Views
No views yet
| Attribute | Description |
|---|---|
| Model Type | Vision-Language Multimodal Model |
| Architecture | InternVL3.5-style encoder-decoder |
| Languages | Chinese, English |
| Precision | bfloat16 |
| File Format | safetensors |
| Frameworks | PyTorch, ModelScope, Transformers |
| Primary Tasks | Image Question Answering, Optical Character Recognition, Multimodal Dialogue |
1import math
2import numpy as np
3import torch
4import torchvision.transforms as T
5from decord import VideoReader, cpu
6from PIL import Image
7from torchvision.transforms.functional import InterpolationMode
8from modelscope import AutoModel, AutoTokenizer
9
10IMAGENET_MEAN = (0.485, 0.456, 0.406)
11IMAGENET_STD = (0.229, 0.224, 0.225)
12
13def build_transform(input_size):
14 transform = T.Compose([
15 T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
16 T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
17 T.ToTensor(),
18 T.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
19 ])
20 return transform
21
22def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
23 best_ratio_diff = float('inf')
24 best_ratio = (1, 1)
25 area = width * height
26 for ratio in target_ratios:
27 target_aspect_ratio = ratio[0] / ratio[1]
28 ratio_diff = abs(aspect_ratio - target_aspect_ratio)
29 if ratio_diff < best_ratio_diff:
30 best_ratio_diff = ratio_diff
31 best_ratio = ratio
32 elif ratio_diff == best_ratio_diff:
33 if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
34 best_ratio = ratio
35 return best_ratio
36
37def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
38 orig_width, orig_height = image.size
39 aspect_ratio = orig_width / orig_height
40 target_ratios = set(
41 (i, j) for n in range(min_num, max_num + 1)
42 for i in range(1, n + 1)
43 for j in range(1, n + 1)
44 if i * j <= max_num and i * j >= min_num)
45 target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
46 target_aspect_ratio = find_closest_aspect_ratio(aspect_ratio, target_ratios, orig_width, orig_height, image_size)
47
48 target_width = image_size * target_aspect_ratio[0]
49 target_height = image_size * target_aspect_ratio[1]
50 blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
51
52 resized_img = image.resize((target_width, target_height))
53 processed_images = []
54 for i in range(blocks):
55 box = (
56 (i % (target_width // image_size)) * image_size,
57 (i // (target_width // image_size)) * image_size,
58 ((i % (target_width // image_size)) + 1) * image_size,
59 ((i // (target_width // image_size)) + 1) * image_size
60 )
61 split_img = resized_img.crop(box)
62 processed_images.append(split_img)
63 assert len(processed_images) == blocks
64 if use_thumbnail and len(processed_images) != 1:
65 thumbnail_img = image.resize((image_size, image_size))
66 processed_images.append(thumbnail_img)
67 return processed_images
68
69def load_image(image_file, input_size=448, max_num=12):
70 image = Image.open(image_file).convert('RGB')
71 transform = build_transform(input_size=input_size)
72 images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
73 pixel_values = [transform(im) for im in images]
74 pixel_values = torch.stack(pixel_values)
75 return pixel_values
76
77# -------------------------------------------
78# Load model & tokenizer
79path = "tianfu-lab/TianJiangZhuGe-8B" # Replace with your HF repo name
80model = AutoModel.from_pretrained(
81 path,
82 torch_dtype=torch.bfloat16,
83 trust_remote_code=True,
84 device_map="auto"
85).eval()
86
87tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
88
89# Load example image
90pixel_values = load_image("example_image.jpg", max_num=12).to(torch.bfloat16).cuda()
91generation_config = dict(max_new_tokens=1024, do_sample=True)
92
93question = "<image>\nWhat is shown in this picture?"
94response = model.chat(tokenizer, pixel_values, question, generation_config)
95print(f"User: {question}\nAssistant: {response}")| Task | Description |
|---|---|
| Image Captioning | Generate a natural language description for an image |
| Visual Question Answering (VQA) | Answer open-ended questions about images |
| Multimodal Dialogue | Conduct context-aware conversations conditioned on visual input |
| OCR-based Reasoning | Understand and reason over textual contents in images |