Implementation of "A MLP-like Architecture for Dense Prediction"

Related tags

Deep LearningCycleMLP
Overview

A MLP-like Architecture for Dense Prediction (arXiv)

License: MIT Python 3.8

    

Updates

  • (22/07/2021) Initial release.

Model Zoo

We provide CycleMLP models pretrained on ImageNet 2012.

Model Parameters FLOPs Top 1 Acc. Download
CycleMLP-B1 15M 2.1G 78.9% model
CycleMLP-B2 27M 3.9G 81.6% model
CycleMLP-B3 38M 6.9G 82.4% model
CycleMLP-B4 52M 10.1G 83.0% model
CycleMLP-B5 76M 12.3G 83.2% model

Usage

Install

  • PyTorch 1.7.0+ and torchvision 0.8.1+
  • timm:
pip install 'git+https://github.com/rwightman/[email protected]'

or

git clone https://github.com/rwightman/pytorch-image-models
cd pytorch-image-models
git checkout c2ba229d995c33aaaf20e00a5686b4dc857044be
pip install -e .
  • fvcore (optional, for FLOPs calculation)
  • mmcv, mmdetection, mmsegmentation (optional)

Data preparation

Download and extract ImageNet train and val images from http://image-net.org/. The directory structure is:

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

Evaluation

To evaluate a pre-trained CycleMLP-B5 on ImageNet val with a single GPU run:

python main.py --eval --model CycleMLP_B5 --resume path/to/CycleMLP_B5.pth --data-path /path/to/imagenet

Training

To train CycleMLP-B5 on ImageNet on a single node with 8 gpus for 300 epochs run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --model CycleMLP_B5 --batch-size 128 --data-path /path/to/imagenet --output_dir /path/to/save

Acknowledgement

This code is based on DeiT and pytorch-image-models. Thanks for their wonderful works

Citing

@article{chen2021cyclemlp,
  title={CycleMLP: A MLP-like Architecture for Dense Prediction},
  author={Chen, Shoufa and Xie, Enze and Ge, Chongjian and Liang, Ding and Luo, Ping},
  journal={arXiv preprint arXiv:2107.10224},
  year={2021}
}

License

CycleMLP is released under MIT License.

Comments
  • detection result

    detection result

    Applying PVT detection framework, I tried a CycleMLP-B1 based detector with RetinaNet 1x. I got AP=27.1, fairly inferior to the reported 38.6. Could you give some advices to reproduce the reported result?

    The specific configure is as follows

    base = [ 'base/models/retinanet_r50_fpn.py', 'base/datasets/coco_detection.py', 'base/schedules/schedule_1x.py', 'base/default_runtime.py' ] #optimizer model = dict( pretrained='./pretrained/CycleMLP_B1.pth', backbone=dict( type='CycleMLP_B1_feat', style='pytorch'), neck=dict( type='FPN', in_channels=[64, 128, 320, 512], out_channels=256, start_level=1, add_extra_convs='on_input', num_outs=5)) #optimizer optimizer = dict(delete=True, type='AdamW', lr=0.0001, weight_decay=0.0001) optimizer_config = dict(grad_clip=None)

    find_unused_parameters = True

    opened by mountain111 6
  • Compiling CycleMLP

    Compiling CycleMLP

    Thank you for this great repo and interesting paper.

    I tried compiling CycleMLP to onnx and not surpassingly the process failed since CycleMLP include dynamic offset creation in https://github.com/ShoufaChen/CycleMLP/blob/main/cycle_mlp.py#L132 and as such cannot be converted to a frozen graph. Were you able to convert CycleMLP to onnx or any other frozen graph framework?

    Thanks in advance.

    opened by shairoz-deci 6
  • Questions about offset calculation

    Questions about offset calculation

    Hi, thanks for your wonderful work.

    I'm currently studying your work, and come up with some question about the offset calculations.

    I understood the offset calculation mentioned on the paper, but can't understand about how generated offset is being used in the code.

