enabling transformers
Browse files- config.json +8 -0
- marqo_fashionSigLIP.py +237 -0
- model.safetensors +3 -0
- preprocessor_config.json +27 -0
- spiece.model +3 -0
config.json
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"MarqoFashionSigLIP"
|
4 |
+
],
|
5 |
+
"open_clip_model_name": "hf-hub:Marqo/marqo-ecommerce-embeddings-L",
|
6 |
+
"torch_dtype": "float32",
|
7 |
+
"transformers_version": "4.45.1"
|
8 |
+
}
|
marqo_fashionSigLIP.py
ADDED
@@ -0,0 +1,237 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from open_clip import create_model
|
3 |
+
from transformers import PretrainedConfig, PreTrainedModel
|
4 |
+
from transformers.models.siglip.modeling_siglip import SiglipOutput
|
5 |
+
from typing import Optional, Tuple, Union, List
|
6 |
+
from transformers.feature_extraction_utils import BatchFeature
|
7 |
+
from transformers.image_utils import ImageInput
|
8 |
+
from transformers.processing_utils import ProcessorMixin
|
9 |
+
from transformers.tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
|
10 |
+
from transformers.utils import TensorType
|
11 |
+
import string
|
12 |
+
import ftfy
|
13 |
+
import html
|
14 |
+
|
15 |
+
def basic_clean(text):
|
16 |
+
text = ftfy.fix_text(text)
|
17 |
+
text = html.unescape(html.unescape(text))
|
18 |
+
return text.strip()
|
19 |
+
|
20 |
+
def canonicalize_text(
|
21 |
+
text,
|
22 |
+
*,
|
23 |
+
keep_punctuation_exact_string=None,
|
24 |
+
trans_punctuation: dict = str.maketrans("", "", string.punctuation),
|
25 |
+
):
|
26 |
+
"""Returns canonicalized `text` (lowercase and punctuation removed).
|
27 |
+
|
28 |
+
From: https://github.com/google-research/big_vision/blob/53f18caf27a9419231bbf08d3388b07671616d3d/big_vision/evaluators/proj/image_text/prompt_engineering.py#L94
|
29 |
+
|
30 |
+
Args:
|
31 |
+
text: string to be canonicalized.
|
32 |
+
keep_punctuation_exact_string: If provided, then this exact string kept.
|
33 |
+
For example providing '{}' will keep any occurrences of '{}' (but will
|
34 |
+
still remove '{' and '}' that appear separately).
|
35 |
+
"""
|
36 |
+
text = text.replace("_", " ")
|
37 |
+
if keep_punctuation_exact_string:
|
38 |
+
text = keep_punctuation_exact_string.join(
|
39 |
+
part.translate(trans_punctuation)
|
40 |
+
for part in text.split(keep_punctuation_exact_string)
|
41 |
+
)
|
42 |
+
else:
|
43 |
+
text = text.translate(trans_punctuation)
|
44 |
+
text = text.lower()
|
45 |
+
text = " ".join(text.split())
|
46 |
+
return text.strip()
|
47 |
+
|
48 |
+
def _clean_canonicalize(x):
|
49 |
+
# basic, remove whitespace, remove punctuation, lower case
|
50 |
+
return canonicalize_text(basic_clean(x))
|
51 |
+
|
52 |
+
class MarqoFashionSigLIPConfig(PretrainedConfig):
|
53 |
+
def __init__(
|
54 |
+
self,
|
55 |
+
open_clip_model_name: str = "",
|
56 |
+
**kwargs,
|
57 |
+
):
|
58 |
+
super().__init__(**kwargs)
|
59 |
+
self.open_clip_model_name = open_clip_model_name
|
60 |
+
|
61 |
+
class MarqoFashionSigLIPProcessor(ProcessorMixin):
|
62 |
+
r"""
|
63 |
+
Constructs a Siglip processor which wraps a Siglip image processor and a Siglip tokenizer into a single processor.
|
64 |
+
|
65 |
+
[`SiglipProcessor`] offers all the functionalities of [`SiglipImageProcessor`] and [`SiglipTokenizer`]. See the
|
66 |
+
[`~SiglipProcessor.__call__`] and [`~SiglipProcessor.decode`] for more information.
|
67 |
+
|
68 |
+
Args:
|
69 |
+
image_processor ([`SiglipImageProcessor`]):
|
70 |
+
The image processor is a required input.
|
71 |
+
tokenizer ([`T5TokenizerFast`]):
|
72 |
+
The tokenizer is a required input.
