In this repo, the training set is divided into two parts: one is for training with size 50,000 and the other is for evaluating with size 10,000.
The processed evaluating dataset is located in './data/processed/val-set' with the name 'test.pth', while the testing dataset is located in './data/processed/test-set' with the name 'test.pth'. Before you train or test the net, you should move the according 'test.pth' to './data/processed'.
MNIST data is also supported in this repo, and the data can be downloaded and processed automatically if you set --data MNIST in train script.
Training
To train drn_a_50 Net using the train script simply specify the parameters listed in train_drn.py as a flag or manually change them. To train resnet50 is in the same way except using the 'train-resnet50.py' script.
python train-drn.py
Note:
--patience :early stopping
--batch_size : batch size
--nepochs: max epochs
--nworkers: number of workers
--seed : random seed
--data : FashionMNIST or MNIST
Evaluation
You can test the drn_a_50 Net with the code below, and to test resnet50 is as the same except using 'test-resnet50.py'.