Pytorch implementation of "Training a 85.4% Top-1 Accuracy Vision Transformer with 56M Parameters on ImageNet"

Overview

Token Labeling: Training an 85.4% Top-1 Accuracy Vision Transformer with 56M Parameters on ImageNet (arxiv)

This is a Pytorch implementation of our technical report.

Compare

Comparison between the proposed LV-ViT and other recent works based on transformers. Note that we only show models whose model sizes are under 100M.

Training Pipeline

Pipeline

Our codes are based on the pytorch-image-models by Ross Wightman.

LV-ViT Models

Model layer dim Image resolution Param Top 1 Download
LV-ViT-S 16 384 224 26.15M 83.3 link
LV-ViT-S 16 384 384 26.30M 84.4 link
LV-ViT-M 20 512 224 55.83M 84.0 link
LV-ViT-M 20 512 384 56.03M 85.4 link
LV-ViT-L 24 768 448 150.47M 86.2 link

Requirements

torch>=1.4.0 torchvision>=0.5.0 pyyaml timm==0.4.5

data prepare: ImageNet with the following folder structure, you can extract imagenet by this script.

│imagenet/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

Validation

Replace DATA_DIR with your imagenet validation set path and MODEL_DIR with the checkpoint path

CUDA_VISIBLE_DEVICES=0 bash eval.sh /path/to/imagenet/val /path/to/checkpoint

Label data

We provide NFNet-F6 generated dense label map here. As NFNet-F6 are based on pure ImageNet data, no extra training data is involved.

Training

Coming soon

Reference

If you use this repo or find it useful, please consider citing:

@misc{jiang2021token,
      title={Token Labeling: Training an 85.4% Top-1 Accuracy Vision Transformer with 56M Parameters on ImageNet}, 
      author={Zihang Jiang and Qibin Hou and Li Yuan and Daquan Zhou and Xiaojie Jin and Anran Wang and Jiashi Feng},
      year={2021},
      eprint={2104.10858},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Related projects

T2T-ViT, Re-labeling ImageNet.

Comments
  • error: download the pretrained model but couldn't be unzipped

    error: download the pretrained model but couldn't be unzipped

    tar -xvf lvvit_s-26M-384-84-4.pth.tar tar: This does not look like a tar archive tar: Skipping to next header tar: Exiting with failure status due to previous errors

    opened by Williamlizl 10
  • The accuracy of the validation set is 0,and the loss is always around 13

    The accuracy of the validation set is 0,and the loss is always around 13

    Hello! I use ILSVRC2012_img_train and ILSVRC2012_img_val, and use the provided label_top5_train_nfnet from Google Drive. I train lv-vit-s with batch_size 64 without apex for one epoch. Thanks for your advice.

    opened by yifanQi98 7
  • Pretrained weights for LV-ViT-T

    Pretrained weights for LV-ViT-T

    Hi,

    Thanks for sharing your work. Could you also provide the pre-trained weights for the LV-ViT-T model variant, the one that achieves 79.1% top1-acc. as mentioned in Table 1 of your paper?

    All the best, Marc

    opened by marc345 5
  • train error: AttributeError: 'tuple' object has no attribute 'log_softmax'

    train error: AttributeError: 'tuple' object has no attribute 'log_softmax'

    Hi, thanks for you great work. When I train script, some error occurs: AttributeError: 'tuple' object has no attribute 'log_softmax'

    with amp_autocast():   
                output = model(input)  
                loss = loss_fn(output, target)  # error occurs
    
    

    and loss function is train_loss_fn = LabelSmoothingCrossEntropy(smoothing=0.0).cuda()

    by the way: Could you please tell me why we need to specify smoothing=0.0?

