeze the base model. Defaults to True.
        gradient_checkpointing (Optional[bool], optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()

    # Load the BLIP model configuration
    config = BlipConfig.from_json_file(config_path)
    text_config = config.text_config
    text_config.gradient_checkpointing = gradient_checkpointing

    self.projection_dim = projection_dim
    self.text_embed_dim = text_config.hidden_size

    # Initialize the BLIP text model
    self.text_model = BlipTextModel(text_config)

    # Project text embeddings to the desired dimension
    self.text_projection = nn.Linear(
        self.text_embed_dim,
        self.projection_dim,
        bias=False,
    )

    # Classifier layer for classification task
    self.classifier = nn.Linear(self.projection_dim, num_classes)

    # Initialize the model weights
    self.init_weights()

    # Freeze the base model if specified
    if freeze_base_model:
        for p in self.text_model.parameters():
            p.requires_grad = False

    # Set gradient checkpointing for the encoder
    self.text_model.encoder.gradient_checkpointing = gradient_checkpointing

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Tensor,
    position_ids: Tensor,
)

Forward pass of the BlipForTextClassification model.

Parameters:

Name Type Description Default
input_ids Tensor

Input token IDs.

 required
attention_mask Tensor

Attention mask for the input.

 required
position_ids Tensor

Position IDs for the input.

 required

Returns:

Type Description
Tensor

Output logits for classification.
Source code in src/unitorch/models/blip/modeling.py 
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: torch.Tensor,
    position_ids: torch.Tensor,
):
    """
    Forward pass of the BlipForTextClassification model.

    Args:
        input_ids (torch.Tensor): Input token IDs.
        attention_mask (torch.Tensor): Attention mask for the input.
        position_ids (torch.Tensor): Position IDs for the input.

    Returns:
        (torch.Tensor):Output logits for classification.
    """
    # Pass the input through the BLIP text model
    text_outputs = self.text_model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )

    # Extract text embeddings and project to desired dimension
    text_embeds = text_outputs[1]
    text_embeds = self.text_projection(text_embeds)

    # Apply ReLU activation and pass through the classifier layer
    return self.classifier(nn.functional.relu(text_embeds))

BlipForImageClassification¤

Bases: GenericModel

Initializes the BlipForImageClassification model.

Parameters:

Name Type Description Default
config_path str

Path to the configuration file.

 required
projection_dim Optional[int]

Dimension of the projection. Defaults to 512.

 512
num_classes Optional[int]

Number of classes for classification. Defaults to 1.

 1
freeze_base_model Optional[bool]

Whether to freeze the base model. Defaults to True.

 True
gradient_checkpointing Optional[bool]

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/blip/modeling.py 
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
def __init__(
    self,
    config_path: str,
    projection_dim: Optional[int] = 512,
    num_classes: Optional[int] = 1,
    freeze_base_model: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initializes the BlipForImageClassification model.

    Args:
        config_path (str): Path to the configuration file.
        projection_dim (Optional[int], optional): Dimension of the projection. Defaults to 512.
        num_classes (Optional[int], optional): Number of classes for classification. Defaults to 1.
        freeze_base_model (Optional[bool], optional): Whether to freeze the base model. Defaults to True.
        gradient_checkpointing (Optional[bool], optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()

    # Load the BLIP model configuration
    config = BlipConfig.from_json_file(config_path)
    vision_config = config.vision_config
    vision_config.gradient_checkpointing = gradient_checkpointing

    self.projection_dim = projection_dim
    self.vision_embed_dim = vision_config.hidden_size

    self.vision_model = BlipVisionModel(vision_config)

    self.visual_projection = nn.Linear(
        self.vision_embed_dim,
        self.projection_dim,
        bias=False,
    )

    self.classifier = nn.Linear(self.projection_dim, num_classes)

    self.init_weights()

    if freeze_base_model:
        for p in self.vision_model.parameters():
            p.requires_grad = False

    self.vision_model.encoder.gradient_checkpointing = gradient_checkpointing

forward ¤

forward(pixel_values: Tensor)

Forward pass of the BlipForImageClassification model.

Parameters:

Name Type Description Default
pixel_values Tensor

Input pixel values.

 required

Returns:

Type Description
Tensor

Output logits for classification.
Source code in src/unitorch/models/blip/modeling.py 
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
def forward(
    self,
    pixel_values: torch.Tensor,
):
    """
    Forward pass of the BlipForImageClassification model.

    Args:
        pixel_values (torch.Tensor): Input pixel values.

    Returns:
        (torch.Tensor):Output logits for classification.
    """
    vision_outputs = self.vision_model(
        pixel_values=pixel_values,
    )

    image_embeds = vision_outputs[1]

    image_embeds = self.visual_projection(image_embeds)

    image_embeds = F.relu(image_embeds)

    return self.classifier(image_embeds)

BlipForImageCaption¤

Bases: GenericModel

Initializes the BlipForImageCaption model.

Parameters:

Name Type Description Default
config_path str

Path to the configuration file.

 required
gradient_checkpointing Optional[bool]

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/blip/modeling.py 
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initializes the BlipForImageCaption model.

    Args:
        config_path (str): Path to the configuration file.
        gradient_checkpointing (Optional[bool], optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()
    self.config = BlipConfig.from_json_file(config_path)
    self.config.vision_config.gradient_checkpointing = gradient_checkpointing
    self.config.text_config.gradient_checkpointing = gradient_checkpointing
    self.model = BlipForConditionalGeneration(self.config)
    self.init_weights()

forward ¤

forward(
    pixel_values: Tensor,
    input_ids: Optional[Tensor] = None,
    attention_mask: Optional[Tensor] = None,
)

Forward pass of the BlipForImageCaption model.

