Dynamic Slimmable Network (CVPR 2021, Oral)

Last update: Dec 09, 2022

Overview

Dynamic Slimmable Network (DS-Net)

This repository contains PyTorch code of our paper: Dynamic Slimmable Network (CVPR 2021 Oral).

Architecture of DS-Net. The width of each supernet stage is adjusted adaptively by the slimming ratio ρ predicted by the gate.

Accuracy vs. complexity on ImageNet.

Usage

1. Requirements

Install PyTorch 1.2.0+, for example:

conda install -c pytorch pytorch torchvision

Install pytorch-image-models 0.3.2, for example:
```
pip install timm==0.3.2
```
Download ImageNet from http://image-net.org/. Move validation images to labeled subfolders using following script: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.shvalprep.sh

2. Stage I: Supernet Training

For example, train dynamic slimmable MobileNet supernet with 8 GPUs (takes about 2 days):

python -m torch.distributed.launch --nproc_per_node=8 train.py /PATH/TO/ImageNet -c ./configs/mobilenetv1_bn_uniform.yml

3. Stage II: Gate Training

Will be available soon

Citation

If you use our code for your paper, please cite:

@inproceedings{li2021dynamic,
  author = {Changlin Li and
            Guangrun Wang and
            Bing Wang and
            Xiaodan Liang and
            Zhihui Li and
            Xiaojun Chang},
  title = {Dynamic Slimmable Network},
  booktitle = {CVPR},
  year = {2021}
}

Comments

The usage of gumbel softmax in DS-Net

Thank you for your very nice work,I want to know that the effect of gumble softmax，because I think the network can be trained without gumble softmax. Is the gumbel softmax just aimed to increase the randomness of channel choice?
discussion

opened by LinyeLi60 7
UserWarning: Argument interpolation should be of type InterpolationMode instead of int. Please, use InterpolationMode enum.

Why I get an warning: /home/chauncey/.local/lib/python3.8/site-packages/torchvision/transforms/functional.py:364: UserWarning: Argument interpolation should be of type InterpolationMode instead of int. Please, use InterpolationMode enum. warnings.warn( when I use python3 -m torch.distributed.launch --nproc_per_node=1 train.py ./imagenet -c ./configs/mobilenetv1_bn_uniform.yml

opened by Chauncey-Wang 3
Question about calculating MAdds of dynamic network in the paper

Thank you for your great work, and I have a question about how to calculate MAdds in your paper. The dynamic network has different widths and MAdds for each instance, but you denoted MAdds for your networks. Are they the average MAdds for the whole dataset?
discussion

opened by sseung0703 3
why not set ensemble_ib to True?

Hi,

I found that ensemble_ib is set to False for both slim training and gate training from the configs, but from paper it would boost the performance when set toTrue.

Any idea?

opened by twmht 2
MAdds of Pretrained Supernet

Hi Changlin, your work is excellent. I have a question about the calculation of MAdds, in README.md the MAdds of Subnetwork 13 is 565M, but I think the MAdds of Subnetwork 13 should be 821M observed in my experiments, because the channel number of Subnetwork 13 is larger than the original MobileNetV1, and the original MobileNetV1 1.0's MAdds should be 565M. Looking forward to your reply.

opened by LinyeLi60 2
Error of change the num_choice in mobilenetv1_bn_uniform_reset_bn.yml

I follow your suggestion to set the num_choice in mobilenetv1_bn_uniform_reset_bn.yml to 14, but get an expected error when I use python -m torch.distributed.launch --nproc_per_node=8 train.py /PATH/TO/ImageNet -c ./configs/mobilenetv1_bn_uniform_reset_bn.yml.