    opened by lxy5513 5
  • RuntimeError: CUDA error: device-side assert triggered

    RuntimeError: CUDA error: device-side assert triggered

    I am a green hand of DL. When I run the code of volo with tlt in a single or multi GPU, I get an error as follows: /pytorch/aten/src/ATen/native/cuda/ScatterGatherKernel.cu:312: operator(): block: [0,0,0], thread: [25,0,0] Assertion idx_dim >= 0 && idx_dim < index_size && "index out of bounds" failed. Traceback (most recent call last): File "main.py", line 949, in main() File "main.py", line 664, in main optimizers=optimizers) File "main.py", line 773, in train_one_epoch label_size=args.token_label_size) File "/opt/conda/lib/python3.6/site-packages/tlt/data/mixup.py", line 90, in mixup_target y1 = get_labelmaps_with_coords(target, num_classes, on_value=on_value, off_value=off_value, device=device, label_size=label_size) File "/opt/conda/lib/python3.6/site-packages/tlt/data/mixup.py", line 64, in get_labelmaps_with_coords num_classes=num_classes,device=device) File "/opt/conda/lib/python3.6/site-packages/tlt/data/mixup.py", line 16, in get_featuremaps _label_topk[1][:, :, :].long(), RuntimeError: CUDA error: device-side assert triggered.

    I can't fix this problem right now.

    opened by JIAOJIAYUASD 4
  • Generating label for custom dataset

    Generating label for custom dataset

    Hello,

    Thank you for sharing your work. I am currently trying to generate token label to a custom dataset for model lvvit_s, but I keep getting the loss close to 7 and the Accuracy 0 (not pre-trained and using 1 GPU in Google Colab). I also tried using the pre-trained model with --transfer but got 0 in both Loss and Acc . What option should I use for a custom dataset? image

    opened by AleMaiaF 2
  • generate_label.py unable to find model lvvit_s

    generate_label.py unable to find model lvvit_s

    Hi,

    When I tried to run the label generation script for the model lvvit_s it returned an error "RuntimeError: Unknown model".

    Solution: It worked when I added the line "import tlt.models" in the file generate_label.py.

    opened by AleMaiaF 2
  • Can Token labeling reach higher than annotator model?

    Can Token labeling reach higher than annotator model?

    Greetings,

    Thank you for this incredible research.

    I would like to know if it is possible to use Token Labeling to achieve scores higher than that of the annotator model, I believe this was the case with VOLO D5 model where it achieved higher score than NFNet, model used for annotation.

    opened by ErenBalatkan 1
  • label_map does not do the same augmentation (random crop) as the input image

    label_map does not do the same augmentation (random crop) as the input image

    Hi Thanks so much for the nice work! I am curious if you could share the insight on processing of the label_map. If I understand it correctly, after we load image and the corresponding, we shall do the same cropping/ flip/ resize, but in https://github.com/zihangJiang/TokenLabeling/blob/aa438eff9b9fc2daa8c8b4cc6bfaa6e3721f995e/tlt/data/label_transforms_factory.py#L58-L73 Seems only image was cropped, but the label map does not do the same cropping, which make the label map not match with the image?

    Shall we do

            return torchvision_F.resized_crop(
                    img, i, j, h, w, self.size, interpolation
            ), torchvision_F.resized_crop(
                    label_map, i / ratio, j / ratio, h / ratio, w / ratio, self.size, interpolation
            )
    

