visheratin
commited on
Commit
•
2f07d8b
1
Parent(s):
d91b3a3
Update nllb_mrl.py
Browse files- nllb_mrl.py +55 -35
nllb_mrl.py
CHANGED
@@ -1,26 +1,21 @@
|
|
|
|
1 |
from typing import List, Union
|
2 |
|
3 |
import torch
|
4 |
import torch.nn as nn
|
5 |
import torch.nn.functional as F
|
6 |
-
from
|
7 |
-
from open_clip
|
8 |
from PIL import Image
|
9 |
-
from transformers import PretrainedConfig
|
10 |
|
11 |
|
|
|
12 |
class MatryoshkaNllbClipConfig(PretrainedConfig):
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
mrl_resolutions: List[int] = [],
|
18 |
-
**kwargs,
|
19 |
-
):
|
20 |
-
super().__init__(**kwargs)
|
21 |
-
self.clip_model_name = clip_model_name
|
22 |
-
self.target_resolution = target_resolution
|
23 |
-
self.mrl_resolutions = mrl_resolutions
|
24 |
|
25 |
|
26 |
class MatryoshkaLayer(nn.Module):
|
@@ -42,23 +37,16 @@ class MatryoshkaLayer(nn.Module):
|
|
42 |
return outputs
|
43 |
|
44 |
|
45 |
-
class MatryoshkaNllbClip(
|
46 |
-
config_class = MatryoshkaNllbClipConfig
|
47 |
-
|
48 |
def __init__(self, config: MatryoshkaNllbClipConfig, device):
|
49 |
-
super().__init__(
|
50 |
if isinstance(device, str):
|
51 |
device = torch.device(device)
|
52 |
self.config = config
|
53 |
-
self.model =
|
54 |
-
config.clip_model_name, output_dict=True
|
55 |
)
|
56 |
-
|
57 |
-
self.transform = image_transform_v2(
|
58 |
-
pp_cfg,
|
59 |
-
is_train=False,
|
60 |
-
)
|
61 |
-
self._device = device
|
62 |
self.model.to(device)
|
63 |
self.matryoshka_layer = MatryoshkaLayer(
|
64 |
config.mrl_resolutions, config.target_resolution
|
@@ -67,8 +55,8 @@ class MatryoshkaNllbClip(PreTrainedModel):
|
|
67 |
self.tokenizer = get_tokenizer(config.clip_model_name)
|
68 |
|
69 |
def forward(self, image_inputs, input_ids, resolution: Union[int, None] = None):
|
70 |
-
image_inputs = image_inputs.to(self.
|
71 |
-
input_ids = input_ids.to(self.
|
72 |
outputs = self.model(
|
73 |
image=image_inputs,
|
74 |
text=input_ids,
|
@@ -91,14 +79,46 @@ class MatryoshkaNllbClip(PreTrainedModel):
|
|
91 |
"logit_bias": outputs["logit_bias"],
|
92 |
}
|
93 |
|
94 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
self,
|
96 |
images: List[Image.Image],
|
97 |
normalize=False,
|
98 |
resolution: Union[int, None] = None,
|
99 |
):
|
100 |
image_inputs = [self.transform(image) for image in images]
|
101 |
-
image_inputs = torch.stack(image_inputs, dim=0).to(self.
|
102 |
with torch.inference_mode():
|
103 |
features = self.model.visual(image_inputs)
|
104 |
if resolution is not None:
|
@@ -109,7 +129,7 @@ class MatryoshkaNllbClip(PreTrainedModel):
|
|
109 |
features = self.matryoshka_layer.layers[str(resolution)](features)
|
110 |
return F.normalize(features, dim=-1) if normalize else features
|
111 |
|
112 |
-
def
|
113 |
self,
|
114 |
texts: List[str],
|
115 |
langs: Union[List[str], None] = None,
|
@@ -118,10 +138,10 @@ class MatryoshkaNllbClip(PreTrainedModel):
|
|
118 |
):
|
119 |
if langs is None:
|
120 |
langs = ["eng_Latn"] * len(texts)
|
121 |
-
texts = [f"{lang}{text}" for lang, text in zip(langs, texts)]
|
122 |
input_ids = self.tokenizer.tokenizer.batch_encode_plus(
|
123 |
texts, return_tensors="pt", padding="longest", add_special_tokens=False
|
124 |
-
)["input_ids"].to(self.
|
125 |
with torch.inference_mode():
|
126 |
features = self.model.text(input_ids)
|
127 |
if resolution is not None:
|
@@ -139,10 +159,10 @@ class MatryoshkaNllbClip(PreTrainedModel):
|
|
139 |
langs: Union[List[str], None] = None,
|
140 |
resolution: Union[int, None] = None,
|
141 |
):
|
142 |
-
image_features = self.
|
143 |
images, normalize=True, resolution=resolution
|
144 |
)
|
145 |
-
text_features = self.
