PyTorch-Office_Finetune

A PyTorch implementation for fine-tuning AlexNet and ResNet on Office dataset.

Environment

• Python 3.6
• PyTorch 0.4.0

Result

	A-W
this(alexnet)	0.6000
this(resnet50)	0.7597

Note

• alexnet pretrained model is converted from caffe pretrained model (bvlc_reference_caffenet.caffemodel), using https://github.com/leelabcnbc/pytorch-caffe-models. Converted model can be download here, inference.py can be used as inference validation. inference.py under caffe directory is the Caffe version inference code.
• LRN layer is officially supported by PyTorch now
• Caffe's AlexNet implementation has different LRN/Pool layer order from original paper, this repo uses conv -> pool -> LRN order (better results). Refer to BVLC/caffe#296 for details
• tried https://github.com/jiecaoyu/pytorch_imagenet, results is bad (<50%)
• tried torchvision pretrained alexnet model, results is bad (~54%))
• tried correct order of classifier layers, refer to pytorch/vision#550, no improve
• fc/final0 and fc/final2 parameter init is important, best at 61%

links

• https://discuss.pytorch.org/t/pytorch-alexnet-not-as-good-as-original-alexnet-implementation/16680
• https://www.qin.ee/2018/06/25/da-office/