.tinynet
├── README.md               # descriptions about tinynet
├── script
│   ├── eval.sh             # evaluation script
│   ├── train_1p_gpu.sh     # training script on single GPU
│   └── train_distributed_gpu.sh    # distributed training script on multiple GPUs
├── src
│   ├── callback.py         # loss, ema, and checkpoint callbacks
│   ├── dataset.py          # data preprocessing
│   ├── loss.py             # label-smoothing cross-entropy loss function
│   ├── tinynet.py          # tinynet architecture
│   └── utils.py            # utility functions
├── eval.py                 # evaluation interface
└── train.py                # training interface

# training on single GPU
  bash train_1p_gpu.sh
# training on multiple GPUs, the number after -n indicates how many GPUs will be used for training
  bash train_distributed_gpu.sh -n 8

--model tinynet_c               model to be used for training
--drop 0.2                      dropout rate
--drop-connect 0                drop connect rate
--num-classes 1000              number of classes for training
--opt-eps 0.001                 optimizer's epsilon
--lr 0.048                      learning rate
--batch-size 128                batch size
--decay-epochs 2.4              learning rate decays every 2.4 epoch
--warmup-lr 1e-6                warm up learning rate
--warmup-epochs 3               learning rate warm up epoch
--decay-rate 0.97               learning rate decay rate
--ema-decay 0.9999              decay factor for model weights moving average
--weight-decay 1e-5             optimizer's weight decay
--epochs 450                    number of epochs to be trained
--ckpt_save_epoch 1             checkpoint saving interval
--workers 8                     number of processes for loading data
--amp_level O0                  training auto-mixed precision
--opt rmsprop                   optimizers, currently we support SGD and RMSProp
--data_path /path_to_ImageNet/
--GPU                           using GPU for training
--dataset_sink                  using sink mode

# infer example

bash eval.sh

--num-classes 1000
--batch-size 128
--workers 8
--data_path /path_to_ImageNet/
--GPU
--ckpt /path_to_EMA_checkpoint/
--dataset_sink > tinynet_c_eval.log 2>&1 &