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.

Statistics of supported models

No.	Model	Validation Accuracy	Parameters	State Dict Size
1	vgg11_bn	92.61%	128.813 M	491 MB
2	vgg13_bn	94.27%	128.998 M	492 MB
3	vgg16_bn	94.07%	134.310 M	512 MB
4	vgg19_bn	94.25%	139.622 M	533 MB
5	resnet18	93.48%	11.174 M	43 MB
6	resnet34	93.82%	21.282 M	81 MB
7	resnet50	94.38%	23.521 M	90 MB
8	densenet121	94.76%	6.956 M	27 MB
9	densenet161	94.96%	26.483 M	102 MB
10	densenet169	94.74%	12.493 M	48 MB
11	mobilenet_v2	93.85%	2.237 M	9 MB
12	googlenet	95.08%	5.491 M	21 MB
13	inception_v3	95.41%	21.640 M	83 MB

How To Use

Download the weights

Download weights from Google Drive Link, and put the weights in models/state_dicts/ folder.

from models import *

# Untrained model
my_model = vgg11_bn()

# Pretrained model
my_model = vgg11_bn(pretrained=True)

Remember to normalize data before feeding to model

transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010])]

Training Hyper-paramters

• Batch size: 256
• Number of epochs: 600
• Learning rate: 0.05, multiply by factor 0.1 every 200 epochs
• Weight decay: 0.001
• Nesterov SGD optimizer with momentum = 0.9