Views
No views yet
1import math
2import numpy as np
3import torch
4import torchvision.transforms as T
5from decord import VideoReader, cpu
6import requests
7from io import BytesIO
8from PIL import Image
9from torchvision.transforms.functional import InterpolationMode
10from transformers import AutoModel, AutoTokenizer, AutoConfig
11
12IMAGENET_MEAN = (0.485, 0.456, 0.406)
13IMAGENET_STD = (0.229, 0.224, 0.225)
14
15
16def build_transform(input_size):
17 MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
18 transform = T.Compose([
19 T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
20 T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
21 T.ToTensor(),
22 T.Normalize(mean=MEAN, std=STD)
23 ])
24 return transform
25
26
27def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
28 best_ratio_diff = float('inf')
29 best_ratio = (1, 1)
30 area = width * height
31 for ratio in target_ratios:
32 target_aspect_ratio = ratio[0] / ratio[1]
33 ratio_diff = abs(aspect_ratio - target_aspect_ratio)
34 if ratio_diff < best_ratio_diff:
35 best_ratio_diff = ratio_diff
36 best_ratio = ratio
37 elif ratio_diff == best_ratio_diff:
38 if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
39 best_ratio = ratio
40 return best_ratio
41
42
43def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
44 orig_width, orig_height = image.size
45 aspect_ratio = orig_width / orig_height
46
47 # calculate the existing image aspect ratio
48 target_ratios = set(
49 (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
50 i * j <= max_num and i * j >= min_num)
51 target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
52
53 # find the closest aspect ratio to the target
54 target_aspect_ratio = find_closest_aspect_ratio(
55 aspect_ratio, target_ratios, orig_width, orig_height, image_size)
56
57 # calculate the target width and height
58 target_width = image_size * target_aspect_ratio[0]
59 target_height = image_size * target_aspect_ratio[1]
60 blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
61
62 # resize the image
63 resized_img = image.resize((target_width, target_height))
64 processed_images = []
65 for i in range(blocks):
66 box = (
67 (i % (target_width // image_size)) * image_size,
68 (i // (target_width // image_size)) * image_size,
69 ((i % (target_width // image_size)) + 1) * image_size,
70 ((i // (target_width // image_size)) + 1) * image_size
71 )
72 # split the image
73 split_img = resized_img.crop(box)
74 processed_images.append(split_img)
75 assert len(processed_images) == blocks
76 if use_thumbnail and len(processed_images) != 1:
77 thumbnail_img = image.resize((image_size, image_size))
78 processed_images.append(thumbnail_img)
79 return processed_images
80
81
82def load_image(image_file, input_size=448, max_num=12):
83 if isinstance(image_file, str) and image_file.startswith(('http://', 'https://')):
84 response = requests.get(image_file, timeout=10)
85 response.raise_for_status()
86 image = Image.open(BytesIO(response.content)).convert('RGB')
87 else:
88 image = Image.open(image_file).convert('RGB')
89
90 transform = build_transform(input_size=input_size)
91 images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
92 pixel_values = [transform(image) for image in images]
93 pixel_values = torch.stack(pixel_values)
94 return pixel_values
95
96
97def split_model(model_name):
98 device_map = {}
99 world_size = torch.cuda.device_count()
100 config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
101 num_layers = config.llm_config.num_hidden_layers
102 # Since the first GPU will be used for ViT, treat it as half a GPU.
103 num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
104 num_layers_per_gpu = [num_layers_per_gpu] * world_size
105 num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
106 layer_cnt = 0
107 for i, num_layer in enumerate(num_layers_per_gpu):
108 for j in range(num_layer):
109 device_map[f'language_model.model.layers.{layer_cnt}'] = i
110 layer_cnt += 1
111 device_map['vision_model'] = 0
112 device_map['mlp1'] = 0
113 device_map['language_model.model.tok_embeddings'] = 0
114 device_map['language_model.model.embed_tokens'] = 0
115 device_map['language_model.output'] = 0
116 device_map['language_model.model.norm'] = 0
117 device_map['language_model.model.rotary_emb'] = 0
118 device_map['language_model.lm_head'] = 0
119 device_map[f'language_model.model.layers.{num_layers - 1}'] = 0
120
121 return device_map
122
123
124# If you set `load_in_8bit=True`, you will need two 80GB GPUs.
125# If you set `load_in_8bit=False`, you will need at least three 80GB GPUs.
126path = 'jihyoung/AQuA-InternVL'
127device_map = split_model(path)
128model = AutoModel.from_pretrained(
129 path,
130 torch_dtype=torch.bfloat16,
131 load_in_8bit=False,
132 low_cpu_mem_usage=True,
133 use_flash_attn=True,
134 trust_remote_code=True,
135 device_map=device_map).eval()
136tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
137
138# set the max number of tiles in `max_num`
139pixel_values = load_image('http://images.cocodataset.org/train2017/000000539311.jpg', max_num=12).to(torch.bfloat16).cuda()
140generation_config = dict(max_new_tokens=1024, do_sample=True)
141
142question = '<image>\nWhat color is this bat?'
143response = model.chat(tokenizer, pixel_values, question, generation_config)
144print(f'User: {question}\nAssistant: {response}')1@inproceedings{
2 jang2026aqua,
3 title={{AQ}uA: Toward Strategic Response Generation for Ambiguous Visual Questions},
4 author={Jihyoung Jang and Hyounghun Kim},
5 booktitle={The Fourteenth International Conference on Learning Representations},
6 year={2026},
7 url={https://openreview.net/forum?id=7b1MpD6IF8}
8}