EfficientNet-B3

This implements training of Efficientnet-B3 on the ImageNet dataset, mainly modified from pycls.

EfficientNet-B3 Detail

For details, seepycls.

Requirements

• Install PyTorch (pytorch.org)
• pip install pycls
• git clone https://github.com/facebookresearch/pycls
• pip install -r requirements.txt
• modify path of dataset in pycls/datasets/loader.py, you can modify the variable _DATA_DIR to your path of imagenet dataset.

Training

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

# 1p train 1p
bash test/train_full_1p.sh --data_path={data/path} # train accuracy

bash test/train_performance_1p.sh --data_path={data/path} # train performance

#  8p train 8p
bash test/train_full_8p.sh --data_path={data/path} # train accuracy

bash test/train_performance_8p.sh --data_path={data/path} # train performance

# 1p eval 1p
bash test/train_eval_8p.sh --data_path={data/path}

# online inference demo 
python3 demo.py

# To ONNX
python3 pthtar2onnx.py

EfficientNet-B3 training result

Acc@1	FPS	Npu_nums	Epochs	AMP_Type
-	267	1	100	O2
77.3418	1558	8	100	O2