Parameters:

Name Type Description Default
pixel_values Tensor

Input pixel values.

 required
input_ids torch.Tensor optional

Input token IDs. Defaults to None.

 None
attention_mask torch.Tensor optional

Attention mask. Defaults to None.

 None

Returns:

Type Description
Tensor

Logits for caption generation.
Source code in src/unitorch/models/blip/modeling.py 
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
def forward(
    self,
    pixel_values: torch.Tensor,
    input_ids: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the BlipForImageCaption model.

    Args:
        pixel_values (torch.Tensor): Input pixel values.
        input_ids (torch.Tensor optional): Input token IDs. Defaults to None.
        attention_mask (torch.Tensor optional): Attention mask. Defaults to None.

    Returns:
        (torch.Tensor):Logits for caption generation.
    """
    outputs = self.model(
        pixel_values=pixel_values,
        input_ids=input_ids,
        attention_mask=attention_mask,
        return_dict=True,
    )
    logits = outputs.decoder_logits
    return logits

generate ¤

generate(
    pixel_values: Tensor,
    input_ids: Optional[Tensor] = None,
    attention_mask: Optional[Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 101,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 102,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
)

Generates captions for the given input images.

Parameters:

Name Type Description Default
pixel_values Tensor

Input pixel values.

 required
input_ids torch.Tensor optional

Input token IDs. Defaults to None.

 None
attention_mask torch.Tensor optional

Attention mask. Defaults to None.

 None
num_beams Optional[int]

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id Optional[int]

ID of the start token for decoding. Defaults to 30522.

 101
decoder_end_token_id int or List[int]

ID of the end token for decoding. Defaults to 2.

 102
num_return_sequences Optional[int]

Number of caption sequences to return. Defaults to 1.

 1
min_gen_seq_length Optional[int]

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length Optional[int]

Maximum length of generated sequences. Defaults to 48.

 48
repetition_penalty Optional[float]

Repetition penalty value. Defaults to 1.0.

 1.0
no_repeat_ngram_size Optional[int]

Size of n-grams to avoid repetition. Defaults to 0.

 0
early_stopping Optional[bool]

Whether to stop generation early. Defaults to True.

 True
length_penalty Optional[float]

Length penalty value. Defaults to 1.0.

 1.0
num_beam_groups Optional[int]

Number of groups for diverse beam search. Defaults to 1.

 1
diversity_penalty Optional[float]

Diversity penalty value. Defaults to 0.0.

 0.0
do_sample Optional[bool]

Whether to use sampling for generation. Defaults to False.

 False
temperature Optional[float]

Temperature value for sampling. Defaults to 1.0.

 1.0
top_k Optional[int]

Value of k for top-k sampling. Defaults to 50.

 50
top_p Optional[float]

Value of p for top-p sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated caption sequences and their scores.
Source code in src/unitorch/models/blip/modeling.py 
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
@torch.no_grad()
def generate(
    self,
    pixel_values: torch.Tensor,
    input_ids: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 101,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 102,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
):
    """
    Generates captions for the given input images.

    Args:
        pixel_values (torch.Tensor): Input pixel values.
        input_ids (torch.Tensor optional): Input token IDs. Defaults to None.
        attention_mask (torch.Tensor optional): Attention mask. Defaults to None.
        num_beams (Optional[int], optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (Optional[int], optional): ID of the start token for decoding. Defaults to 30522.
        decoder_end_token_id (int or List[int], optional): ID of the end token for decoding. Defaults to 2.
        num_return_sequences (Optional[int], optional): Number of caption sequences to return. Defaults to 1.
        min_gen_seq_length (Optional[int], optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (Optional[int], optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (Optional[float], optional): Repetition penalty value. Defaults to 1.0.
        no_repeat_ngram_size (Optional[int], optional): Size of n-grams to avoid repetition. Defaults to 0.
        early_stopping (Optional[bool], optional): Whether to stop generation early. Defaults to True.
        length_penalty (Optional[float], optional): Length penalty value. Defaults to 1.0.
        num_beam_groups (Optional[int], optional): Number of groups for diverse beam search. Defaults to 1.
        diversity_penalty (Optional[float], optional): Diversity penalty value. Defaults to 0.0.
        do_sample (Optional[bool], optional): Whether to use sampling for generation. Defaults to False.
        temperature (Optional[float], optional): Temperature value for sampling. Defaults to 1.0.
        top_k (Optional[int], optional): Value of k for top-k sampling. Defaults to 50.
        top_p (Optional[float], optional): Value of p for top-p sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated caption sequences and their scores.
    """
    outputs = self.model.generate(
        pixel_values=pixel_values,
        input_ids=input_ids,
        attention_mask=attention_mask,
        max_length=max_gen_seq_length,
        min_length=min_gen_seq_length,
        num_beams=num_beams,
        do_sample=do_sample,
        decoder_start_token_id=decoder_start_token_id,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        repetition_penalty=repetition_penalty,
        num_return_sequences=num_return_sequences,
        bos_token_id=decoder_start_token_id,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
        return_dict_in_generate=True,
        output_scores=True,
    )

    sequences = outputs.sequences.reshape(
        -1, num_return_sequences, outputs.sequences.size(-1)
    )
    outputs.sequences = torch.zeros(
        sequences.size(0), num_return_sequences, max_gen_seq_length
    ).to(device=sequences.device)
    outputs.sequences[:, :, : sequences.size(-1)].copy_(sequences)

    if num_return_sequences == 1:
        outputs.sequences = outputs.sequences.reshape(-1, max_gen_seq_length)

    return GenericOutputs(
        sequences=outputs.sequences, sequences_scores=outputs.sequences_scores
    )