08/25 10:15:57 AM Recalibrating BatchNorm statistics... 08/25 10:16:10 AM Finish recalibrating BatchNorm statistics. 08/25 10:16:19 AM Finish recalibrating BatchNorm statistics. 08/25 10:16:21 AM Test: [ 0/0] Mode: 0 Time: 0.344 (0.344) Loss: 6.9204 (6.9204) [email protected]: 0.0000 ( 0.0000) [email protected]: 0.0000 ( 0.0000) Flops: 132890408 (132890408) 08/25 10:16:22 AM Test: [ 0/0] Mode: 1 Time: 0.406 (0.406) Loss: 6.9189 (6.9189) [email protected]: 0.0000 ( 0.0000) [email protected]: 0.0000 ( 0.0000) Flops: 152917440 (152917440) 08/25 10:16:22 AM Test: [ 0/0] Mode: 2 Time: 0.381 (0.381) Loss: 6.9187 (6.9187) [email protected]: 0.0000 ( 0.0000) [email protected]: 0.0000 ( 0.0000) Flops: 175152224 (175152224) 08/25 10:16:23 AM Test: [ 0/0] Mode: 3 Time: 0.389 (0.389) Loss: 6.9134 (6.9134) [email protected]: 0.0000 ( 0.0000) [email protected]: 0.0000 ( 0.0000) Flops: 199594752 (199594752) Traceback (most recent call last): File "train.py", line 658, in main() File "train.py", line 635, in main eval_metrics.append(validate_slim(model, File "/home/chauncey/PycharmProjects/DS-Net-main/dyn_slim/apis/train_slim.py", line 215, in validate_slim output = model(input) File "/home/chauncey/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl result = self.forward(*input, **kwargs) File "/home/chauncey/PycharmProjects/DS-Net-main/dyn_slim/models/dyn_slim_net.py", line 191, in forward x = self.forward_features(x) File "/home/chauncey/PycharmProjects/DS-Net-main/dyn_slim/models/dyn_slim_net.py", line 178, in forward_features x = stage(x) File "/home/chauncey/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl result = self.forward(*input, **kwargs) File "/home/chauncey/PycharmProjects/DS-Net-main/dyn_slim/models/dyn_slim_stages.py", line 48, in forward x = self.first_block(x) File "/home/chauncey/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl result = self.forward(*input, **kwargs) File "/home/chauncey/PycharmProjects/DS-Net-main/dyn_slim/models/dyn_slim_blocks.py", line 240, in forward x = self.conv_pw(x) File "/home/chauncey/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl result = self.forward(*input, **kwargs) File "/home/chauncey/PycharmProjects/DS-Net-main/dyn_slim/models/dyn_slim_ops.py", line 94, in forward self.running_outc = self.out_channels_list[self.channel_choice] IndexError: list index out of range

It looks like we should make some adjustment in other py files.

opened by chaunceywx 2
Why the num_choice in different yml is different?

Why you set num_choice in mobilenetv1_bn_uniform_reset_bn.yml as 4, but set this parameter as 14 in the other two yml file?

老哥，如果你也是中国人，咱们还是用中文交流吧，我英语水平比较感人。。。

opened by chaunceywx 2
运行问题

请问大佬下面这个问题是为什么 Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

/root/anaconda3/envs/0108/lib/python3.6/site-packages/torchvision/io/image.py:11: UserWarning: Failed to load image Python extension: /root/anaconda3/envs/0108/lib/python3.6/site-packages/torchvision/image.so: undefined symbol: _ZNK3c106IValue23reportToTensorTypeErrorEv warn(f"Failed to load image Python extension: {e}") /root/anaconda3/envs/0108/lib/python3.6/site-packages/torchvision/io/image.py:11: UserWarning: Failed to load image Python extension: /root/anaconda3/envs/0108/lib/python3.6/site-packages/torchvision/image.so: undefined symbol: _ZNK3c106IValue23reportToTensorTypeErrorEv warn(f"Failed to load image Python extension: {e}") 01/21 05:42:18 AM Added key: store_based_barrier_key:1 to store for rank: 1 01/21 05:42:18 AM Added key: store_based_barrier_key:1 to store for rank: 0 01/21 05:42:18 AM Training in distributed mode with multiple processes, 1 GPU per process. Process 0, total 2. 01/21 05:42:18 AM Training in distributed mode with multiple processes, 1 GPU per process. Process 1, total 2. 01/21 05:42:20 AM Model slimmable_mbnet_v1_bn_uniform created, param count: 7676204 01/21 05:42:20 AM Data processing configuration for current model + dataset: 01/21 05:42:20 AM input_size: (3, 224, 224) 01/21 05:42:20 AM interpolation: bicubic 01/21 05:42:20 AM mean: (0.485, 0.456, 0.406) 01/21 05:42:20 AM std: (0.229, 0.224, 0.225) 01/21 05:42:20 AM crop_pct: 0.875 01/21 05:42:20 AM NVIDIA APEX not installed. AMP off. 01/21 05:42:21 AM Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP. 01/21 05:42:21 AM Scheduled epochs: 40 01/21 05:42:21 AM Training folder does not exist at: images/train 01/21 05:42:21 AM Training folder does not exist at: images/train Killing subprocess 239 Killing subprocess 240 Traceback (most recent call last): File "/root/anaconda3/envs/0108/lib/python3.6/runpy.py", line 193, in _run_module_as_main "main", mod_spec) File "/root/anaconda3/envs/0108/lib/python3.6/runpy.py", line 85, in _run_code exec(code, run_globals) File "/root/anaconda3/envs/0108/lib/python3.6/site-packages/torch/distributed/launch.py", line 340, in main() File "/root/anaconda3/envs/0108/lib/python3.6/site-packages/torch/distributed/launch.py", line 326, in main sigkill_handler(signal.SIGTERM, None) # not coming back File "/root/anaconda3/envs/0108/lib/python3.6/site-packages/torch/distributed/launch.py", line 301, in sigkill_handler raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd) subprocess.CalledProcessError: Command '['/root/anaconda3/envs/0108/bin/python', '-u', 'train.py', '--local_rank=1', 'images', '-c', './configs/mobilenetv1_bn_uniform_reset_bn.yml']' returned non-zero exit status 1.

