lilujunai / coarse-to-fine-curriculum

Coarse-to-Fine Curriculum Learning

Coarse-to-Fine Curriculum Learning

This code repository contains an implementation of the work described in:

O. Stretcu, E.A. Platanios, T. Mitchell, B. Poczos. "Coarse-to-Fine Curriculum Learning." Bridging AI and Cognitive Science (BAICS) Workshop at ICLR 2020.

To cite this work please use the following citation:

@inproceedings{coarse-to-fine:2020,
  title={{Coarse-to-Fine Curriculum Learning for Classification}},
  author={Stretcu, Otilia and Platanios, Emmanouil Antonios and Mitchell, Tom and P{\'o}czos, Barnab{\'a}s},
  booktitle={International Conference on Learning Representations (ICLR) Workshop on Bridging AI and Cognitive Science (BAICS)},
  year={2020},
}

The code is organized into the following folders:

• data: contains classes and methods for loading and preprocessing data.
• learning:
  • models.py: model architectures implementations.
  • trainer.py: a class in charge of training a provided model.
  • hierarchy_computation.py: code for computing the label hierarchy.
• utils: other utility functions, including our implementation of the Affinity clustering algorithm.

More details can be found in our paper, video & slides,

Requirements

Our code was implemented in Python 3.7 with Tensorflow 1.14.0 using eager execution. It also requires the following Python packages:

• matplotlib
• numpy
• tensorflow-datasets
• tqdm
• urllib
• yaml

How to run

Our code currently supports the datasets included in the tensorflow-datasets package, as well as Tiny ImageNet and the synthetic Shapes dataset described in our paper.

CIFAR-100, CIFAR-10 and more

To run on tensorflow-datasets such as CIFAR-100, from the base repository folder run the command:

python3.7 -m src.run \
  --dataset=cifar100 \
  --normalization=norm

Shapes

To run on the Shapes dataset, from the base repository folder run the command:

python3.7 -m src.run \
  --dataset=shapes \
  --data_path=data/shapes \
  --normalization=norm

We provide the Shapes dataset used in our experiments in the folder data/shapes.

Tiny ImageNet

To run on the Tiny ImageNet dataset, from the base repository folder run the command:

python3.7 -m src.run \
  --dataset=tiny_imagenet \
  --data_path=data/tiny_imagent \
  --normalization=norm

This will automatically download the Tiny ImageNet dataset from the official website to the folder specified by the --data_path flag.

Advice for running the code

Note that our code offers a range of commandline arguments, which are described in run.py. An important argument is normalization which specifies what type of preprocessing to perform on the data (e.g, in the examples above --normalization=norm specifies that inputs should be converted from integer pixels in [0, 255] to float numbers in [0, 1]).

Another useful argument is max_num_epochs_auxiliary which can limit the maximum number of epochs for each of the auxiliary functions (when training on coarser labels). In our experiments we set this to a number lower than num_epochs.

We recommend running on a GPU. With CUDA, this can be done by prepending CUDA_VISIBLE_DEVICES=<your-gpu-number> in front of the run command.

Visualizing the results

To visualize the results in Tensorboard, use the following command, adjusting the dataset name accordingly: tensorboard --logdir=outputs/summaries/cifar100

An example of such visualization on the CIFAR-100 dataset is the following:

The command we ran for this experiment was:

python3.7 -m src.run \
  --dataset=cifar100 \
  --normalization=norm \
  --num_epochs=300 \
  --max_num_epochs_auxiliary=30 \
  --max_num_epochs_no_val_improvement=50

The image above shows the accuracy per epoch on the test set for the baseline model trained without curriculum (the model with the suffix "original" [green]), as well as the accuracy of the curriculum models trained at each level of the label hierarchy (identified through the suffixes "level_1" [grey], "level_2" [orange], etc.). The model trained at the highest level ("level_4" in this case [red]) is trained on the original label space, and is thus comparable to the "original" [green] model.