|
73 |
+
"""
|
74 |
+
|
75 |
+
attributes = ["image_processor", "tokenizer"]
|
76 |
+
image_processor_class = "SiglipImageProcessor"
|
77 |
+
tokenizer_class = "T5TokenizerFast"
|
78 |
+
|
79 |
+
def __init__(self, image_processor, tokenizer):
|
80 |
+
super().__init__(image_processor, tokenizer)
|
81 |
+
|
82 |
+
def __call__(
|
83 |
+
self,
|
84 |
+
text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
|
85 |
+
images: ImageInput = None,
|
86 |
+
padding: Union[bool, str, PaddingStrategy] = False,
|
87 |
+
truncation: Union[bool, str, TruncationStrategy] = None,
|
88 |
+
max_length: int = None,
|
89 |
+
return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
|
90 |
+
) -> BatchFeature:
|
91 |
+
"""
|
92 |
+
Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
|
93 |
+
and `kwargs` arguments to SiglipTokenizer's [`~SiglipTokenizer.__call__`] if `text` is not `None` to encode
|
94 |
+
the text. To prepare the image(s), this method forwards the `images` argument to
|
95 |
+
SiglipImageProcessor's [`~SiglipImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
|
96 |
+
of the above two methods for more information.
|
97 |
+
|
98 |
+
Args:
|
99 |
+
text (`str`, `List[str]`, `List[List[str]]`):
|
100 |
+
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
|
101 |
+
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
|
102 |
+
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
|
103 |
+
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
|
104 |
+
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
|
105 |
+
tensor. Both channels-first and channels-last formats are supported.
|
106 |
+
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
|
107 |
+
Select a strategy to pad the returned sequences (according to the model's padding side and padding
|
108 |
+
index) among:
|
109 |
+
- `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
|
110 |
+
sequence if provided).
|
111 |
+
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
|
112 |
+
acceptable input length for the model if that argument is not provided.
|
113 |
+
- `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
|
114 |
+
lengths).
|
115 |
+
max_length (`int`, *optional*):
|
116 |
+
Maximum length of the returned list and optionally padding length (see above).
|
117 |
+
truncation (`bool`, *optional*):
|
118 |
+
Activates truncation to cut input sequences longer than `max_length` to `max_length`.
|
119 |
+
return_tensors (`str` or [`~utils.TensorType`], *optional*):
|
120 |
+
If set, will return tensors of a particular framework. Acceptable values are:
|
121 |
+
|
122 |
+
- `'tf'`: Return TensorFlow `tf.constant` objects.
|
123 |
+
- `'pt'`: Return PyTorch `torch.Tensor` objects.
|
124 |
+
- `'np'`: Return NumPy `np.ndarray` objects.
|
125 |
+
- `'jax'`: Return JAX `jnp.ndarray` objects.
|
126 |
+
|
127 |
+
Returns:
|
128 |
+
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
|
129 |
+
|
130 |
+
- **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
|
131 |
+
- **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
|
132 |
+
`return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
|
133 |
+
`None`).
|
134 |
+
- **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
|
135 |
+
"""
|
136 |
+
|
137 |
+
if text is None and images is None:
|
138 |
+
raise ValueError("You have to specify either text or images. Both cannot be none.")
|
139 |
+
|
140 |
+
if text is not None:
|
141 |
+
if isinstance(text, str):
|
142 |
+
text = [text]
|
143 |
+
text = [_clean_canonicalize(raw_text) for raw_text in text]
|
144 |
+
encoding = self.tokenizer(
|
145 |
+
text, return_tensors=return_tensors, padding=padding, truncation=truncation, max_length=max_length
|
146 |
+
)
|
147 |
+
|
148 |
+
if images is not None:
|
149 |
+
try:
|
150 |
+
images = [image.convert('RGB') for image in images] if isinstance(images, list) else images.convert('RGB')
|
151 |
+
except:
|
152 |
+
images = images
|
153 |
+
image_features = self.image_processor(images, return_tensors=return_tensors)
|
154 |
+
|
155 |
+
if text is not None and images is not None:
|
156 |
+
encoding["pixel_values"] = image_features.pixel_values
|
157 |
+
return encoding
|
158 |
+
elif text is not None:
|
159 |
+
return encoding
|
160 |
+
else:
|
161 |
+
return BatchFeature(data=dict(**image_features), tensor_type=return_tensors)
|
162 |
+
|
163 |
+
def decode(self, *args, **kwargs):
|
164 |
+
"""
|
165 |
+
This method forwards all its arguments to SiglipTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to
|
166 |
+
the docstring of this method for more information.
|
167 |
+
"""
|
168 |
+
return self.tokenizer.decode(*args, **kwargs)
|
169 |
+
|
170 |
+
def batch_decode(self, *args, **kwargs):
|
171 |
+
"""
|
172 |
+
This method forwards all its arguments to SiglipTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
|
173 |
+
refer to the docstring of this method for more information.