    Thanks

    opened by haooooooqi 1
  • Python3.6, ok; Python3.8, error

    Python3.6, ok; Python3.8, error

    Test: [ 0/1] Time: 11.293 (11.293) Loss: 0.7043 (0.7043) [email protected]: 42.1875 (42.1875) [email protected]: 100.0000 (100.0000) Test: [ 1/1] Time: 0.108 (5.701) Loss: 0.5847 (0.6689) [email protected]: 89.8148 (56.3187) [email protected]: 100.0000 (100.0000) free(): invalid pointer free(): invalid pointer Traceback (most recent call last): File "/opt/conda/lib/python3.8/runpy.py", line 194, in _run_module_as_main return _run_code(code, main_globals, None, File "/opt/conda/lib/python3.8/runpy.py", line 87, in _run_code exec(code, run_globals) File "/opt/conda/lib/python3.8/site-packages/torch/distributed/launch.py", line 303, in <module> main() File "/opt/conda/lib/python3.8/site-packages/torch/distributed/launch.py", line 294, in main raise subprocess.CalledProcessError(returncode=process.returncode, subprocess.CalledProcessError: Command '['/opt/conda/bin/python3.8', '-u', 'main.py', '--local_rank=1', './dataset/c/c', '--model', 'lvvit_s', '-b', '128', '--apex-amp', '--img-size', '224', '--drop-path', '0.1', '--token-label', '--token-label-size', '14', '--dense-weight', '0.0', '--num-classes', '2', '--finetune', './pretrained/lvvit_s-26M-384-84-4.pth.tar']' died with <Signals.SIGABRT: 6>. [email protected]:/puxin_libochao/TokenLabeling# CUDA_VISIBLE_DEVICES=0,1 bash ./distributed_train.sh 2 ./dataset/c/c --model lvvit_s -b 128 --apex-amp --img-size 224 --drop-path 0.1 --token-label --token-label-size 14 --dense-weight 0.0 --num-classes 2 --finetune ./pretrained/lvvit_s-26M-384-84-4.pth.tar

    opened by Williamlizl 1
  • A Bag of Training Techniques for ViT

    A Bag of Training Techniques for ViT

    Hi, thanks for your wonderful work. I have a question that whether training techniques mentioned in the LV-Vit can be used in other downstream task like object detection? In your paper, I see that many of this techniques are used in ImageNet. Thanks!

    opened by qdd1234 1
  • how to apply token labeling to CNN ?

    how to apply token labeling to CNN ?

    Hello ~ I'm interested in your token labeling technique, So I want to apply this technique in CNN based model because ViT is very heavy to train.

    can I get the your code with CNN token labeling? if you're not give me some detail for implementing

    thank you.

    opened by HoJ00n2 0
  • Model settings for Cifar10

    Model settings for Cifar10

    I am interested if there is any LV-ViT- model setup you have tested for Cifar10. I would like to know the proper setup of all blocks in none pretrained weights settings.

    opened by Aminullah6264 0
Owner
蒋子航
Now a Ph.D. student supervised by Prof. Feng Jiashi in ECE, NUS.
蒋子航
Official implementation of the method ContIG, for self-supervised learning from medical imaging with genomics

ContIG: Self-supervised Multimodal Contrastive Learning for Medical Imaging with Genetics This is the code implementation of the paper "ContIG: Self-s

Digital Health & Machine Learning 22 Dec 13, 2022
Official Implementation of "DialogLM: Pre-trained Model for Long Dialogue Understanding and Summarization."

DialogLM Code for AAAI 2022 paper: DialogLM: Pre-trained Model for Long Dialogue Understanding and Summarization. Pre-trained Models We release two ve

Microsoft 92 Dec 19, 2022
Gym-TORCS is the reinforcement learning (RL) environment in TORCS domain with OpenAI-gym-like interface.

Gym-TORCS Gym-TORCS is the reinforcement learning (RL) environment in TORCS domain with OpenAI-gym-like interface. TORCS is the open-rource realistic

naoto yoshida 400 Dec 27, 2022
A graph-to-sequence model for one-step retrosynthesis and reaction outcome prediction.

Graph2SMILES A graph-to-sequence model for one-step retrosynthesis and reaction outcome prediction. 1. Environmental setup System requirements Ubuntu:

29 Nov 18, 2022
An implementation for `Text2Event: Controllable Sequence-to-Structure Generation for End-to-end Event Extraction`

Text2Event An implementation for Text2Event: Controllable Sequence-to-Structure Generation for End-to-end Event Extraction Please contact Yaojie Lu (@

Roger 153 Jan 07, 2023
PyTorch implementation of Masked Autoencoders Are Scalable Vision Learners for self-supervised ViT.