    For ex) if $S_H \times S_W : 3 \times 1$; I understood how the offset is applied in this figure 스크린샷 2022-06-13 오후 9 18 20

    by calculate like this: 스크린샷 2022-06-13 오후 9 19 57

    However, when I run the offset generating code, I can't figure out how this offset is being used in deform_conv2d 스크린샷 2022-06-13 오후 9 21 57

    Can you provide more detailed information about this??

    And also, the paper contains how $S_H \times S_W: 3 \times 3$ works, but in the code, it seems like either one ofkernel_size[0] or kernel_size[1] has to be 1. So, if I want to use $S_H \times S_W : 3 \times 3$, do I have to make $3 \times 1$ and $1 \times 3$ offsets and add those together?

    Thank you again for your work. I really learned a lot.

    opened by tae-mo 5
  • Example of CycleMLP Configuration for Dense Prediction

    Example of CycleMLP Configuration for Dense Prediction

    Hello.

    First of all, thank you for curating this interesting work. I was wondering, are there any working examples of how I can use CycleMLP for dense prediction while maintaining the original input size (e.g., predict a 0 or 1 value for each pixel in an input image)? In addition, I am interested in only a single ("annotated") output image, although I noticed the model definitions given in this repository output multiple downsampled versions of the original input image. Any thoughts on this?

    Thank you in advance for your time.

    opened by amorehead 2
  • Swin-B vs CycleMLP-B on image classification

    Swin-B vs CycleMLP-B on image classification

    For classificaion on ImageNet-1k, the acuracy of Swin-B is 83.5, which is 0.1 higher than the proposed CycleMLP-B. But, in this paper, the authors reprot that the accuracy of Swin-B is 83.3, which is 0.1 lower than the proposed CycleMLP-B. Why are these accuracies different?

    opened by hkzhang91 1
  • question about the offset

    question about the offset

    Thanks for your work!

    The implementation of this code inspired me. But the calculation of offset here is confusing. Although this issue (https://github.com/ShoufaChen/CycleMLP/issues/10) has asked similar questions, I haven't found a reasonable explanation.

    https://github.com/ShoufaChen/CycleMLP/blob/2f76a1f6e3cc6672143fdac46e3db5f9a7341253/cycle_mlp.py#L127-L136

    kernel_size = (1, 3)
    start_idx = (kernel_size[0] * kernel_size[1]) // 2
    for i in range(num_channels):
        offset[0, 2 * i + 0, 0, 0] = 0
        # relative offset
        offset[0, 2 * i + 1, 0, 0] = (i + start_idx) % kernel_size[1] - (kernel_size[1] // 2)
    offset.reshape(num_channels, 2)
    
    tensor([[ 0.,  0.],
            [ 0.,  1.],
            [ 0., -1.],
            [ 0.,  0.],
            [ 0.,  1.],
            [ 0., -1.]])
    

    the results are different with the figure in paper:

    image

    Some codes for verification:

    import torch
    from torchvision.ops import deform_conv2d
    
    num_channels = 6
    
    data = torch.arange(1, 6).reshape(1, 1, 1, 5).expand(-1, num_channels, -1, -1)
    data
    """
    tensor([[[[1, 2, 3, 4, 5]],
             [[1, 2, 3, 4, 5]],
             [[1, 2, 3, 4, 5]],
             [[1, 2, 3, 4, 5]],
             [[1, 2, 3, 4, 5]],
             [[1, 2, 3, 4, 5]]]])
    """
    
    weight = torch.eye(num_channels).reshape(num_channels, num_channels, 1, 1)
    weight.reshape(num_channels, num_channels)
    """
    tensor([[1., 0., 0., 0., 0., 0.],
            [0., 1., 0., 0., 0., 0.],
            [0., 0., 1., 0., 0., 0.],
            [0., 0., 0., 1., 0., 0.],
            [0., 0., 0., 0., 1., 0.],
            [0., 0., 0., 0., 0., 1.]])
    """
    
