/pytorch-office_finetune

A PyTorch implementation for fine-tuning AlexNet on Office dataset

Primary LanguagePythonMIT LicenseMIT

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