Commit
·
e08f928
1
Parent(s):
9530eee
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,270 @@
|
|
1 |
---
|
2 |
license: creativeml-openrail-m
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: creativeml-openrail-m
|
3 |
---
|
4 |
+
|
5 |
+
|
6 |
+
# AltCLIP-m18
|
7 |
+
|
8 |
+
|
9 |
+
| 名称 Name | 任务 Task | 语言 Language(s) | 模型 Model | Github |
|
10 |
+
|:------------------:|:----------:|:-------------------:|:--------:|:------:|
|
11 |
+
| AltCLIP-m18 | Text-Image | Multilingual | CLIP | [FlagAI](https://github.com/FlagAI-Open/FlagAI) |
|
12 |
+
|
13 |
+
## 简介 Brief Introduction
|
14 |
+
|
15 |
+
继双语模型AltCLIP与9语模型AltCLIP-m9之后,我们训练了18语CLIP模型。命名为AltCLIP-m18。它支持英语、中文、日语、泰语、韩语、印地语、乌克兰语、阿拉伯语、土耳其语、越南语、波兰语、荷兰语、葡萄牙语、意大利语、西班牙语、德语、法语和俄语。
|
16 |
+
|
17 |
+
AltCLIP-m18模型可以为AltDiffusion-m18模型提供支持,关于AltDiffusion模型的具体信息可查看[此教程](https://github.com/FlagAI-Open/FlagAI/tree/master/examples/AltDiffusion/README.md) 。
|
18 |
+
|
19 |
+
模型代码已经在 [FlagAI](https://github.com/FlagAI-Open/FlagAI/tree/master/examples/AltCLIP-m18) 上开源,权重位于我们搭建的 [modelhub](https://model.baai.ac.cn/model-detail/100095) 上。我们还提供了微调,推理,验证的脚本,欢迎试用。
|
20 |
+
|
21 |
+
Following the bilingual model AltCLIP and the nine-language model AltCLIP-m9, we trained the eighteen-language CLIP model, Named AltCLIP-m18. It supports English, Chinese, Japanese, Thai, Korean, Hindi, Ukrainian, Arabic, Turkish, Vietnamese, Polish, Dutch, Portuguese, Italian, Spanish, German, French, and Russian.
|
22 |
+
|
23 |
+
The AltCLIP-m18 model can provide support for the AltDiffusion-m18 model. Specific information on the AltDiffusion modle can be found in [this tutorial](https://github.com/FlagAI-Open/FlagAI/tree/master/examples/AltDiffusion/README.md).
|
24 |
+
|
25 |
+
The model code has been open sourced on [FlagAI](https://github.com/FlagAI-Open/FlagAI/tree/master/examples/AltCLIP-m18) and the weights are located on [modelhub](https://model.baai.ac.cn/model-detail/100095). We also provide scripts for fine-tuning, inference, and validation, so feel free to try them out.
|
26 |
+
|
27 |
+
## 引用
|
28 |
+
关于AltCLIP,我们已经推出了相关报告,有更多细节可以查阅,如对您的工作有帮助,欢迎引用。
|
29 |
+
|
30 |
+
If you find this work helpful, please consider to cite
|
31 |
+
```
|
32 |
+
@article{https://doi.org/10.48550/arxiv.2211.06679,
|
33 |
+
doi = {10.48550/ARXIV.2211.06679},
|
34 |
+
url = {https://arxiv.org/abs/2211.06679},
|
35 |
+
author = {Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell},
|
36 |
+
keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences},
|
37 |
+
title = {AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities},
|
38 |
+
publisher = {arXiv},
|
39 |
+
year = {2022},
|
40 |
+
copyright = {arXiv.org perpetual, non-exclusive license}
|
41 |
+
}
|
42 |
+
```
|
43 |
+
|
44 |
+
## AltCLIP-m18评测 AltCLIP-m18 evaluation
|
45 |
+
|
46 |
+
部分数据集评测结果展示:
|
47 |
+
|
48 |
+
Partial dataset evaluation results are displayed:
|
49 |
+
|
50 |
+
| | birdsnap | caltech101 | cars | cifar10 | cifar100 | country211 | dtd | eurosat | fer2013 |
|
51 |
+
| :----: | :----: | :----: | :----: | :----: | :----: | :----: | :----: | :----: | :----: |
|
52 |
+
| AltCLIP-M18 | 41.57 | 88.25 | 92.75 | 97.44 | 84.83 | 30.52 | 68.62 | 67.46 | 54.4 |
