Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- zh
|
4 |
+
pipeline_tag: image-to-text
|
5 |
+
tags:
|
6 |
+
- vit
|
7 |
+
- gpt
|
8 |
+
---
|
9 |
+
```python
|
10 |
+
from transformers import VisionEncoderDecoderModel, ViTFeatureExtractor, AutoTokenizer
|
11 |
+
import torch
|
12 |
+
from PIL import Image
|
13 |
+
import pathlib
|
14 |
+
import pandas as pd
|
15 |
+
import numpy as np
|
16 |
+
from IPython.core.display import HTML
|
17 |
+
import os
|
18 |
+
import requests
|
19 |
+
|
20 |
+
class Image2Caption(object):
|
21 |
+
def __init__(self ,model_path = "nlpconnect/vit-gpt2-image-captioning",
|
22 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu"),
|
23 |
+
overwrite_encoder_checkpoint_path = None,
|
24 |
+
overwrite_token_model_path = None
|
25 |
+
):
|
26 |
+
assert type(overwrite_token_model_path) == type("") or overwrite_token_model_path is None
|
27 |
+
assert type(overwrite_encoder_checkpoint_path) == type("") or overwrite_encoder_checkpoint_path is None
|
28 |
+
if overwrite_token_model_path is None:
|
29 |
+
overwrite_token_model_path = model_path
|
30 |
+
if overwrite_encoder_checkpoint_path is None:
|
31 |
+
overwrite_encoder_checkpoint_path = model_path
|
32 |
+
self.device = device
|
33 |
+
self.model = VisionEncoderDecoderModel.from_pretrained(model_path)
|
34 |
+
self.feature_extractor = ViTFeatureExtractor.from_pretrained(overwrite_encoder_checkpoint_path)
|
35 |
+
self.tokenizer = AutoTokenizer.from_pretrained(overwrite_token_model_path)
|
36 |
+
self.model = self.model.to(self.device)
|
37 |
+
|
38 |
+
def predict_to_df(self, image_paths):
|
39 |
+
img_caption_pred = self.predict_step(image_paths)
|
40 |
+
img_cation_df = pd.DataFrame(list(zip(image_paths, img_caption_pred)))
|
41 |
+
img_cation_df.columns = ["img", "caption"]
|
42 |
+
return img_cation_df
|
43 |
+
#img_cation_df.to_html(escape=False, formatters=dict(Country=path_to_image_html))
|
44 |
+
|
45 |
+
def predict_step(self ,image_paths, max_length = 128, num_beams = 4):
|
46 |
+
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
|
47 |
+
images = []
|
48 |
+
for image_path in image_paths:
|
49 |
+
#i_image = Image.open(image_path)
|
50 |
+
if image_path.startswith("http"):
|
51 |
+
i_image = Image.open(
|
52 |
+
requests.get(image_path, stream=True).raw
|
53 |
+
)
|
54 |
+
else:
|
55 |
+
i_image = Image.open(image_path)
|
56 |
+
|
57 |
+
if i_image.mode != "RGB":
|
58 |
+
i_image = i_image.convert(mode="RGB")
|
59 |
+
images.append(i_image)
|
60 |
+
|
61 |
+
pixel_values = self.feature_extractor(images=images, return_tensors="pt").pixel_values
|
62 |
+
pixel_values = pixel_values.to(self.device)
|
63 |
+
|
64 |
+
output_ids = self.model.generate(pixel_values, **gen_kwargs)
|
65 |
+
|
66 |
+
preds = self.tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
67 |
+
preds = [pred.strip() for pred in preds]
|
68 |
+
return preds
|
69 |
+
|
70 |
+
def path_to_image_html(path):
|
71 |
+
return '<img src="'+ path + '" width="60" >'
|
72 |
+
|
73 |
+
i2c_tiny_zh_obj = Image2Caption("svjack/vit-gpt-diffusion-zh",
|
74 |
+
overwrite_encoder_checkpoint_path = "google/vit-base-patch16-224",
|
75 |
+
overwrite_token_model_path = "IDEA-CCNL/Wenzhong-GPT2-110M"
|
76 |
+
)
|
77 |
+
|
78 |
+
i2c_tiny_zh_obj.predict_step(
|
79 |
+
["https://datasets-server.huggingface.co/assets/poloclub/diffusiondb/--/2m_all/train/28/image/image.jpg"]
|
80 |
+
)
|
81 |
+
```
|
82 |
+
|
83 |
+
</br>
|
84 |
+
|
85 |
+
<div><img src='https://datasets-server.huggingface.co/assets/poloclub/diffusiondb/--/2m_all/train/28/image/image.jpg' width="550" height="450" /></div>
|
86 |
+
|
87 |
+
</br>
|
88 |
+
|
89 |
+
```json
|
90 |
+
['"一个年轻男人的肖像,由Greg Rutkowski创作"。Artstation上的趋势"。"《刀锋战士》的艺术作品"。高度细节化。"电影般的灯光"。超现实主义。锐利的焦点。辛烷�']
|
91 |
+
```
|