SE-ResNeXt101_32x4d

This implements training of SE-ResNeXt101_32x4d on the ImageNet dataset, mainly modified from pytorch/examples

SE-ResNeXt101_32x4d Detail

As of the current date, Ascend-Pytorch is still inefficient for contiguous operations.Therefore, SE-ResNeXt101_32x4d is re-implemented using semantics such as custom OP.

Requirements

• Install PyTorch (pytorch.org)
• pip install -r requirements.txt Note: pillow recommends installing a newer version. If the corresponding torchvision version cannot be installed directly, you can use the source code to install the corresponding version. The source code reference link: Suggestion the pillow is 9.1.0 and the torchvision is 0.6.0

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=/opt/npu/imagenet

# training 1p performance
bash test/train_performance_1p.sh --data_path=/opt/npu/imagenet

# training 8p accuracy
bash test/train_full_8p.sh --data_path=/opt/npu/imagenet

# training 8p performance
bash test/train_performance_8p.sh --data_path=/opt/npu/imagenet

#test 8p accuracy
bash test/train_eval_8p.sh --data_path=/opt/npu/imagenet --pth_path="./checkpointmodel_best.pth"

# finetuning 1p 
bash test/train_finetune_1p.sh --data_path=/opt/npu/imagenet --pth_path="checkpointmodel_best.pth"

Log path:

test/output/devie_id/train_${device_id}.log           # training detail log
test/output/devie_id/Se-ResNext101_bs1024_8p_perf.log  # 8p training performance result log
test/output/devie_id/Se-ResNext101_bs1024_8p_acc.log   # 8p training accuracy result log

online inference demo

python3 demo.py

SE-ResNeXt101_32x4d training result

Acc@1	FPS	Platform	Device_nums	Epochs	Type
-	221	GPU	1	1	O2
-	395	NPU	1	1	O2
78.34	1480	GPU	8	100	O2
77.75	1978	NPU	8	100	O2

公网地址说明

代码涉及公网地址参考 public_address_statement.md