Code for reproducing experiments in AutoLoss: Learning Discrete Schedules for Alternate Optimization.

python>=3.6, tensorflow-gpu==1.8.0, matplotlib==2.2.2, imageio==2.4.1, sklearn==0.20.2, nltk==3.4, subword-nmt==0.3.6

Install and Activate a Virtual Environment

virtualenv --python=python3.6 ./env
source ./env/bin/activate

Install Dependencies

pip install -r requirements.txt

The regression and classification tasks use synthetic data. To synthesize a dataset for the regression or the classification experiment, run the following script:

# regression
python dataio/gen_reg_data.py

# classification
python dataio/gen_cls_data.py

Our GANs experiments use MNIST and CIFAR-10 datasets. Set a desired path for the argument data_dir in config file config/gan/cfg.py. The training scripts will automatically download the data at its first run.

The CIFAR-10 data is available at Download page of CIFAR10. You need to download the python version from this page and unzip it. Set the argument 'data_dir' in config file 'config/gan_cifar.py' to where you save the data.

You can directly use the preprocessed data we have provided in this repository under the path data/nmt/. Or you can prepare the data by running the following script:

python dataio/gen_nmt_data.py

In this case, you need to download the tigercorpus-2.2.xml.tar.gz from the TIGER corpus and uncompress it to the path data/nmt/pos. The script will expect an XML file named tiger\_release\_aug07.corrected.16012013.xml.

Caution: The data preparation for the multi-task neural machine translation task could take 10-15 mins on a desktop with Intel i7-6800K CPU @ 3.40GHz x 12 CPU.

Use the following script to launch an experiment:

python trainer.py
    --task_name=[task_name] 
    --task_mode=[task_mode] 
    --exp_name=[exp_name]

where

• task_name is one of: reg, cls, gan, gan_cifar10, nmt.
• task_mode is one of: train, test, baseline.
• exp_name can be any string you would like use to name this experiment.

Train a controller on the regression task:

python trainer.py --task_name=reg --task_mode=train --exp_name=reg_train

After the training is done, we use the trained controller to guide the training of the regression model on a new dataset:

# Make sure the `exp_name` is set as the one used in the controller training experiment. 
python trainer.py --task_name=reg --task_mode=test --exp_name=reg_train

The script will automatically load the controller trained in the reg_train experiment.

Alternatively, you can specify a specific checkpoint you want to test with by:

python trainer.py --task_name=reg --task_mode=test --exp_name=reg_test --load_ctrl=/path/to/checkpoint/folder/

To compare the results with a baseline training schedule:

python trainer.py --task_name=reg --task_mode=baseline --exp_name=reg_baseline

You can design your own training schedule through the class controller_designed in models/reg.py

It may take days to train a controller on GANs task. We provide a pretrained controller for MNIST GAN. To test this controller:

python trainer.py --task_name=gan --task_mode=test --exp_name=[any-experiment-name] --load_ctrl=/path/to/AutoLossRelease/weights/gan_2l_adam_short_ctrl --model_dir=/path/to/AutoLossRelease/weights

If you use any part of this code in your research, please cite our paper:

@misc{xu2018autoloss,
    title={AutoLoss: Learning Discrete Schedules for Alternate Optimization},
    author={Haowen Xu and Hao Zhang and Zhiting Hu and Xiaodan Liang and Ruslan Salakhutdinov and Eric Xing},
    year={2018},
    eprint={1810.02442},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}