Commit
•
5319378
1
Parent(s):
f6b8980
Upload README.md
Browse files
README.md
ADDED
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
tags:
|
3 |
+
- image-classification
|
4 |
+
library_name: generic
|
5 |
+
---
|
6 |
+
|
7 |
+
## Example
|
8 |
+
|
9 |
+
The model is by no means a state-of-the-art model, but nevertheless
|
10 |
+
produces reasonable image captioning results. It was mainly fine-tuned
|
11 |
+
as a proof-of-concept for the 🤗 FlaxVisionEncoderDecoder Framework.
|
12 |
+
|
13 |
+
The model can be used as follows:
|
14 |
+
|
15 |
+
**In PyTorch**
|
16 |
+
```python
|
17 |
+
|
18 |
+
import torch
|
19 |
+
import requests
|
20 |
+
from PIL import Image
|
21 |
+
from transformers import ViTFeatureExtractor, AutoTokenizer, VisionEncoderDecoderModel
|
22 |
+
|
23 |
+
|
24 |
+
loc = "ydshieh/vit-gpt2-coco-en"
|
25 |
+
|
26 |
+
feature_extractor = ViTFeatureExtractor.from_pretrained(loc)
|
27 |
+
tokenizer = AutoTokenizer.from_pretrained(loc)
|
28 |
+
model = VisionEncoderDecoderModel.from_pretrained(loc)
|
29 |
+
model.eval()
|
30 |
+
|
31 |
+
|
32 |
+
def predict(image):
|
33 |
+
|
34 |
+
pixel_values = feature_extractor(images=image, return_tensors="pt").pixel_values
|
35 |
+
|
36 |
+
with torch.no_grad():
|
37 |
+
output_ids = model.generate(pixel_values, max_length=16, num_beams=4, return_dict_in_generate=True).sequences
|
38 |
+
|
39 |
+
preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
40 |
+
preds = [pred.strip() for pred in preds]
|
41 |
+
|
42 |
+
return preds
|
43 |
+
|
44 |
+
|
45 |
+
# We will verify our results on an image of cute cats
|
46 |
+
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
47 |
+
with Image.open(requests.get(url, stream=True).raw) as image:
|
48 |
+
preds = predict(image)
|
49 |
+
|
50 |
+
print(preds)
|
51 |
+
# should produce
|
52 |
+
# ['a cat laying on top of a couch next to another cat']
|
53 |
+
|
54 |
+
```
|
55 |
+
|
56 |
+
**In Flax**
|
57 |
+
```python
|
58 |
+
|
59 |
+
import jax
|
60 |
+
import requests
|
61 |
+
from PIL import Image
|
62 |
+
from transformers import ViTFeatureExtractor, AutoTokenizer, FlaxVisionEncoderDecoderModel
|
63 |
+
|
64 |
+
|
65 |
+
loc = "ydshieh/vit-gpt2-coco-en"
|
66 |
+
|
67 |
+
feature_extractor = ViTFeatureExtractor.from_pretrained(loc)
|
68 |
+
tokenizer = AutoTokenizer.from_pretrained(loc)
|
69 |
+
model = FlaxVisionEncoderDecoderModel.from_pretrained(loc)
|
70 |
+
|
71 |
+
gen_kwargs = {"max_length": 16, "num_beams": 4}
|
72 |
+
|
73 |
+
|
74 |
+
# This takes sometime when compiling the first time, but the subsequent inference will be much faster
|
75 |
+
@jax.jit
|
76 |
+
def generate(pixel_values):
|
77 |
+
output_ids = model.generate(pixel_values, **gen_kwargs).sequences
|
78 |
+
return output_ids
|
79 |
+
|
80 |
+
|
81 |
+
def predict(image):
|
82 |
+
|
83 |
+
pixel_values = feature_extractor(images=image, return_tensors="np").pixel_values
|
84 |
+
output_ids = generate(pixel_values)
|
85 |
+
preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
86 |
+
preds = [pred.strip() for pred in preds]
|
87 |
+
|
88 |
+
return preds
|
89 |
+
|
90 |
+
|
91 |
+
# We will verify our results on an image of cute cats
|
92 |
+
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
93 |
+
with Image.open(requests.get(url, stream=True).raw) as image:
|
94 |
+
preds = predict(image)
|
95 |
+
|
96 |
+
print(preds)
|
97 |
+
# should produce
|
98 |
+
# ['a cat laying on top of a couch next to another cat']
|
99 |
+
|
100 |
+
```
|