Views
No views yet
1
2from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
3import torch
4from PIL import Image
5
6model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
7feature_extractor = ViTImageProcessor.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
8tokenizer = AutoTokenizer.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
9
10device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
11model.to(device)
12
13
14
15max_length = 16
16num_beams = 4
17gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
18def 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
37predict_step(['doctor.e16ba4e4.jpg']) # ['a woman in a hospital bed with a woman in a hospital bed']
381
2from transformers import pipeline
3
4image_to_text = pipeline("image-to-text", model="nlpconnect/vit-gpt2-image-captioning")
5
6image_to_text("https://ankur3107.github.io/assets/images/image-captioning-example.png")
7
8# [{'generated_text': 'a soccer game with a player jumping to catch the ball '}]
9
10