Skip to content

unitorch.models.kolors¤

KolorsMPSProcessor¤

Bases: HfImageClassificationProcessor, HfTextClassificationProcessor

Initializes the KolorsMPSProcessor.

Parameters:

Name Type Description Default
vocab_path str

Path to the vocabulary file.

 None
merge_path str

Path to the merge file.

 None
vision_config_path str

Path to the vision processor configuration file.

 None
max_seq_length int

Maximum sequence length. Defaults to 77.

 77
position_start_id int

Starting position ID. Defaults to 0.

 0
Source code in src/unitorch/models/kolors/processing.py 
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
def __init__(
    self,
    vocab_path: Optional[str] = None,
    merge_path: Optional[str] = None,
    vision_config_path: Optional[str] = None,
    max_seq_length: Optional[int] = 77,
    position_start_id: Optional[int] = 0,
):
    """
    Initializes the KolorsMPSProcessor.

    Args:
        vocab_path (str, optional): Path to the vocabulary file.
        merge_path (str, optional): Path to the merge file.
        vision_config_path (str, optional): Path to the vision processor configuration file.
        max_seq_length (int, optional): Maximum sequence length. Defaults to 77.
        position_start_id (int, optional): Starting position ID. Defaults to 0.
    """
    if vision_config_path is not None:
        vision_processor = CLIPImageProcessor.from_json_file(vision_config_path)
    else:
        vision_processor = CLIPImageProcessor.from_pretrained(
            "laion/CLIP-ViT-H-14-laion2B-s32B-b79K"
        )
    HfImageClassificationProcessor.__init__(self, vision_processor=vision_processor)

    if vocab_path is not None and merge_path is not None:
        tokenizer = CLIPTokenizer(vocab_file=vocab_path, merges_file=merge_path)
    else:
        tokenizer = CLIPTokenizer.from_pretrained(
            "laion/CLIP-ViT-H-14-laion2B-s32B-b79K"
        )
    tokenizer.cls_token = tokenizer.bos_token
    tokenizer.sep_token = tokenizer.eos_token
    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,
    )

text_classification ¤

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

Processes text for classification.

Parameters:

Name Type Description Default
text str

Input text.

 required
max_seq_length int

Maximum sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs

Processed text inputs.
Source code in src/unitorch/models/kolors/processing.py 
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
def text_classification(
    self,
    text: str,
    max_seq_length: Optional[int] = None,
):
    """
    Processes text for classification.

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

    Returns:
        GenericOutputs: Processed text inputs.
    """
    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,
    )

image_classification ¤

image_classification(image: Union[Image, str])

Processes an image for classification.

Parameters:

Name Type Description Default
image Image or str

Input image or path.

 required

Returns:

Name Type Description
GenericOutputs

Processed image inputs.
Source code in src/unitorch/models/kolors/processing.py 
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
def image_classification(
    self,
    image: Union[Image.Image, str],
):
    """
    Processes an image for classification.

    Args:
        image (PIL.Image.Image or str): Input image or path.

    Returns:
        GenericOutputs: Processed image inputs.
    """
    outputs = HfImageClassificationProcessor.classification(self, image=image)
    return GenericOutputs(pixel_values=outputs.pixel_values)

classification ¤

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

Processes text, image, and condition for multimodal classification.

Parameters:

Name Type Description Default
text str

Input text.

 required
image Image or str

Input image or path.

 required
condition str

Condition text.

 required
max_seq_length int

Maximum sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs

Processed text, image, and condition inputs.
Source code in src/unitorch/models/kolors/processing.py 
 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
def classification(
    self,
    text: str,
    image: Union[Image.Image, str],
    condition: str,
    max_seq_length: Optional[int] = None,
):
    """
    Processes text, image, and condition for multimodal classification.

    Args:
        text (str): Input text.
        image (PIL.Image.Image or str): Input image or path.
        condition (str): Condition text.
        max_seq_length (int, optional): Maximum sequence length. Defaults to None.