    offset = torch.empty(1, 2 * num_channels * 1 * 1, 1, 1)
    kernel_size = (1, 3)
    start_idx = (kernel_size[0] * kernel_size[1]) // 2
    for i in range(num_channels):
        offset[0, 2 * i + 0, 0, 0] = 0
        # relative offset
        offset[0, 2 * i + 1, 0, 0] = (
            (i + start_idx) % kernel_size[1] - (kernel_size[1] // 2)
        )
    offset.reshape(num_channels, 2)
    """
    tensor([[ 0.,  0.],
            [ 0.,  1.],
            [ 0., -1.],
            [ 0.,  0.],
            [ 0.,  1.],
            [ 0., -1.]])
    """
    
    deform_conv2d(
        data.float(), 
        offset=offset.expand(-1, -1, -1, 5).float(), 
        weight=weight.float(), 
        bias=None,
    )
    """
    tensor([[[[1., 2., 3., 4., 5.]],
             [[2., 3., 4., 5., 0.]],
             [[0., 1., 2., 3., 4.]],
             [[1., 2., 3., 4., 5.]],
             [[2., 3., 4., 5., 0.]],
             [[0., 1., 2., 3., 4.]]]])
    """
    
    opened by lartpang 1
  • question about the offset

    question about the offset

    Hi, thank you very much for your excellent work. In Fig.4 of your paper, you show the pseudo-kernel when kernel size is 1x3. But I when I find that function "gen_offset" does not generate the same offset as Fig.4. The offset it generates is "0,1,0,-1,0,0,0,1..." instead of "0,1,0,-1,0,1,0,-1', which is shown in Fig.4. So could you please tell me the reason? image image

    opened by linjing7 1
  • About

    About "crop_pct"

    Hi, thanks for your great work and code. I wonder the parameter crop_pct actually works in which part of code. When I go throught the timm, I can't find out how this crop_pct is loaded.

    opened by ggjy 1
  • How to deploy CycleMLP-T for training?

    How to deploy CycleMLP-T for training?

    Thank you very much for such a wonderful work!

    After learning the cycle_mlp source code in the repository, I am very confused to deploy CycleMLP Block based on Swin Transformer. Is it convenient for you to release swin-based CycleMLP? Looking forward to your reply, Thanks!

    opened by Pak287 0
Owner
Shoufa Chen
Shoufa Chen
Real-time face detection and emotion/gender classification using fer2013/imdb datasets with a keras CNN model and openCV.

Real-time face detection and emotion/gender classification using fer2013/imdb datasets with a keras CNN model and openCV.

Octavio Arriaga 5.3k Dec 30, 2022
make ASCII Art by Deep Learning

DeepAA This is convolutional neural networks generating ASCII art. This repository is under construction. This work is accepted by NIPS 2017 Workshop,

OsciiArt 1.4k Dec 28, 2022
Using Clinical Drug Representations for Improving Mortality and Length of Stay Predictions

Using Clinical Drug Representations for Improving Mortality and Length of Stay Predictions Usage Clone the code to local. https://github.com/tanlab/MI

Computational Biology and Machine Learning lab @ TOBB ETU 3 Oct 18, 2022
Vignette is a face tracking software for characters using osu!framework.

Vignette is a face tracking software for characters using osu!framework. Unlike most solutions, Vignette is: Made with osu!framework, the game framewo

Vignette 412 Dec 28, 2022
A general 3D Object Detection codebase in PyTorch.

Det3D is the first 3D Object Detection toolbox which provides off the box implementations of many 3D object detection algorithms such as PointPillars, SECOND, PIXOR, etc, as well as state-of-the-art

Benjin Zhu 1.4k Jan 05, 2023
Cossim - Sharpened Cosine Distance implementation in PyTorch

Sharpened Cosine Distance PyTorch implementation of the Sharpened Cosine Distanc

Istvan Fehervari 10 Mar 22, 2022
A Pytorch implementation of "Splitter: Learning Node Representations that Capture Multiple Social Contexts" (WWW 2019).

