TNT

This implements training of TNT on the ImageNet dataset, mainly modified from CV-Backbones/tnt_pytorch.

TNT Detail

TNT divides the image patches into sub-patches. With the structure of visual sentences and visual words, TNT can better perform the image classification task.

Requirements

• Install PyTorch (pytorch.org)
• pip install -r requirements.txt Note:Install the torchvision that corresponds to the torch version
• Download the ImageNet dataset
  • Then, and move validation images to labeled subfolders, using the following shell script
• Please add the following shapes into /usr/local/Ascend/ascend-toolkit/5.0.3/arm64-linux/opp/op_impl/built-in/ai_core/tbe/impl/dynamic/transpose.py for better performance.
  • [2, 25088, 4, 16, 6], [25088, 16, 2, 4, 6]

Training

To train a model, run main.py with the desired model architecture and the path to the ImageNet dataset:

# training 1p accuracy
bash ./test/train_full_1p.sh --data_path=real_data_path

# training 1p performance
bash ./test/train_performance_1p.sh --data_path=real_data_path

# training 8p accuracy
bash ./test/train_full_8p.sh --data_path=real_data_path

# training 8p performance
bash ./test/train_performance_8p.sh --data_path=real_data_path

#test 8p accuracy
bash test/train_eval_8p.sh --data_path=real_data_path --pth_path=real_pre_train_model_path

# finetuning 1p 
bash test/train_finetune_1p.sh --data_path=real_data_path --pth_path=real_pre_train_model_path

Log path: test/output/devie_id/train_${device_id}.log # training detail log test/output/devie_id/WideReesnet50_2_bs8192_8p_perf.log # 8p training performance result log test/output/devie_id/WideReesnet50_2_bs8192_8p_acc.log # 8p training accuracy result log

TNT training result

Acc@1	FPS	Npu_nums	Epochs	AMP_Type
-	57	1	1	O1
71.4%	410	8	79	O1

自测结果说明：

• 项目交付精度要求：Acc@1 81.5%，目前在gpu中Acc@1为81.2%，已达目标精度的99%以上。

由于跑完310个epoch训练时间过长，本项目中测试npu上79个epoch后Acc@1为71.4%，此时gpu的Acc@1为71.7%，Atlas（NPU）精度大于竞品V（GPU）精度的99%（即71.7% * 99% = 71.0%），说明精度已经对齐，符合验收标准。

Statement

For details about the public address of the code in this repository, you can get from the file public_address_statement.md