|
174 |
+
"""
|
175 |
+
return self.tokenizer.batch_decode(*args, **kwargs)
|
176 |
+
|
177 |
+
@property
|
178 |
+
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names with CLIP->Siglip, T5->Siglip
|
179 |
+
def model_input_names(self):
|
180 |
+
tokenizer_input_names = self.tokenizer.model_input_names
|
181 |
+
image_processor_input_names = self.image_processor.model_input_names
|
182 |
+
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
|
183 |
+
|
184 |
+
class MarqoFashionSigLIP(PreTrainedModel):
|
185 |
+
config_class = MarqoFashionSigLIPConfig
|
186 |
+
|
187 |
+
def __init__(self, config: MarqoFashionSigLIPConfig):
|
188 |
+
super().__init__(config)
|
189 |
+
self.config = config
|
190 |
+
self.model = create_model(config.open_clip_model_name, output_dict=True)
|
191 |
+
self.model.eval()
|
192 |
+
self.model.to(self.device)
|
193 |
+
|
194 |
+
def get_image_features(
|
195 |
+
self,
|
196 |
+
pixel_values: torch.FloatTensor,
|
197 |
+
normalize: bool = False,
|
198 |
+
**kwargs
|
199 |
+
) -> torch.FloatTensor:
|
200 |
+
|
201 |
+
with torch.inference_mode():
|
202 |
+
image_features = self.model.encode_image(pixel_values, normalize=normalize)
|
203 |
+
return image_features
|
204 |
+
|
205 |
+
def get_text_features(
|
206 |
+
self,
|
207 |
+
input_ids: torch.Tensor,
|
208 |
+
normalize: bool = False,
|
209 |
+
**kwargs
|
210 |
+
) -> torch.FloatTensor:
|
211 |
+
|
212 |
+
with torch.inference_mode():
|
213 |
+
text_features = self.model.encode_text(input_ids, normalize=normalize)
|
214 |
+
return text_features
|
215 |
+
|
216 |
+
def forward(
|
217 |
+
self,
|
218 |
+
input_ids: Optional[torch.LongTensor] = None,
|
219 |
+
pixel_values: Optional[torch.FloatTensor] = None,
|
220 |
+
return_dict: Optional[bool] = None,
|
221 |
+
) -> Union[Tuple, SiglipOutput]:
|
222 |
+
|
223 |
+
vision_outputs = self.get_image_features(pixel_values=pixel_values, normalize=True)
|
224 |
+
text_outputs = self.get_text_features(input_ids=input_ids, normalize=True)
|
225 |
+
|
226 |
+
logits_per_text = text_outputs @ vision_outputs.T
|
227 |
+
logits_per_image = logits_per_text.T
|
228 |
+
|
229 |
+
if not return_dict:
|
230 |
+
return logits_per_image, logits_per_text, text_outputs, vision_outputs
|
231 |
+
|
232 |
+
return SiglipOutput(
|
233 |
+
logits_per_image=logits_per_image,
|
234 |
+
logits_per_text=logits_per_text,
|
235 |
+
text_embeds=text_outputs,
|
236 |
+
image_embeds=vision_outputs
|
237 |
+
)
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5f54e3323fc98caddba9626aa9771efd873c3cb9d63cc65b4619c2ccb6213e4e
|
3 |
+
size 2608674872
|
preprocessor_config.json
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"auto_map": {
|
3 |
+
"AutoProcessor": "marqo_fashionSigLIP.MarqoFashionSigLIPProcessor"
|
4 |
+
},
|
5 |
+
"do_normalize": true,
|
6 |
+
"do_rescale": true,
|
7 |
+
"do_resize": true,
|
8 |
+
"do_convert_rgb": true,
|
9 |
+
"image_processor_type": "SiglipImageProcessor",
|
10 |
+
"image_mean": [
|
11 |
+
0.5,
|
12 |
+
0.5,
|
13 |
+
0.5
|
14 |
+
],
|
15 |
+
"processor_class": "marqo_fashionSigLIP.MarqoFashionSigLIPProcessor",
|
16 |
+
"resample": 3,
|
17 |
+
"rescale_factor": 0.00392156862745098,
|
18 |
+
"size": {
|
19 |
+
"height": 224,
|
20 |
+
"width": 224
|
21 |
+
},
|
22 |
+
"image_std": [
|
23 |
+
0.5,
|
24 |
+
0.5,
|
25 |
+
0.5
|
26 |
+
]
|
27 |
+
}
|
spiece.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d60acb128cf7b7f2536e8f38a5b18a05535c9e14c7a355904270e15b0945ea86
|
3 |
+
size 791656
|