Splitter ⠀⠀ A PyTorch implementation of Splitter: Learning Node Representations that Capture Multiple Social Contexts (WWW 2019). Abstract Recent inte

Benedek Rozemberczki 201 Nov 09, 2022
Analyses of the individual electric field magnitudes with Roast.

Aloi Davide - PhD Student (UoB) Analysis of electric field magnitudes (wp2a dataset only at the moment) and correlation analysis with Dynamic Causal M

Davide Aloi 7 Dec 15, 2022
Video Frame Interpolation with Transformer (CVPR2022)

VFIformer Official PyTorch implementation of our CVPR2022 paper Video Frame Interpolation with Transformer Dependencies python = 3.8 pytorch = 1.8.0

DV Lab 63 Dec 16, 2022
🧠 A PyTorch implementation of 'Deep CORAL: Correlation Alignment for Deep Domain Adaptation.', ECCV 2016

Deep CORAL A PyTorch implementation of 'Deep CORAL: Correlation Alignment for Deep Domain Adaptation. B Sun, K Saenko, ECCV 2016' Deep CORAL can learn

Andy Hsu 200 Dec 25, 2022
Convert Mission Planner (ArduCopter) Waypoint Missions to Litchi CSV Format to execute on DJI Drones

Mission Planner to Litchi Convert Mission Planner (ArduCopter) Waypoint Surveys to Litchi CSV Format to execute on DJI Drones Litchi doesn't support S

Yaros 24 Dec 09, 2022
Official repository with code and data accompanying the NAACL 2021 paper "Hurdles to Progress in Long-form Question Answering" (https://arxiv.org/abs/2103.06332).

Hurdles to Progress in Long-form Question Answering This repository contains the official scripts and datasets accompanying our NAACL 2021 paper, "Hur

Kalpesh Krishna 41 Nov 08, 2022
Contrastive Learning for Compact Single Image Dehazing, CVPR2021

AECR-Net Contrastive Learning for Compact Single Image Dehazing, CVPR2021. Official Pytorch based implementation. Paper arxiv Pytorch Version TODO: mo

glassy 253 Jan 01, 2023
Open-source implementation of Google Vizier for hyper parameters tuning

Advisor Introduction Advisor is the hyper parameters tuning system for black box optimization. It is the open-source implementation of Google Vizier w

tobe 1.5k Jan 04, 2023
UltraGCN: An Ultra Simplification of Graph Convolutional Networks for Recommendation

UltraGCN This is our Pytorch implementation for our CIKM 2021 paper: Kelong Mao, Jieming Zhu, Xi Xiao, Biao Lu, Zhaowei Wang, Xiuqiang He. UltraGCN: A

XUEPAI 93 Jan 03, 2023
Large-Scale Pre-training for Person Re-identification with Noisy Labels (LUPerson-NL)

LUPerson-NL Large-Scale Pre-training for Person Re-identification with Noisy Labels (LUPerson-NL) The repository is for our CVPR2022 paper Large-Scale

43 Dec 26, 2022
FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection

FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection arXi

59 Nov 29, 2022
EMNLP 2021 paper The Devil is in the Detail: Simple Tricks Improve Systematic Generalization of Transformers.

Codebase for training transformers on systematic generalization datasets. The official repository for our EMNLP 2021 paper The Devil is in the Detail:

Csordás Róbert 57 Nov 21, 2022
Offcial implementation of "A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction, ICCV-2021".

HF2-VAD Offcial implementation of "A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Predictio

76 Dec 21, 2022
Specification language for generating Generalized Linear Models (with or without mixed effects) from conceptual models

tisane Tisane: Authoring Statistical Models via Formal Reasoning from Conceptual and Data Relationships TL;DR: Analysts can use Tisane to author gener

Eunice Jun 11 Nov 15, 2022