├── TNT
  ├── README.md              # descriptions about TNT
  ├── configs
  │   ├── tnt_b_patch16_224_imagenet.yaml  # example of configuration for TNT-B
  │   ├── tnt_s_patch16_224_imagenet.yaml  # example of configuration for TNT-B
  ├── scripts
  │   ├── run_distribute_train_gpu.sh  # shell script for distributed training on multiple GPUs
  │   ├── run_eval_gpu.sh              # shell script for MindSpore model evaluation
  │   ├── run_eval_onnx.sh             # shell script for ONNX model evaluation
  │   ├── run_infer_gpu.sh             # shell script for MindSpore model inference
  │   ├── run_infer_onnx.sh            # shell script for ONNX model inference
  │   ├── run_standalone_train_gpu.sh  # example of shell script for training on 1 GPU
  ├── src
  │   ├── configs
  |   │   ├── config.py          # YAML configuration parser
  │   ├── data
  │   |   ├── augment
  |   │   |   ├── auto_augment.py    # augmentation set builder
  |   │   |   ├── mixup.py           # MixUp augmentation
  |   │   |   ├── random_erasing.py  # 'random erasing' augmentation
  │   |   ├── data_utils
  |   │   |   ├── moxing_adapter.py  # DS synchronization for distributed training
  |   |   ├── imagenet.py            # wrapper for reading ImageNet dataset
  │   ├── models
  │   |   ├── tnt
  │   |   |   ├── layers
  |   │   |   |   ├── attention.py  # Attention layer
  |   │   |   |   ├── misc.py  # extra tools and layers
  |   │   |   |   ├── patch_embed.py  # PatchEmbed layer
  |   │   |   |   ├── unfold_kernel.py  # Unfold
  |   │   |   ├── tnt.py  # implementation of TNT architecture
  │   ├── tools
  │   |   ├── callbacks.py  # callback functions (implementation)
  │   |   ├── cell.py  # tune model layers/parameters
  │   |   ├── common.py  # initialize callbacks
  │   |   ├── criterion.py  # model training objective function (implementation)
  │   |   ├── get_misc.py  # initialize optimizers and other arguments for training process
  │   |   ├── optimizer.py  # model optimizer function (implementation)
  │   |   ├── schedulers.py  # training (LR) scheduling function (implementation)
  │   ├── trainers
  │   |   ├── train_one_step_with_scale_and_clip_global_norm.py  # example training script

  ├── convert_from_pt.py  # convert model weights from PyTorch model
  ├── eval.py             # evaluation script (MindSpore)
  ├── eval_onnx.py        # evaluation script (ONNX)
  ├── export.py           # export MindIR/ONNX script
  ├── infer.py            # run inference on single file/folder (MindSpore)
  ├── infer_onnx.py       # run inference on single file/folder (ONNX)
  ├── train.py            # training script
  ├── tune_params.py      # tune hyperparameters for trainable model

# training example
  python:
      GPU: python train.py --tnt_config configs/tnt_s.yaml \
             --dir_ckpt /mindspore/save/ckpt/ --dir_best_ckpt /mindspore/save/best_ckpt/ \
             --dir_summary /mindspore/save/summary/
  shell:
      GPU: bash scripts/run_train.sh

epoch: [  0/200], step:[  624/  625], loss:[5.258/5.258], time:[140412.236], lr:[0.100]
epoch time: 140522.500, per step time: 224.836, avg loss: 5.258
epoch: [  1/200], step:[  624/  625], loss:[3.917/3.917], time:[138221.250], lr:[0.200]
epoch time: 138331.250, per step time: 221.330, avg loss: 3.917

# infer example
  python:
    GPU: python eval.py --dataset_path ~/imagenet/val/ --pretrained pretrained.ckpt --device_target GPU

result: {'acc': 0.7252} ckpt=/path/to/checkpoint/tnt_s_patch16_224-200_625.ckpt

python export.py --pretrained [CKPT_PATH] --tnt_config configs/tnt_s.yaml --file_format [FILE_FORMAT]