|
53 |
+
|
54 |
+
Cifar10 dataset evaluation
|
55 |
+
|
56 |
+
```python
|
57 |
+
# Copyright © 2022 BAAI. All rights reserved.
|
58 |
+
#
|
59 |
+
# Licensed under the Apache License, Version 2.0 (the "License")
|
60 |
+
import torch
|
61 |
+
from flagai.auto_model.auto_loader import AutoLoader
|
62 |
+
import zeroshot_classification
|
63 |
+
import json
|
64 |
+
import os
|
65 |
+
from torchvision.datasets import CIFAR10
|
66 |
+
|
67 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
68 |
+
maxlen = 256
|
69 |
+
|
70 |
+
dataset_root = "./clip_benchmark_datasets/"
|
71 |
+
dataset_name = "cifar10"
|
72 |
+
|
73 |
+
auto_loader = AutoLoader(
|
74 |
+
task_name="txt_img_matching",
|
75 |
+
model_dir="./checkpoints/",
|
76 |
+
model_name="AltCLIP-XLMR-L-m18" # Load the checkpoints from Modelhub(model.baai.ac.cn/models)
|
77 |
+
)
|
78 |
+
|
79 |
+
model = auto_loader.get_model()
|
80 |
+
model.to(device)
|
81 |
+
model.eval()
|
82 |
+
tokenizer = auto_loader.get_tokenizer()
|
83 |
+
transform = auto_loader.get_transform()
|
84 |
+
|
85 |
+
dataset = CIFAR10(root=os.path.join(dataset_root, dataset_name),
|
86 |
+
transform=transform,
|
87 |
+
download=True)
|
88 |
+
batch_size = 128
|
89 |
+
num_workers = 4
|
90 |
+
|
91 |
+
template = {"cifar10": [
|
92 |
+
"a photo of a {c}.",
|
93 |
+
"a blurry photo of a {c}.",
|
94 |
+
"a black and white photo of a {c}.",
|
95 |
+
"a low contrast photo of a {c}.",
|
96 |
+
"a high contrast photo of a {c}.",
|
97 |
+
"a bad photo of a {c}.",
|
98 |
+
"a good photo of a {c}.",
|
99 |
+
"a photo of a small {c}.",
|
100 |
+
"a photo of a big {c}.",
|
101 |
+
"a photo of the {c}.",
|
102 |
+
"a blurry photo of the {c}.",
|
103 |
+
"a black and white photo of the {c}.",
|
104 |
+
"a low contrast photo of the {c}.",
|
105 |
+
"a high contrast photo of the {c}.",
|
106 |
+
"a bad photo of the {c}.",
|
107 |
+
"a good photo of the {c}.",
|
108 |
+
"a photo of the small {c}.",
|
109 |
+
"a photo of the big {c}."
|
110 |
+
],
|
111 |
+
}
|
112 |
+
def evaluate():
|
113 |
+
if dataset:
|
114 |
+
dataloader = torch.utils.data.DataLoader(
|
115 |
+
dataset,
|
116 |
+
batch_size=batch_size,
|
117 |
+
shuffle=False,
|
118 |
+
num_workers=num_workers,
|
119 |
+
)
|
120 |
+
|
121 |
+
zeroshot_templates = template["cifar10"]
|
122 |
+
classnames = dataset.classes if hasattr(dataset, "classes") else None
|
123 |
+
|
124 |
+
metrics = zeroshot_classification.evaluate(
|
125 |
+
model,
|
126 |
+
dataloader,
|
127 |
+
tokenizer,
|
128 |
+
classnames,
|
129 |
+
zeroshot_templates,
|
130 |
+
device=device,
|
131 |
+
amp=True,
|
132 |
+
)
|
133 |
+
|
134 |
+
dump = {
|
135 |
+
"dataset": dataset_name,
|
136 |
+
"metrics": metrics
|
137 |
+
}
|
138 |
+
|
139 |
+
print(dump)
|
140 |
+
with open("./result.txt", "w") as f:
|
141 |
+
json.dump(dump, f)
|
142 |
+
return metrics
|
143 |
+
|
144 |
+
if __name__ == "__main__":
|
145 |
+
evaluate()
|
146 |
+
|
147 |
+
```
|
148 |
+
|
149 |
+
|
150 |
+
|
151 |
+
## 推理脚本 inference
|
152 |
+
|
153 |
+
```python
|
154 |
+
import torch
|
155 |
+
from PIL import Image
|
156 |
+
from flagai.auto_model.auto_loader import AutoLoader
|
157 |
+
|
158 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
159 |
+
|
160 |
+
loader = AutoLoader(
|
161 |
+
task_name="txt_img_matching",
|
162 |
+
model_name="AltCLIP-XLMR-L-m18", # Load the checkpoints from Modelhub(model.baai.ac.cn/models)
|
163 |
+
model_dir="./checkpoints"
|
164 |
+
)
|
165 |
+
|
166 |
+
model = loader.get_model()
|
167 |
+
tokenizer = loader.get_tokenizer()
