intheashtray commited on
Commit
ff68b4e
·
verified ·
1 Parent(s): 3251538

Create script/preprocess.py

Browse files
Files changed (1) hide show
  1. script/preprocess.py +80 -0
script/preprocess.py ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ import os
3
+ from datasets import Dataset, Image
4
+ from transformers import CLIPProcessor, CLIPModel, AutoTokenizer
5
+ from PIL import Image as PILImage
6
+ import torch
7
+ from tqdm import tqdm
8
+
9
+ def preprocess_multimodal_data(raw_json_path, image_dir, dataset_name, tokenizer_name, clip_model_name, output_dir):
10
+ """
11
+ Preprocesses raw JSON and image data into a Hugging Face Dataset and saves it to a repository.
12
+
13
+ Args:
14
+ raw_json_path (str): Path to the raw JSON file.
15
+ image_dir (str): Directory containing the images.
16
+ dataset_name (str): Name of the Hugging Face Dataset.
17
+ tokenizer_name (str): Name of the tokenizer from Hugging Face Transformers.
18
+ clip_model_name (str): Name of the CLIP model from Hugging Face Transformers.
19
+ output_dir (str): Directory where the dataset repository will be saved.
20
+ """
21
+
22
+ # Load data from JSON
23
+ with open(raw_json_path, 'r') as f:
24
+ data = json.load(f)
25
+
26
+ # Initialize lists to store data
27
+ images = []
28
+ questions = []
29
+ answers = []
30
+ ids = [] #add an id field.
31
+
32
+ # Load CLIP and Tokenizer
33
+ clip_processor = CLIPProcessor.from_pretrained(clip_model_name)
34
+ clip_model = CLIPModel.from_pretrained(clip_model_name).to("cuda")
35
+ tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
36
+
37
+ # Process each data point
38
+ for i, item in enumerate(tqdm(data, desc="Processing data")):
39
+ image_path = os.path.join(image_dir, item['image'])
40
+ question = item['question']
41
+ answer = item.get('answer', '') #handle cases where answer is missing.
42
+ id_field = str(i) #create id field.
43
+
44
+ try:
45
+ image = PILImage.open(image_path).convert("RGB") # Ensure RGB
46
+ images.append(image)
47
+ questions.append(question)
48
+ answers.append(answer)
49
+ ids.append(id_field)
50
+
51
+ except FileNotFoundError:
52
+ print(f"Warning: Image not found at {image_path}")
53
+ except Exception as e:
54
+ print(f"Warning: Error processing image {image_path}: {e}")
55
+
56
+ # Create Hugging Face Dataset
57
+ dataset_dict = {
58
+ 'image': images,
59
+ 'question': questions,
60
+ 'answer': answers,
61
+ 'id': ids,
62
+ }
63
+
64
+ dataset = Dataset.from_dict(dataset_dict)
65
+ dataset = dataset.cast_column("image", Image())
66
+
67
+ # Save to Hugging Face Dataset repository
68
+ dataset.save_to_disk(os.path.join(output_dir, dataset_name))
69
+
70
+ print(f"Dataset '{dataset_name}' saved to {os.path.join(output_dir, dataset_name)}")
71
+
72
+ # Example Usage:
73
+ raw_json_path = 'your_data.json' # replace with your json path.
74
+ image_dir = 'your_image_dir' # replace with your image directory.
75
+ dataset_name = 'my_multimodal_dataset'
76
+ tokenizer_name = 'deepseek-ai/deepseek-coder-1.3b-instruct'
77
+ clip_model_name = 'openai/clip-vit-base-patch32'
78
+ output_dir = 'my_datasets'
79
+
80
+ preprocess_multimodal_data(raw_json_path, image_dir, dataset_name, tokenizer_name, clip_model_name, output_dir)