This repository contains the code for the paper Decoupled Weight Decay Regularization (old title: Fixing Weight Decay Regularization in Adam) by Ilya Loshchilov and Frank Hutter, ICLR 2019 arXiv.

The code represents a tiny modification of the source code provided for the Shake-Shake regularization by Xavier Gastaldi arXiv. Since the usage of both is very similar, the introduction and description of the original Shake-Shake code is given below. Please consider to first run the Shake-Shake code and then our code.

Find below a few examples to train a 26 2x96d "Shake-Shake-Image" ResNet on CIFAR-10 with 1 GPU. To run on 4 GPUs, set CUDA_VISIBLE_DEVICES=0,1,2,3 and -nGPU 4. For test purposes you may reduce -nEpochs from 1500 to e.g. 150 and set -widenFactor to 4 to use a smaller network. To run on ImageNet32x32, set -dataset to imagenet32 and reduce -nEpochs to 150. You may consider to use -weightDecay=0.05 for CIFAR-10.

Importantly, please copy with replacement adam.lua and sgd.lua from UPDATETORCHFILES to YOURTORCHFOLDER/install/share/lua/5.1/optim/

To run AdamW for nEpochs=1500 epochs without restarts with initial learning rate LR=0.001, normalized weight decay weightDecay=0.025

CUDA_VISIBLE_DEVICES=0 th main.lua -algorithmType ADAMW -nEpochs 1500 -Te 1500 -Tmult 2 -widenFactor 6 -LR 0.001 -weightDecay 0.025 -dataset cifar10 -nGPU 1 -depth 26 -irun 1 -batchSize 128 -momentum 0.9 -shareGradInput false -optnet true -netType shakeshake -forwardShake true -backwardShake true -shakeImage true -lrShape cosine -LRdec true

To run AdamW for nEpochs=1500 epochs with restarts, where the first restart will happen after Te=100 epochs and the second restart after 200 more epochs because 100*Tmult=200.

CUDA_VISIBLE_DEVICES=0 th main.lua -algorithmType ADAMW -nEpochs 1500 -Te 100 -Tmult 2 -widenFactor 6 -LR 0.001 -weightDecay 0.025 -dataset cifar10 -nGPU 1 -depth 26 -irun 1 -batchSize 128 -momentum 0.9 -shareGradInput false -optnet true -netType shakeshake -forwardShake true -backwardShake true -shakeImage true -lrShape cosine -LRdec true

To run SGDW for nEpochs=150 epochs without restarts with initial learning rate LR=0.05, normalized weight decay weightDecay=0.025

CUDA_VISIBLE_DEVICES=0 th main.lua -algorithmType SGDW -nEpochs 1500 -Te 1500 -Tmult 2 -widenFactor 6 -LR 0.05 -weightDecay 0.025 -dataset cifar10 -nGPU 1 -depth 26 -irun 1 -batchSize 128 -momentum 0.9 -shareGradInput false -optnet true -netType shakeshake -forwardShake true -backwardShake true -shakeImage true -lrShape cosine -LRdec true

To run SGDW for nEpochs=150 epochs with restarts, where the first restart will happen after Te=100 epochs and the second restart after 200 more epochs because 100*Tmult=200.

CUDA_VISIBLE_DEVICES=0 th main.lua -algorithmType SGDW -nEpochs 1500 -Te 100 -Tmult 2 -widenFactor 6 -LR 0.001 -weightDecay 0.025 -dataset cifar10 -nGPU 1 -depth 26 -irun 1 -batchSize 128 -momentum 0.9 -shareGradInput false -optnet true -netType shakeshake -forwardShake true -backwardShake true -shakeImage true -lrShape cosine -LRdec true

Acknowledgments: We thank Patryk Chrabaszcz for creating functions dealing with ImageNet32x32 dataset.

This repository contains the code for the paper Shake-Shake regularization of 3-branch residual networks.

The code is based on [fb.resnet.torch] (https://github.com/facebook/fb.resnet.torch).

0. Introduction
1. Results
2. Usage
3. Contact

This method aims at helping computer vision practitioners faced with an overfit problem. The idea is to replace, in a 3-branch ResNet, the standard summation of residual branches by a stochastic affine combination. The largest tested model improves on the best single shot published result on CIFAR-10 by reaching 2.72% test error.

Figure 1: Left: Forward training pass. Center: Backward training pass. Right: At test time.

The base network is a 26 2x32d ResNet (i.e. the network has a depth of 26, 2 residual branches and the first residual block has a width of 32). "Shake" means that all scaling coefficients are overwritten with new random numbers before the pass. "Even" means that all scaling coefficients are set to 0.5 before the pass. "Keep" means that we keep, for the backward pass, the scaling coefficients used during the forward pass. "Batch" means that, for each residual block, we apply the same scaling coefficient for all the images in the mini-batch. "Image" means that, for each residual block, we apply a different scaling coefficient for each image in the mini-batch. The numbers in the Table below represent the average of 3 runs except for the 96d models which were run 5 times.

Table 1: Error rates (%) on CIFAR-10

0. Install [fb.resnet.torch] (https://github.com/facebook/fb.resnet.torch), optnet and lua-stdlib.
1. Download Shake-Shake

git clone https://github.com/xgastaldi/shake-shake.git

2. Copy the elements in the shake-shake folder and paste them in the fb.resnet.torch folder. This will overwrite 5 files (main.lua, train.lua, opts.lua, checkpoints.lua and models/init.lua) and add 3 new files (models/shakeshake.lua, models/shakeshakeblock.lua and models/mulconstantslices.lua).
3. You can train a 26 2x32d "Shake-Shake-Image" ResNet on CIFAR-10+ using

th main.lua -dataset cifar10 -nGPU 1 -batchSize 128 -depth 26 -shareGradInput false -optnet true -nEpochs 1800 -netType shakeshake -lrShape cosine -widenFactor 2 -LR 0.2 -forwardShake true -backwardShake true -shakeImage true

You can train a 26 2x96d "Shake-Shake-Image" ResNet on 2 GPUs using

CUDA_VISIBLE_DEVICES=0,1 th main.lua -dataset cifar10 -nGPU 2 -batchSize 128 -depth 26 -shareGradInput false -optnet true -nEpochs 1800 -netType shakeshake -lrShape cosine -widenFactor 6 -LR 0.2 -forwardShake true -backwardShake true -shakeImage true

A widenFactor of 2 corresponds to 32d, 4 to 64d, etc..

Changes made to fb.resnet.torch files:

main.lua
Ln 17, 54-59, 81-88: Adds a log (courtesy of Sergey Zagoruyko)

train.lua
Ln 36-38 58-60 206-213: Adds the cosine learning rate function
Ln 88-89: Adds the learning rate to the elements printed on screen

opts.lua
Ln 57-62: Adds Shake-Shake options

checkpoints.lua
Ln 15-16: Adds require 'models/shakeshakeblock' and require 'std'
Ln 60-61: Avoids using the fb.renet.torch deepcopy (it doesn't seem to be compatible with the BN in shakeshakeblock) and replaces it with the deepcopy from stdlib
Ln 67-81: Saves only the best model

init.lua
Ln 91-92: Adds require 'models/mulconstantslices' and require 'models/shakeshakeblock'

The main model is in shakeshake.lua. The residual block model is in shakeshakeblock.lua. mulconstantslices.lua is just an extension of nn.mulconstant that multiplies elements of a vector with image slices of a mini-batch tensor.

xgastaldi.mba2011 at london.edu
Any discussions, suggestions and questions are welcome!