File size: 9,962 Bytes
4fe9c30 f764292 4fe9c30 f764292 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 |
---
license: mit
base_model:
- OpenGVLab/InternViT-300M-448px
- internlm/internlm2_5-7b-chat
new_version: OpenGVLab/PVC-InternVL-8B
language:
- multilingual
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- internvl
- video
- token compression
---
# PVC-InternVL2-8B
[\[📂 GitHub\]](https://github.com/OpenGVLab/PVC)
## Introduction
We introduce the **Progressive Visual Token Compression (PVC)** in large vision-language models (VLMs), which unifies the visual inputs as videos and progressively compresses vision tokens across video frames. Our PVC achieves:
* Preserve spatial details and temporal dynamics for both images and videos.
* Effectively reduce the tokens used for each video frame and image tile.
* SoTA performance on various video benchmarks, including long and fine-grained short video tasks.
* No performance loss on image benchmarks, especially on detail-sensitive tasks.
<div style="text-align: center;">
<img src="./assets/overview.png" width="70%"/>
</div>
## Results
Our implementation is based on the [InternVL2](https://github.com/OpenGVLab/InternVL) model, referred to as **PVC<sub>InternVL2</sub>**
### Video Understanding Benckmarks
| Model | LLaVA-OneVision-7B | Qwen2-VL-7B | InternVL2-8B | PVC<sub>InternVL2</sub>-8B |
| :--------------: | :--: | :--: | :--: | :--: |
| \# token/frame | 196 | - | 256 | 64 |
| | | | | |
| MVbench | 56.7 | 67.0 | 66.4 | 73.8 |
| VideoMME w/o-sub | 58.2 | 63.3 | 54.0 | 64.1 |
| VideoMME w-sub | 61.5 | 69.0 | 56.9 | 69.7 |
| MLVU | 64.7 | - | 52.0 | 72.4 |
| LongVideoBench | 56.5 | - | - | 59.2 |
| NextQA | 79.4 | - | - | 82.0 |
| Egoschema | 60.1 | 66.7 | 55.0 | 59.6 |
| PercepTest | 57.1 | 62.3 | 52.0 | 68.4 |
| AcNet-QA | 56.6 | - | - | 57.1 |
### Image Understanding Benckmarks
| Model | LLaVA-OneVision-7B | Qwen2-VL-7B | InternVL2-8B | PVC<sub>InternVL2</sub>-8B |
| :--------------------: | :--: | :--: | :--: | :--: |
| \# token/image tile | 729 | - | 256 | 64 |
| | | | | |
| AI2D<sub>test</sub> | 81.4 | 83.0 | 83.8 | 83.8 |
| ChartQA<sub>test</sub> | 80.0 | 83.0 | 83.3 | 84.1 |
| DocVQA<sub>test</sub> | 87.5 | 94.5 | 91.6 | 92.5 |
| InfoVQA<sub>test</sub> | 68.8 | 76.5 | 74.8 | 75.0 |
| SQA<sub>test</sub> | 96.0 | - | 97.1 | 97.7 |
| TextVQA<sub>val</sub> | - | 84.3 | 77.4 | 80.0 |
| MMB<sub>en-test</sub> | - | 83.0 | 81.7 | 83.9 |
| MME<sub>sum</sub> | 1998 | 2327 | 2210 | 2282 |
| MMMU<sub>val</sub> | 48.8 | 54.1 | 49.3 | 50.9 |
| SEED<sub>I</sub> | 75.4 | - | 76.2 | 77.2 |
| OCRBench | - | 866 | 794 | 807 |
## Quick Start
```python
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(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
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
path = 'OpenGVLab/PVC-InternVL2-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
generation_config = dict(max_new_tokens=1024, do_sample=True)
# single-image conversation
pixel_values = load_image('./assets/example_image1.jpg', max_num=12).to(torch.bfloat16).cuda()
data_flag = torch.tensor([1], dtype=torch.long).cuda()
question = '<image>\nWhat is in the image?'
response = model.chat(tokenizer, pixel_values, question, generation_config, data_flag=data_flag)
print(f'User: {question}\nAssistant: {response}')
# multi-image conversation
pixel_values1 = load_image('./assets/example_image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./assets/example_image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
data_flag = torch.tensor([2], dtype=torch.long).cuda()
num_patches_list = [pixel_values1.shape[0], pixel_values2.shape[0]]
question = 'Image-1: <image>\nImage-2: <image>\nWhat are the similarities and differences between these two images.'
response = model.chat(tokenizer, pixel_values, question, generation_config, data_flag=data_flag, num_patches_list=num_patches_list)
print(f'User: {question}\nAssistant: {response}')
# video conversation
pixel_values, num_patches_list = load_video('./assets/example_video.mp4', num_segments=64, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
# Frame1: <image>\nFrame2: <image>\n...\nFrameN: <image>\n{question}
data_flag = torch.tensor([3], dtype=torch.long).cuda()
question = video_prefix + 'Describe this video in detail.'
response = model.chat(tokenizer, pixel_values, question, generation_config, data_flag=data_flag, num_patches_list=num_patches_list)
print(f'User: {question}\nAssistant: {response}')
```
## License
This project is released under the MIT license. Parts of this project contain code and models from other sources, which are subject to their respective licenses.
|