opened by 6imust 1
project environment

Hi,could you provide the environment for the project?I try to train the network with python=3.8 pytorch=1.7.1,cuda=10.2.Shortly after starting training,there's a RuntimeError: CUDA error: device-side assert triggered happened,and some other environment also lead to this error.I'm not sure whether the problem is caused by the difference of environment.

opened by singularity97 1
Softmax twice for SGS loss？

Dear authors, thanks for this nice work.

I wonder why the calculation of the SGS loss is using the softmaxed data rather than the logits, considering the PyTorch CrossEntropyLoss already contains a softmax inside.

https://github.com/changlin31/DS-Net/blob/15cd3036970ec27d2c306014344fd50d9e9b888b/dyn_slim/apis/train_slim_gate.py#L98 https://github.com/changlin31/DS-Net/blob/15cd3036970ec27d2c306014344fd50d9e9b888b/dyn_slim/models/dyn_slim_blocks.py#L324-L355

opened by Yu-Zhewen 0
Can we futher improve autoalim without gate?

It is not easy to deploy gate operator with some other backends, like TensorRT.

So my question is can we futher improve autoalim without the dynamic gate when inference?Any ongoing work are doing this?

opened by twmht 3
DS-Net for object detection

Hello. Thanks for your work. I noticed that you also conducted some experiments in object detection. I wonder whether or when you will release the code

opened by NoLookDefense 8
Dynamic path for DS-mobilenet

Hi. Thanks for your work. I am reading your paper and trying to reimplement, and I feel confused about some details. You mentioned in your paper that the slimming ratio ρ∈[0.35 : 0.05 : 1.25], which have 18 paths. However, in your code, there are only 14 paths ρ∈[0.35 : 0.05 : 1] as mentioned in https://github.com/changlin31/DS-Net/blob/15cd3036970ec27d2c306014344fd50d9e9b888b/dyn_slim/models/dyn_slim_net.py#L36 . And also, when conducting gate training, the gate function only has a 4-dimension output, meaning that there is only 4 paths and the slimming ratio is restricted to ρ∈[0.35 : 0.05 : 0.5]. https://github.com/changlin31/DS-Net/blob/15cd3036970ec27d2c306014344fd50d9e9b888b/dyn_slim/models/dyn_slim_blocks.py#L204 Why the dynamic path for larger network is not used?

opened by NoLookDefense 1

Releases(v0.0.1)

v0.0.1(Nov 30, 2021)

