A classification repo implemented with PyTorch on CIFAR-10 and ImageNet under different training conditions.

• Python 3.6+
• PyTorch 1.5.0-cu10.1

Training different architectures (PyTorch) on the CIFAR10 dataset without any tricks i.e., auto-augmentation, cutout, droppath, dropout. The learning rate is adjusted by the consine learning schedular, start from 0.1 with 300 epochs.

Training the models in basic train with:

cd ./pytorch-cifar-basic; python main.py --model_name resnet18

Code base: darts

Torch version: 1.5.0+cu101

Training different architectures (PyTorch) on the CIFAR10 dataset with and without tricks i.e., cutout, droppath, dropout. The learning rate is adjusted by the consine learning schedular, start from 0.025, batch size 96 with 600 epochs.

Model: NASNet

Training condition

Namespace(auxiliary=True, auxiliary_weight=0.4, batch_size=96, cutout=True, cutout_length=16, data='/gdata/cifar10', drop_path_prob=0.2, epochs=600, gpu=0, grad_clip=5, init_channels=36, layers=20, learning_rate=0.025, model_name='nasnet', momentum=0.9, report_freq=50, seed=0, weight_decay=0.0003)

Code base: pt.darts

Torch version: 1.5.0+cu101

Training different architectures (PyTorch) on the CIFAR10 dataset with and without tricks i.e., cutout, droppath, dropout. The learning rate is adjusted by the consine learning schedular, start from 0.025, batch size 96 with 600 epochs.

Code base: py_cls

Torch version: 1.5.0+cu101

• pt.darts
• darts
• pytorch-cifar
• py_cls