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
Dynamica causal Bayesian optimisation

Dynamic Causal Bayesian Optimization This is a Python implementation of Dynamic Causal Bayesian Optimization as presented at NeurIPS 2021. Abstract Th

nd308 18 Nov 22, 2022
MVFNet: Multi-View Fusion Network for Efficient Video Recognition (AAAI 2021)

MVFNet: Multi-View Fusion Network for Efficient Video Recognition (AAAI 2021) Overview We release the code of the MVFNet (Multi-View Fusion Network).

2 Jan 29, 2022
Fast image augmentation library and an easy-to-use wrapper around other libraries

Albumentations Albumentations is a Python library for image augmentation. Image augmentation is used in deep learning and computer vision tasks to inc

11.4k Jan 09, 2023
Iterative Training: Finding Binary Weight Deep Neural Networks with Layer Binarization

Iterative Training: Finding Binary Weight Deep Neural Networks with Layer Binarization This repository contains the source code for the paper (link wi

Rakuten Group, Inc. 0 Nov 19, 2021
PoseViz – Multi-person, multi-camera 3D human pose visualization tool built using Mayavi.

PoseViz – 3D Human Pose Visualizer Multi-person, multi-camera 3D human pose visualization tool built using Mayavi. As used in MeTRAbs visualizations.

István Sárándi 79 Dec 30, 2022
Elastic weight consolidation technique for incremental learning.

Overcoming-Catastrophic-forgetting-in-Neural-Networks Elastic weight consolidation technique for incremental learning. About Use this API if you dont

Shivam Saboo 89 Dec 22, 2022
Official Repo for ICCV2021 Paper: Learning to Regress Bodies from Images using Differentiable Semantic Rendering

[ICCV2021] Learning to Regress Bodies from Images using Differentiable Semantic Rendering Getting Started DSR has been implemented and tested on Ubunt

Sai Kumar Dwivedi 83 Nov 27, 2022
I-BERT: Integer-only BERT Quantization

I-BERT: Integer-only BERT Quantization HuggingFace Implementation I-BERT is also available in the master branch of HuggingFace! Visit the following li

Sehoon Kim 139 Dec 27, 2022
Material for my PyConDE & PyData Berlin 2022 Talk "5 Steps to Speed Up Your Data-Analysis on a Single Core"

5 Steps to Speed Up Your Data-Analysis on a Single Core Material for my talk at the PyConDE & PyData Berlin 2022 Description Your data analysis pipeli

Jonathan Striebel 9 Dec 12, 2022
Useful materials and tutorials for 110-1 NTU DBME5028 (Application of Deep Learning in Medical Imaging)

Useful materials and tutorials for 110-1 NTU DBME5028 (Application of Deep Learning in Medical Imaging)

7 Jun 22, 2022
PyTorch implementation of "Efficient Neural Architecture Search via Parameters Sharing"

Efficient Neural Architecture Search (ENAS) in PyTorch PyTorch implementation of Efficient Neural Architecture Search via Parameters Sharing. ENAS red

Taehoon Kim 2.6k Dec 31, 2022
EMNLP 2021: Single-dataset Experts for Multi-dataset Question-Answering

MADE (Multi-Adapter Dataset Experts) This repository contains the implementation of MADE (Multi-adapter dataset experts), which is described in the pa

Princeton Natural Language Processing 68 Jul 18, 2022
Event-forecasting - Event Forecasting Algorithms With Python

event-forecasting Event Forecasting Algorithms Theory Correlating events in comp

Intellia ICT 4 Feb 15, 2022
deep-table implements various state-of-the-art deep learning and self-supervised learning algorithms for tabular data using PyTorch.

deep-table implements various state-of-the-art deep learning and self-supervised learning algorithms for tabular data using PyTorch.

63 Oct 17, 2022
Python Environment for Bayesian Learning

Pebl is a python library and command line application for learning the structure of a Bayesian network given prior knowledge and observations. Pebl in

Abhik Shah 103 Jul 14, 2022
Pytorch implemenation of Stochastic Multi-Label Image-to-image Translation (SMIT)

SMIT: Stochastic Multi-Label Image-to-image Translation This repository provides a PyTorch implementation of SMIT. SMIT can stochastically translate a

Biomedical Computer Vision Group @ Uniandes 37 Mar 01, 2022
A Dataset for Direct Quotation Extraction and Attribution in News Articles.

DirectQuote - A Dataset for Direct Quotation Extraction and Attribution in News Articles DirectQuote is a corpus containing 19,760 paragraphs and 10,3

THUNLP-MT 9 Sep 23, 2022
Personals scripts using ageitgey/face_recognition

HOW TO USE pip3 install requirements.txt Add some pictures of known people in the folder 'people' : a) Create a folder called by the name of the perso

Antoine Bollengier 1 Jan 06, 2022
UniFormer - official implementation of UniFormer

UniFormer This repo is the official implementation of "Uniformer: Unified Transf

SenseTime X-Lab 573 Jan 04, 2023
Simple tutorials using Google's TensorFlow Framework

TensorFlow-Tutorials Introduction to deep learning based on Google's TensorFlow framework. These tutorials are direct ports of Newmu's Theano Tutorial

Nathan Lintz 6k Jan 06, 2023