|
146 |
texts, langs, normalize=True, resolution=resolution
|
147 |
)
|
148 |
with torch.inference_mode():
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
from typing import List, Union
|
3 |
|
4 |
import torch
|
5 |
import torch.nn as nn
|
6 |
import torch.nn.functional as F
|
7 |
+
from huggingface_hub import PyTorchModelHubMixin
|
8 |
+
from open_clip import create_model_and_transforms, get_tokenizer
|
9 |
from PIL import Image
|
10 |
+
from transformers import PretrainedConfig
|
11 |
|
12 |
|
13 |
+
@dataclass
|
14 |
class MatryoshkaNllbClipConfig(PretrainedConfig):
|
15 |
+
clip_model_name: str
|
16 |
+
clip_model_version: str
|
17 |
+
target_resolution: int
|
18 |
+
mrl_resolutions: List[int]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
|
21 |
class MatryoshkaLayer(nn.Module):
|
|
|
37 |
return outputs
|
38 |
|
39 |
|
40 |
+
class MatryoshkaNllbClip(nn.Module, PyTorchModelHubMixin):
|
|
|
|
|
41 |
def __init__(self, config: MatryoshkaNllbClipConfig, device):
|
42 |
+
super().__init__()
|
43 |
if isinstance(device, str):
|
44 |
device = torch.device(device)
|
45 |
self.config = config
|
46 |
+
self.model, _, self.transform = create_model_and_transforms(
|
47 |
+
config.clip_model_name, config.clip_model_version, output_dict=True
|
48 |
)
|
49 |
+
self.device = device
|
|
|
|
|
|
|
|
|
|
|
50 |
self.model.to(device)
|
51 |
self.matryoshka_layer = MatryoshkaLayer(
|
52 |
config.mrl_resolutions, config.target_resolution
|
|
|
55 |
self.tokenizer = get_tokenizer(config.clip_model_name)
|
56 |
|
57 |
def forward(self, image_inputs, input_ids, resolution: Union[int, None] = None):
|
58 |
+
image_inputs = image_inputs.to(self.device)
|
59 |
+
input_ids = input_ids.to(self.device)
|
60 |
outputs = self.model(
|
61 |
image=image_inputs,
|
62 |
text=input_ids,
|
|
|
79 |
"logit_bias": outputs["logit_bias"],
|
80 |
}
|
81 |
|
82 |
+
def encode_image(
|
83 |
+
self,
|
84 |
+
image,
|
85 |
+
normalize=False,
|
86 |
+
resolution: Union[int, None] = None,
|
87 |
+
):
|
88 |
+
with torch.inference_mode():
|
89 |
+
features = self.model.visual(image)
|
90 |
+
if resolution is not None:
|
91 |
+
if resolution not in self.matryoshka_layer.resolutions:
|
92 |
+
raise ValueError(
|
93 |
+
f"Resolution {resolution} not in {self.matryoshka_layer.resolutions}"
|
94 |
+
)
|
95 |
+
features = self.matryoshka_layer.layers[str(resolution)](features)
|
96 |
+
return F.normalize(features, dim=-1) if normalize else features
|
97 |
+
|
98 |
+
def encode_text(
|
99 |
+
self,
|
100 |
+
text,
|
101 |
+
normalize=False,
|
102 |
+
resolution: Union[int, None] = None,
|
103 |
+
):
|
104 |
+
with torch.inference_mode():
|
105 |
+
features = self.model.text(text)
|
106 |
+
if resolution is not None:
|
107 |
+
if resolution not in self.matryoshka_layer.resolutions:
|
108 |
+
raise ValueError(
|
109 |
+
f"Resolution {resolution} not in {self.matryoshka_layer.resolutions}"
|
110 |
+
)
|
111 |
+
features = self.matryoshka_layer.layers[str(resolution)](features)
|
112 |
+
return F.normalize(features, dim=-1) if normalize else features
|
113 |
+
|
114 |
+
def image_features(
|
115 |
self,
|
116 |
images: List[Image.Image],
|
117 |
normalize=False,
|
118 |
resolution: Union[int, None] = None,
|
119 |
):
|
120 |
image_inputs = [self.transform(image) for image in images]
|
121 |
+
image_inputs = torch.stack(image_inputs, dim=0).to(self.device)
|
122 |
with torch.inference_mode():
|
123 |
features = self.model.visual(image_inputs)
|
124 |
if resolution is not None:
|
|
|
129 |
features = self.matryoshka_layer.layers[str(resolution)](features)
|
130 |
return F.normalize(features, dim=-1) if normalize else features
|
131 |
|
132 |
+
def text_features(
|
133 |
self,
|
134 |
texts: List[str],
|
135 |
langs: Union[List[str], None] = None,
|
|
|
138 |
):
|
139 |
if langs is None:
|
140 |
langs = ["eng_Latn"] * len(texts)
|
141 |
+
texts = [f"{lang} {text}" for lang, text in zip(langs, texts)]
|
142 |
input_ids = self.tokenizer.tokenizer.batch_encode_plus(
|
143 |
texts, return_tensors="pt", padding="longest", add_special_tokens=False
|
144 |
+
)["input_ids"].to(self.device)
|
145 |
with torch.inference_mode():
|
146 |
features = self.model.text(input_ids)
|
147 |
if resolution is not None:
|
|
|
159 |
langs: Union[List[str], None] = None,
|
160 |
resolution: Union[int, None] = None,
|
161 |
):
|
162 |
+
image_features = self.image_features(
|
163 |
images, normalize=True, resolution=resolution
|
164 |
)
|
165 |
+
text_features = self.text_features(
|
166 |
texts, langs, normalize=True, resolution=resolution
|
167 |
)
|
168 |
with torch.inference_mode():
|