MAE for Self-supervised ViT Introduction This is an unofficial PyTorch implementation of Masked Autoencoders Are Scalable Vision Learners for self-sup

36 Oct 30, 2022
Baseline model for "GraspNet-1Billion: A Large-Scale Benchmark for General Object Grasping" (CVPR 2020)

GraspNet Baseline Baseline model for "GraspNet-1Billion: A Large-Scale Benchmark for General Object Grasping" (CVPR 2020). [paper] [dataset] [API] [do

GraspNet 209 Dec 29, 2022
[CVPR2021] The source code for our paper 《Removing the Background by Adding the Background: Towards Background Robust Self-supervised Video Representation Learning》.

TBE The source code for our paper "Removing the Background by Adding the Background: Towards Background Robust Self-supervised Video Representation Le

Jinpeng Wang 150 Dec 28, 2022
Code for generating a single image pretraining dataset

Single Image Pretraining of Visual Representations As shown in the paper A critical analysis of self-supervision, or what we can learn from a single i

Yuki M. Asano 12 Dec 19, 2022
Python library for analysis of time series data including dimensionality reduction, clustering, and Markov model estimation

deeptime Releases: Installation via conda recommended. conda install -c conda-forge deeptime pip install deeptime Documentation: deeptime-ml.github.io

495 Dec 28, 2022
Official repository for the paper "Going Beyond Linear Transformers with Recurrent Fast Weight Programmers"

Recurrent Fast Weight Programmers This is the official repository containing the code we used to produce the experimental results reported in the pape

IDSIA 36 Nov 15, 2022
For encoding a text longer than 512 tokens, for example 800. Set max_pos to 800 during both preprocessing and training.

LongScientificFormer For encoding a text longer than 512 tokens, for example 800. Set max_pos to 800 during both preprocessing and training. Some code

Athar Sefid 6 Nov 02, 2022
Code for our ACL 2021 paper - ConSERT: A Contrastive Framework for Self-Supervised Sentence Representation Transfer

ConSERT Code for our ACL 2021 paper - ConSERT: A Contrastive Framework for Self-Supervised Sentence Representation Transfer Requirements torch==1.6.0

Yan Yuanmeng 478 Dec 25, 2022
This is an official implementation for "SimMIM: A Simple Framework for Masked Image Modeling".

SimMIM By Zhenda Xie*, Zheng Zhang*, Yue Cao*, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai and Han Hu*. This repo is the official implementation of

Microsoft 674 Dec 26, 2022
A simple but complete full-attention transformer with a set of promising experimental features from various papers

x-transformers A concise but fully-featured transformer, complete with a set of promising experimental features from various papers. Install $ pip ins

Phil Wang 2.3k Jan 03, 2023
Easily benchmark PyTorch model FLOPs, latency, throughput, max allocated memory and energy consumption

⏱ pytorch-benchmark Easily benchmark model inference FLOPs, latency, throughput, max allocated memory and energy consumption Install pip install pytor

Lukas Hedegaard 21 Dec 22, 2022
PFENet: Prior Guided Feature Enrichment Network for Few-shot Segmentation (TPAMI).

PFENet This is the implementation of our paper PFENet: Prior Guided Feature Enrichment Network for Few-shot Segmentation that has been accepted to IEE

DV Lab 230 Dec 31, 2022
Populating 3D Scenes by Learning Human-Scene Interaction https://posa.is.tue.mpg.de/

Populating 3D Scenes by Learning Human-Scene Interaction [Project Page] [Paper] License Software Copyright License for non-commercial scientific resea

Mohamed Hassan 81 Nov 08, 2022
DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.

DeepSpeed+Megatron trained the world's most powerful language model: MT-530B DeepSpeed is hiring, come join us! DeepSpeed is a deep learning optimizat

Microsoft 8.4k Dec 28, 2022
Image Super-Resolution by Neural Texture Transfer

SRNTT: Image Super-Resolution by Neural Texture Transfer Tensorflow implementation of the paper Image Super-Resolution by Neural Texture Transfer acce

Zhifei Zhang 413 Nov 30, 2022