sigmaquan / act Goto Github PK

Alternative approach on Adaptive Computation Time for Recurrent Neural Networks by Alex Graves in TensorFlow.

Contributor: Berkay Antmen, Divyam Madaan, Bryan M. Li, Aidan N. Gomez

We use a small trick to help with the training of ACT modules. The idea is to penalize pondering less when the model is incorrect and more when it is correct.

In effect, relaxing the objective constraints when the model is struggling and then asking for computation efficiency only once the model has solved the problem. We implement this simply by scaling the ponder cost term by the inverse of the task loss. See function calculate_ponder_cost()

To use new ponder cost, set flag --use_new_ponder_cost=True

Call pip install -r requirements.txt to install all dependencies.

• parity
• sort
• addition

The settings for the reader used is passed as flag. For example, to run the parity task, call:

python -m ACT.train --model=act --hparam_sets=act_parity  --output_dir=runs/act_parity --data=parity  --train_steps=100000

Task	Accuracy	Sequence error rate	Pondering
Parity
Sort
Addition

• Our ponder cost
• LSTM cell
• Original ponder cost

Results for parity, sort and addition tasks are displayed. Light blue color represents LSTM cell, dark blue represents ACT cell with old ponder cost and green represents ACT cell with our new ponder cost.

The model with the new ponder cost ponders comparatively more than the model with the original ponder cost. The ACT cell with new ponder cost performed better than the old ponder cost for parity task, though performed similarly for sort and addition tasks.

We'd love to accept your contributions to this project. Please feel free to open an issue, or submit a pull request as necessary. If you have implementations of this repository in other ML frameworks, please reach out so we may highlight them here.

The ACT code is build upon DeNeutoy/act-tensorflow