/MILe

Primary LanguagePython

MILe

An unofficial implementation of MILe in Multi-label Iterated Learning for Image Classification with Label Ambiguity.

Features

Backbones

This implementation supports the following models as the backbone of MILe

  • ResNet18 for 1000 classes
  • ResNet50 for 1000 classes
  • ResNet18 for 10 classes

Note that it's easy for you to use other backbones by replacing teacher and student with the backbone in the code. But by that, you may need to modify the code of data preprocessing.

Benchmarks

I have implemented some benchmarks used in the paper.

  • ImageNet1k (for training) and ImageNet Real (for validating)
  • Multi-label MNIST

Get Started

Run the following command to train MILe.

python MILe.py

Configuration

You can modify all the configurations in config.py

  • model: the backbone of MILe. Valid options:

    • resnet18
    • resnet50
    • resnet18-10

  • schema: training strategy used in the paper. Valid options:

    • softmax softmax + cross entropy loss
    • sigmoid sigmoid + BCE loss
    • MILe sigmoid + BCE loss + iterated learning

  • data_path: path of the ImageNet train set. The structure of the directory should be:

    train/
  -- n01440764/
  -- n01443537/
  ...

  • val_path: path of the ImageNet val set. The structure of the directory should be the same as the train set.

  • real_path: path of the Real label file. See here.

  • mnist_path: path of the MNIST dataset. The structure of the directory should be:

    MNIST/
  -- raw/
    -- t10k-images-idx3-ubyte
    ...

    You don't need to manually download it because the torchvision.dataset.MNIST will do that if there is no dataset. So just set this parameter the root directory.

  • dataset: the benchmark you use. Valid options: -- imagenet, then the model should be resnet18 or resnet50 -- mnist, then the model should be resnet18-10

  • lr: learning rate

  • batch_size

  • num_workers: num_worker of dataloader

  • epoch_num: number of training epochs.

  • k_t: iteration number of interactive phase.

  • k_s: iteration number of imitation phase

  • device: training device. Valid options: -- cuda -- cpu

  • rho: threshold of pseudo-label

  • checkpoint_path: saving path of checkpoint

  • max_checkpoint_num: maximum number of checkpoints in the saving path.