I trained ResNet models using the Tiny ImageNet dataset. This implementation uses PyTorch Lightning.
hparams = {
"pretrained": True,
"output_size": 200,
"lr": 0.1,
"max_epochs": 50,
"weight_decay": 5e-4,
"batch_size": 128,
"seed": 12345,
}
python3 train.py \
--dataset_path="tiny-imagenet-200" \
--pretrained \
--model="resnet50" \
--verbose=1
Model | Test accuracy |
---|---|
resnet18 | 0.6781 |
resnet34 | 0.6994 |
resnet50 | 0.7151 |
resnext50 | 0.7256 |
wide_resnet50 | 0.7303 |
The dataset loading code in dataset.py
is adapted from https://github.com/pranavphoenix/TinyImageNetLoader.