PyTorch implementation of the NIPS'17 paper Training Deep Networks without Learning Rates Through Coin Betting.

Paper: https://arxiv.org/abs/1705.07795 by Francesco Orabona, Tatiana Tommasi

Paper being validated by Anand Saha as part of NIPS paper implementation challenge.

• cocob.py: The optimizer in PyTorch
• cocob_report.ipynb: Notebook where the fully connected and CNN based models were tested on the various optimizers including cocob.
• word_language_model/: The LSTM model from PyTorch examples where the Penn Tree Bank corpus was tested on various optimizers including cocob.

The positives

• Very promising optimizer. Not having to decide the learning rate was refreshing change.
• The trends I got matched with the paper with some caveats mentioned below in concerns, but are not as smooth as the paper's
• In most cases, COCOB performed better than other optimizers but is yet to be battle tested.

The concerns

• It has more tensor operations than other optimizers. It does poorly on performance if batchsize is less (~100 in my case) and take almost 30-50% more time than the next expensive optimizer. However on increasing the batch size, it's performance came at par.
• I found it is sensitive to batch size, and somehow overfits easily if batch size is increased. This behaviour is not seen in other optimizers. (See 'Other Experiments' in the notebook cocob_report.ipynb) To compensave for that, I increased the 'L' or epsilon value (a hyper parameter in itself). I got better results in the CIFAR-10 case by increasing this value (see below). Might be a case where we are replacing one hyperparameter with another?

• The authors showed results only on the training set. I have included the results on validation set as well to understand how well it generalized, which can be seen in the notebook cocob_report.ipynb
• What the authors call the Test err is nothing but 1 - accuracy obtained on the Training set (and not test set). Hence I mark it as Train Err instead of Test Err.
• There were some minor differences in the models used by me vs the authors. However I think that should not matter because this is a comparison of the optimizers on a common model. The differences are:
  • For weights initialization, I used samples from normal distribution or xavier initialization rather than truncated_normal. PyTorch doesn't have the TensorFlow equivalent of truncated_normal, and I did not spend time in implementing it.
  • For CIFAR-10 on CNN, I used BatchNorm and did not use LRN as was used by the author's model. PyTorch doesn't have LRN and I did not spend time to implement it.
  • I used different learning rates for non-cocob optimizers than used by the authors so as to present all optimizers to the best of their ability, on my setup/models.
  • For the word prediction test, I used the word_language_model from PyTorch examples.

Check the cocob_report.ipynb for validation cost/errors.

Batch size was 100. The Train Err is nothing but 1 - accuracy. The graphs which shoot down actually could fit the entire training dataset, hence accuracy became 1.

Batch size was 100. The Train Err is nothing but 1 - accuracy.

With batchsize of 100 and L or 1e-8 (default):

The graphs which shoot down actually could fit the entire training dataset, hence accuracy became 1.

With batchsize of 1000 and L or 1e-2 (this smoothened the loss function and made cocob perform better):

The graphs which shoot down actually could fit the entire training dataset, hence accuracy became 1.