PAPER

This is a pytorch implementation of Feedback-Network (CVPR 2017, Zamir et al.)

Requirements

• Pytorch = 0.3.1
• python = 2.7
• numpy >= 1.14.2

Train

• Data：

  • Download CIFAR100 dataset (binary file), and put it under folder ./data/

• Run：

  • Run classifier_train.py

  python FeedbackNet_train.py

  • The trained model will saved in folder ./models/ every 10 epochs。
  • Attention：
    • You can adjust the parameter batch_size to fit your GPU memory。

Evaluation

• Data：

  • Download CIFAR100 dataset (binary file), and put it under folder ./data/

• Evaluate on CIFAR100 dataset：

  • Put the checkpoint files under folder ./models/, and specify the checkpoints file path by setting parameter params.ckpt in file classifier_train.py. You can download my baseline model here
  • Run classifier_train.py​

  python FeedbackNet_test.py

Result

• Val Accuracy

Physical / Virtual Depth	PAPER(tp1)	PAPER(tp5)	ME(tp1)	ME(tp5)
12 / 48 (stack-3;iteration=4)	71.12%	91.51%	70.92%	92.02%

To Do

• Skip connections
• Multi-GPU
• Other virtual depths

Reference

• @amir32002 feedback-networks : The Torch7 repo of Feedback Networks, Zamir et al.
• Thanks @maxspero for the implemention of FeedbackNet. In this repository, I change a few codes in ./network/feedbacknet.py and ./network/convlstmstack.py to fit the model definition in the original paper
• @bzcheeseman_pytorch-feedbacknet