• I modified TorchVision official implementation of popular CNN models, and trained those on CIFAR-10 dataset.
• I changed number of class, filter size, stride, and padding in the the original code so that it works with CIFAR-10.
• I also share the weights of these models, so you can just load the weights and use them.
• The code is highly re-producible and readable by using PyTorch-Lightning.

Automatically download and extract the weights from Box (2.39 GB)

python cifar10_download.py

Or use Google Drive backup link (you have to download and extract manually)

Load model and run

from cifar10_models import *

# Untrained model
my_model = vgg11_bn()

# Pretrained model
my_model = vgg11_bn(pretrained=True)

If you use your own images, all models expect data to be in range [0, 1] then normalize by

mean = [0.4914, 0.4822, 0.4465]
std = [0.2023, 0.1994, 0.2010]

Check the cifar10_train.py to see all available hyper-parameter choices. To reproduce the same accuracy use the default hyper-parameters

python cifar10_train.py --classifier resnet18 --gpu '0,'

python cifar10_test.py --classifier resnet18 --gpu '0,'

Output

TEST RESULTS {'Accuracy': 93.33}

To see the training progress, cd to the tensorboard_logs and run TensorBoard there

tensorboard --logdir=. --port=YOUR_PORT_NUMBER

Then go to http://localhost:YOUR_PORT_NUMBER

Just to use pretrained models

• pytorch = 1.5.0

To train & test

• torchvision = 0.6.0
• tensorboard = 2.2.1
• pytorch-lightning = 0.7.6