Pretrained weights of DS-MBNet supernet. Detailed accuracy of each sub-networks:

| Subnetwork | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | | ----------------- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | MAdds | 133M | 153M | 175M | 200M | 226M | 255M | 286M | 319M | 355M | 393M | 433M | 475M | 519M | 565M | | Top-1 (%) | 70.1 | 70.4 | 70.8 | 71.2 | 71.6 | 72.0 | 72.4 | 72.7 | 73.0 | 73.3 | 73.6 | 73.9 | 74.1 | 74.6 | | Top-5 (%) | 89.4 | 89.6 | 89.9 | 90.2 | 90.3 | 90.6 | 90.9 | 91.0 | 91.2 | 91.4 | 91.5 | 91.7 | 91.8 | 92.0 |
Source code(tar.gz)
Source code(zip)
DS_MBNet-70_1.pth.tar(60.93 MB)
log-DS_MBNet-70_1.txt(6.12 KB)

Owner

Changlin Li

GitHub Repository

Hl classification bc - A Network-Based High-Level Data Classification Algorithm Using Betweenness Centrality

A Network-Based High-Level Data Classification Algorithm Using Betweenness Centr

3 Dec 01, 2022

Bayesian Meta-Learning Through Variational Gaussian Processes

vmgp This is the repository of Vivek Myers and Nikhil Sardana for our CS 330 final project, Bayesian Meta-Learning Through Variational Gaussian Proces

2 Nov 17, 2022

This is a computer vision based implementation of the popular childhood game 'Hand Cricket/Odd or Even' in python

Hand Cricket Table of Content Overview Installation Game rules Project Details Future scope Overview This is a computer vision based implementation of

6 Jan 12, 2022

PyTorch implementation of the Transformer in Post-LN (Post-LayerNorm) and Pre-LN (Pre-LayerNorm).

Transformer-PyTorch A PyTorch implementation of the Transformer from the paper Attention is All You Need in both Post-LN (Post-LayerNorm) and Pre-LN (

22 Feb 27, 2022

PyTorch-LIT is the Lite Inference Toolkit (LIT) for PyTorch which focuses on easy and fast inference of large models on end-devices.

PyTorch-LIT PyTorch-LIT is the Lite Inference Toolkit (LIT) for PyTorch which focuses on easy and fast inference of large models on end-devices. With

157 Dec 11, 2022

Machine learning framework for both deep learning and traditional algorithms

NeoML is an end-to-end machine learning framework that allows you to build, train, and deploy ML models. This framework is used by ABBYY engineers for

704 Dec 27, 2022

Code for reproducing our analysis in the paper titled: Image Cropping on Twitter: Fairness Metrics, their Limitations, and the Importance of Representation, Design, and Agency

Image Crop Analysis This is a repo for the code used for reproducing our Image Crop Analysis paper as shared on our blog post. If you plan to use this

239 Jan 02, 2023

An image processing project uses Viola-jones technique to detect faces and then use SIFT algorithm for recognition.

Attendance_System An image processing project uses Viola-jones technique to detect faces and then use LPB algorithm for recognition. Face Detection Us

8 Jan 11, 2022

Official code of our work, AVATAR: A Parallel Corpus for Java-Python Program Translation.

AVATAR Official code of our work, AVATAR: A Parallel Corpus for Java-Python Program Translation. AVATAR stands for jAVA-pyThon progrAm tRanslation. AV

26 Dec 03, 2022

A benchmark for the task of translation suggestion

WeTS: A Benchmark for Translation Suggestion Translation Suggestion (TS), which provides alternatives for specific words or phrases given the entire d

55 Dec 24, 2022

This is an official implementation of our CVPR 2021 paper "Bottom-Up Human Pose Estimation Via Disentangled Keypoint Regression" (https://arxiv.org/abs/2104.02300)

Bottom-Up Human Pose Estimation Via Disentangled Keypoint Regression Introduction In this paper, we are interested in the bottom-up paradigm of estima

367 Dec 27, 2022

An onlinel learning to rank python codebase.

OLTR Online learning to rank python codebase. The code related to Pairwise Differentiable Gradient Descent (ranker/PDGDLinearRanker.py) is copied from

5 Jul 18, 2022

PyTorch Lightning + Hydra. A feature-rich template for rapid, scalable and reproducible ML experimentation with best practices. ⚡🔥⚡

Lightning-Hydra-Template A clean and scalable template to kickstart your deep learning project 🚀 ⚡ 🔥 Click on Use this template to initialize new re

2.1k Jan 09, 2023