Re-implementation of ConvNets on ILSVRC 2012 classification task with PyTorch

Contact email: [email protected]

Introduction

Here are re-implementations of Convolutional Networks on ILSVRC 2012 with PyTorch.

Evaluation metrics are the top-1 and top-5 error rates.

Requirements

• A single TITAN RTX (24G memory) is used.

• Python 3.7+

• PyTorch 1.0+

Usage

1. Clone this repository

    git clone https://github.com/longrootchen/ILSVRC-2012-classification-pytorch.git
   

2. Train a model, taking alexnet as an example

    python -u train.py --work-dir ./experiments/alexnet --resume ./experiments/alexnet/checkpoints/last_checkpoint.pth
   

3. Evaluate a model on the validation set, taking alexnet as an example

    python -u eval.py --work-dir ./experiments/alexnet --ckpt-name last_checkpoint.pth --test-root ./datasets/val
   

Results

The single AlexNet model converges at 60-th epoch and achieving a top-1 error rate of 41.20% and a top-5 error rate of 18.20% on the validation set.

Error Rate (%)	Top-1 origin	Top-5 origin	Top-1 re-implementation	Top-5 re-implementation
AlexNet [1]	40.7%	18.2%	41.20%	18.20%

Here are visualizations for training loss and error rates (dark blue for train, light blue for val; red for train, pink for val) for AlexNet.

References

[1] Alex Krizhevsky, Ilya Sutskever & Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In NIPS, 2012.