YannDubs / PyTorch_CIFAR10

Pretrained TorchVision models on CIFAR10 dataset (with weights)

PyTorch models trained on CIFAR-10 dataset

• 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.

Statistics of supported models

No.	Model	Val. Acc.	No. Params	Size
1	vgg11_bn	92.09%	128.813 M	491 MB
2	vgg13_bn	94.29%	128.998 M	492 MB
3	vgg16_bn	93.91%	134.310 M	512 MB
4	vgg19_bn	93.80%	139.622 M	533 MB
5	resnet18	93.33%	11.174 M	43 MB
6	resnet34	92.92%	21.282 M	81 MB
7	resnet50	93.86%	23.521 M	90 MB
8	densenet121	94.14%	6.956 M	27 MB
9	densenet161	94.24%	26.483 M	102 MB
10	densenet169	94.00%	12.493 M	48 MB
11	mobilenet_v2	94.17%	2.237 M	9 MB
12	googlenet	92.73%	5.491 M	21 MB
13	inception_v3	93.76%	21.640 M	83 MB

How to use pretrained models

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]

How to train models from scratch

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,'

How to test trained models

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

Output

TEST RESULTS {'Accuracy': 93.33}

Check the TensorBoard logs

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

Requirements

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