Pooling

This is a collection of different pooling methods used in image classification segmentation detection

Features

• Multi-GPU support
• Easy and Useful Training log file
• Easy to test different pooling method on classification task

Requirements

• python3.6
• pytorch1.6.0 + cuda10.1
• tensorboard 2.3.0

Installation

• clone

  git clone https://github.com/rentainhe/pytorch-pooling.git
  

• make data directory for cifar100

  $ cd pytorch-pooling
  $ mkdir data

Usage

1. enter directory

$ cd pytorch-pooling

2. dataset

• Only support cifar100 now (Will support Imagenet Later)
• Using cifar100 dataset from torchvision since it's more convinient

3. run tensorboard

Install tensorboard

$ pip install tensorboard
Run tensorboard
$ tensorboard --logdir runs --port 6006 --host localhost

4. training

Our base backbone is vgg16 with batch_normalization

$ python3 train.py --run train --name test --pooling max

• --run={'train','test','visual'} to set the mode to be executed

• --name=str to set the name of this training

• --pooling=str, e.g, --pooling='max' to set the pooling method in vgg16 to be max_pool2d

• --gpu=str, e.g, --gpu='1' to set the specified GPU for training

The supported pooling args are

max pooling
average pooling
mixed pooling
Lp pooling
lip pooling
soft pooling

5. Add a new pooling method

You should add a new pooling method pool.py in "/Pooling/pooling_method" and update the __init__.py file

6. Addition

• lip pooling: the backbone in original paper is ResNet, But I use VggNet in this repo, so there might be something wrong with the accuracy

Results

The result I can get from this repo, I train every model with the same hyperparam and I don't use any tricks in this repo.

dataset	backbone	pooling	acc	epoch(lr = 0.1)	epoch(lr = 0.02)	epoch(lr = 0.004)	epoch(lr = 0.0008)	total epoch
cifar100	vgg16_bn	max	70.89%	60	60	40	40	200
cifar100	vgg16_bn	avg	70.56%	60	60	40	40	200
cifar100	vgg16_bn	mixed	71.19%	60	60	40	40	200
cifar100	vgg16_bn	Lp(p=2)	70.65%	60	60	40	40	200
cifar100	vgg16_bn	Lp(p=3)	70.67%	60	60	40	40	200
cifar100	vgg16_bn	lip	71.23%	60	60	40	40	200
cifar100	vgg16_bn	softpool	71.39%	60	60	40	40	200

Implementated Pooling

• mixed pooling Mixed pooling for convolutional neural networks
• Lp pooling Convolutional neural networks applied to house numbers digit classification
• lip pooling LIP: Local Importance-based Pooling
• soft pooling Refining activation downsampling with SoftPool