Skip to content

jerett/PyTorch-CIFAR10

Repository files navigation

PyTorch for CIFAR10

This project demonstrates some personal examples with PyTorch on CIFAR10 dataset.


Introduction


The CIFAR10 dataset is 32x32 size, 50000 train images and 10000 test images. The dataset is divided into 40000 train images, 10000 validation images, and 10000 images.

Features

  • Test for many models, each model is a a little different from orgin for 32*32 input, and will contiune to add new model.
  • Visdom realtime visualization of loss, acc, port 8097.
  • Use torchnet for training.
  • Use jupyter book for recording echo model training process.

Train

  • Run visdom first, python -m visdom.server &
  • Open the jupyter file for the corresponding model, and then run all cells.

Requirements

  • torch
  • torchvision
  • numpy
  • torchnet
  • visdom

Result

All result is tested on 10000 test images.You can lookup the jupyter for more details.

Model Accuracy
SVM 34.27%
Softmax 35.67%
small-ResNet20 91.38%
small-ResNet32 92.53%
small-ResNet56 93.31%
vgg11 91.25%
vgg13 92.84%
vgg16 92.94%%
MobileNetV1 92.45%
MobileNetV2 92.47%