    Returns:
        GenericOutputs: Processed text, image, and condition inputs.
    """
    text_outputs = self.text_classification(
        text=text, max_seq_length=max_seq_length
    )
    pixel_outputs = self.image_classification(image=image)
    condition_outputs = self.text_classification(
        text=condition, max_seq_length=max_seq_length
    )
    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,
        condition_input_ids=condition_outputs.input_ids,
        condition_attention_mask=condition_outputs.attention_mask,
        condition_position_ids=condition_outputs.position_ids,
    )

KolorsMPSModel¤

Bases: GenericModel

Initializes the KolorsMPSModel.

Parameters:

Name Type Description Default
config_path str

Path to the CLIP configuration file.

 required
Source code in src/unitorch/models/kolors/modeling.py 
217
218
219
220
221
222
223
224
225
226
227
def __init__(self, config_path: str):
    """
    Initializes the KolorsMPSModel.

    Args:
        config_path (str): Path to the CLIP configuration file.
    """
    super().__init__()
    self.config = CLIPConfig.from_json_file(config_path)
    self.model = CLIPModel(self.config)
    self.cross_model = CrossModel(dim=1024, layer_num=4, heads=16)

config instance-attribute ¤

config = from_json_file(config_path)

model instance-attribute ¤

model = CLIPModel(config)

cross_model instance-attribute ¤

cross_model = CrossModel(dim=1024, layer_num=4, heads=16)

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Tensor,
    position_ids: Tensor,
    pixel_values: Tensor,
    condition_input_ids: Tensor,
    condition_attention_mask: Tensor,
    condition_position_ids: Tensor,
    labels: Optional[Tensor] = None,
)

Forward pass of the KolorsMPSModel.

Parameters:

Name Type Description Default
input_ids Tensor

Text token IDs.

 required
attention_mask Tensor

Text attention mask.

 required
position_ids Tensor

Text position IDs.

 required
pixel_values Tensor

Image pixel values.

 required
condition_input_ids Tensor

Condition text token IDs.

 required
condition_attention_mask Tensor

Condition text attention mask.

 required
condition_position_ids Tensor

Condition text position IDs.

 required
labels Tensor

Labels (unused). Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Scaled similarity scores.
Source code in src/unitorch/models/kolors/modeling.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: torch.Tensor,
    position_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    condition_input_ids: torch.Tensor,
    condition_attention_mask: torch.Tensor,
    condition_position_ids: torch.Tensor,
    labels: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the KolorsMPSModel.

    Args:
        input_ids (torch.Tensor): Text token IDs.
        attention_mask (torch.Tensor): Text attention mask.
        position_ids (torch.Tensor): Text position IDs.
        pixel_values (torch.Tensor): Image pixel values.
        condition_input_ids (torch.Tensor): Condition text token IDs.
        condition_attention_mask (torch.Tensor): Condition text attention mask.
        condition_position_ids (torch.Tensor): Condition text position IDs.
        labels (torch.Tensor, optional): Labels (unused). Defaults to None.

    Returns:
        torch.Tensor: Scaled similarity scores.
    """
    text_outputs = self.model.text_model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )
    text_features = self.model.text_projection(text_outputs[0])
    text_pooled_features = self.model.text_projection(text_outputs[1])

    image_outputs = self.model.vision_model(pixel_values=pixel_values)
    image_features = self.model.visual_projection(image_outputs[0])

    condition_outputs = self.model.text_model(
        input_ids=condition_input_ids,
        attention_mask=condition_attention_mask,
        position_ids=condition_position_ids,
    )
    condition_features = self.model.text_projection(condition_outputs[0])

    sim_text_condition = torch.einsum(
        "b i d, b j d -> b j i", text_features, condition_features
    )
    sim_text_condition = torch.max(sim_text_condition, dim=1, keepdim=True)[0]
    sim_text_condition = sim_text_condition / sim_text_condition.max()
    mask = torch.where(sim_text_condition > 0.01, 0, float("-inf"))
    mask = mask.repeat(1, image_features.shape[1], 1)

    cross_features = self.cross_model(image_features, text_features, mask)[:, 0]
    text_embeds = text_pooled_features / text_pooled_features.norm(
        dim=-1, keepdim=True
    )
    cross_embeds = cross_features / cross_features.norm(dim=-1, keepdim=True)

    scores = torch.sum(text_embeds * cross_embeds, dim=-1, keepdim=True)
    return self.model.logit_scale.exp() * scores