evarest commited on
Commit
f9280b4
·
1 Parent(s): 9a61d5d

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +37 -0
app.py CHANGED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
3
+ import torch
4
+ from PIL import Image
5
+
6
+ model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
7
+ feature_extractor = ViTImageProcessor.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
8
+ tokenizer = AutoTokenizer.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
9
+
10
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
11
+ model.to(device)
12
+
13
+
14
+
15
+ max_length = 16
16
+ num_beams = 4
17
+ gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
18
+ def predict_step(image_paths):
19
+ images = []
20
+ for image_path in image_paths:
21
+ i_image = Image.open(image_path)
22
+ if i_image.mode != "RGB":
23
+ i_image = i_image.convert(mode="RGB")
24
+
25
+ images.append(i_image)
26
+
27
+ pixel_values = feature_extractor(images=images, return_tensors="pt").pixel_values
28
+ pixel_values = pixel_values.to(device)
29
+
30
+ output_ids = model.generate(pixel_values, **gen_kwargs)
31
+
32
+ preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
33
+ preds = [pred.strip() for pred in preds]
34
+ return preds
35
+
36
+
37
+ predict_step(['doctor.e16ba4e4.jpg']) # ['a woman in a hospital bed with a woman in a hospital bed']