In-Place Activated BatchNorm

In-Place Activated BatchNorm for Memory-Optimized Training of DNNs

In-Place Activated BatchNorm (InPlace-ABN) is a novel approach to reduce the memory required for training deep networks. It allows for up to 50% memory savings in modern architectures such as ResNet, ResNeXt and Wider ResNet by redefining BN + non linear activation as a single in-place operation, while smartly dropping or recomputing intermediate buffers as needed.

This repository contains a PyTorch implementation of the InPlace-ABN layer, as well as some training scripts to reproduce the ImageNet classification results reported in our paper.

• Overview
• Installation
• Training on ImageNet

If you use In-Place Activated BatchNorm in your research, please cite:

@article{rotabulo2017place,
  title={In-Place Activated BatchNorm for Memory-Optimized Training of DNNs},
  author={Rota Bul\`o, Samuel and Porzi, Lorenzo and Kontschieder, Peter},
  journal={arXiv preprint arXiv:1712.02616},
  year={2017}
}

Overview

When processing a BN-Activation-Convolution sequence in the forward pass, most deep learning frameworks need to store two big buffers, i.e. the input x of BN and the input z of Conv. This is necessary because the standard implementations of the backward passes of BN and Conv depend on their inputs to calculate the gradients. Using Inplace-ABN to replace the BN-Activation sequence, we can safely discard x, thus saving up to 50% GPU memory at training time. To achieve this, we rewrite the backward pass of BN in terms of its output y, which is in turn reconstructed from z by inverting the activation function.

Installation

Our code has only been tested under Linux with CUDA 8.0 / 9.0 and CUDNN 7.0.

Requirements

To install PyTorch, please refer to https://github.com/pytorch/pytorch#installation.

NOTE: due to unresolved issues with PyTorch master, our code requires PyTorch v0.2 (see pytorch/pytorch#3883).

For all other dependencies, just run:

pip install -r requirements.txt

Compiling

Some parts of InPlace-ABN have a native CUDA implementation, which must be compiled with the following commands:

cd modules
sh build.sh
python build.py

The build.sh script assumes that the nvcc compiler is available in the current system search path. The CUDA kernels are compiled for sm_50, sm_52 and sm_61 by default. To change this (e.g. if you are using a Kepler GPU), please edit the CUDA_GENCODE variable in build.sh.

Training on ImageNet

Here you can find the results from our arXiv paper (top-1 / top-5 scores) with corresponding, trained models and md5 checksums, respectively. The model files provided below are made available under the license attached to ImageNet.

Data preparation

Our script uses torchvision.datasets.ImageFolder for loading ImageNet data, which expects folders organized as follows:

root/train/[class_id1]/xxx.{jpg,png,jpeg}
root/train/[class_id1]/xxy.{jpg,png,jpeg}
root/train/[class_id2]/xxz.{jpg,png,jpeg}
...

root/val/[class_id1]/asdas.{jpg,png,jpeg}
root/val/[class_id1]/123456.{jpg,png,jpeg}
root/val/[class_id2]/__32_.{jpg,png,jpeg}
...

Images can have any name, as long as the extension is that of a recognized image format. Class ids are also free-form, but they are expected to match between train and validation data. Note that the training data in the standard ImageNet distribution is already given in the required format, while validation images need to be split into class sub-folders as described above.

Training

The main training script is train_imagenet.py: this supports training on ImageNet, or any other dataset formatted as described above, while keeping a log of relevant metrics in Tensorboard format and periodically saving snapshots. Most training parameters can be specified as a json-formatted configuration file (look here for a complete list of configurable parameters). All parameters not explicitly specified in the configuration file are set to their defaults, also available in imagenet/config.py.

Our arXiv results can be reproduced by running train_imagenet.py with the configuration files in ./experiments. As an example, the command to train ResNeXt101 with InPlace-ABN, Leaky ReLU and batch_size = 512 is:

python train_imagenet.py --log-dir /path/to/tensorboard/logs experiments/resnext101_ipabn_lr_512.json /path/to/imagenet/root

Validation

Validation is run by train_imagenet.py at the end of every training epoch. To validate a trained model, you can use the test_imagenet.py script, which allows for 10-crops validation and transferring weights across compatible networks (e.g. from ResNeXt101 with ReLU to ResNeXt101 with Leaky ReLU). This script accepts the same configuration files as train_imagenet.py, but note that the scale_val and crop_val parameters are ignored in favour of the --scale and --crop command-line arguments.

As an example, to validate the ResNeXt101 trained above using 10-crops of size 224 from images scaled to 256 pixels, you can run:

python test_imagenet.py --crop 224 --scale 256 --ten_crops experiments/resnext101_ipabn_lr_512.json /path/to/checkpoint /path/to/imagenet/root

Usage for Semantic Segmentation on Cityscapes and Mapillary Vistas

The currently released code can only deal with image classification tasks (like ImageNet), but we have successfully used InPlace-ABN with a DeepLab3 segmentation head that was trained on top of the WideResNet38 model above. Due to InPlace-ABN, we can significantly increase the amount of input data to this model, which eventually allowed us to obtain #1 positions on Cityscapes and Mapillary Vistas segmentation leaderboards. The training settings mostly follow the description in our paper but we strive to release code later this year.