|
168 |
+
transform = loader.get_transform()
|
169 |
+
|
170 |
+
model.eval()
|
171 |
+
model.to(device)
|
172 |
+
tokenizer = loader.get_tokenizer()
|
173 |
+
|
174 |
+
def inference():
|
175 |
+
image = Image.open("./dog.jpeg")
|
176 |
+
image = transform(image)
|
177 |
+
image = torch.tensor(image["pixel_values"]).to(device)
|
178 |
+
tokenizer_out = tokenizer(["a rat", "a dog", "a cat"],
|
179 |
+
padding=True,
|
180 |
+
truncation=True,
|
181 |
+
max_length=77,
|
182 |
+
return_tensors='pt')
|
183 |
+
|
184 |
+
text = tokenizer_out["input_ids"].to(device)
|
185 |
+
attention_mask = tokenizer_out["attention_mask"].to(device)
|
186 |
+
with torch.no_grad():
|
187 |
+
image_features = model.get_image_features(image)
|
188 |
+
text_features = model.get_text_features(text, attention_mask=attention_mask)
|
189 |
+
text_probs = (image_features @ text_features.T).softmax(dim=-1)
|
190 |
+
|
191 |
+
print(text_probs.cpu().numpy()[0].tolist())
|
192 |
+
|
193 |
+
if __name__=="__main__":
|
194 |
+
inference()
|
195 |
+
```
|
196 |
+
|
197 |
+
|
198 |
+
|
199 |
+
## 微调 fintuning
|
200 |
+
|
201 |
+
Cifar10 dataset
|
202 |
+
|
203 |
+
```python
|
204 |
+
# Copyright © 2022 BAAI. All rights reserved.
|
205 |
+
#
|
206 |
+
# Licensed under the Apache License, Version 2.0 (the "License")
|
207 |
+
import torch
|
208 |
+
from flagai.auto_model.auto_loader import AutoLoader
|
209 |
+
import os
|
210 |
+
from flagai.trainer import Trainer
|
211 |
+
from torchvision.datasets import (
|
212 |
+
CIFAR10
|
213 |
+
)
|
214 |
+
|
215 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
216 |
+
|
217 |
+
dataset_root = "./clip_benchmark_datasets"
|
218 |
+
dataset_name = "cifar10"
|
219 |
+
|
220 |
+
batch_size = 4
|
221 |
+
classes = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
|
222 |
+
|
223 |
+
auto_loader = AutoLoader(
|
224 |
+
task_name="txt_img_matching",
|
225 |
+
model_dir="./checkpoints",
|
226 |
+
model_name="AltCLIP-XLMR-L-m18" # Load the checkpoints from Modelhub(model.baai.ac.cn/models)
|
227 |
+
)
|
228 |
+
|
229 |
+
model = auto_loader.get_model()
|
230 |
+
model.to(device)
|
231 |
+
model.eval()
|
232 |
+
tokenizer = auto_loader.get_tokenizer()
|
233 |
+
transform = auto_loader.get_transform()
|
234 |
+
|
235 |
+
trainer = Trainer(env_type="pytorch",
|
236 |
+
pytorch_device=device,
|
237 |
+
experiment_name="clip_finetuning",
|
238 |
+
batch_size=4,
|
239 |
+
lr=1e-4,
|
240 |
+
epochs=10,
|
241 |
+
log_interval=10)
|
242 |
+
|
243 |
+
dataset = CIFAR10(root=os.path.join(dataset_root, dataset_name),
|
244 |
+
transform=transform,
|
245 |
+
download=True)
|
246 |
+
|
247 |
+
def cifar10_collate_fn(batch):
|
248 |
+
# image shape is (batch, 3, 224, 224)
|
249 |
+
images = torch.tensor([b[0]["pixel_values"][0] for b in batch])
|
250 |
+
# text_id shape is (batch, n)
|
251 |
+
input_ids = torch.tensor([tokenizer(f"a photo of a {b[1]}",
|
252 |
+
padding=True,
|
253 |
+
truncation=True,
|
254 |
+
max_length=77)["input_ids"] for b in batch])
|
255 |
+
|
256 |
+
attention_mask = torch.tensor([tokenizer(f"a photo of a {b[1]}",
|
257 |
+
padding=True,
|
258 |
+
truncation=True,
|
259 |
+
max_length=77)["attention_mask"] for b in batch])
|
260 |
+
|
261 |
+
return {
|
262 |
+
"pixel_values": images,
|
263 |
+
"input_ids": input_ids,
|
264 |
+
"attention_mask": attention_mask,
|
265 |
+
}
|
266 |
+
|
267 |
+
if __name__ == "__main__":
|
268 |
+
trainer.train(model=model, train_dataset=dataset, collate_fn=cifar10_collate_fn)
|
269 |
+
```
|
270 |
+
|