brendanofallon / jovian

Detecting variants in NGS short read data by sequence-to-sequence modeling
Other
7 stars 1 forks source link

batch_first=True for encoder, use GeLU activations #12

Closed brendanofallon closed 3 years ago

brendanofallon commented 3 years ago

Fixes a serious bug in the model TransformerEncoderLayer by adding batch_first=True to the encoder (by default it expects the second dimension to be batch??). Also uses "gelu" instead of relu activations for the transformer.

Changes the ENABLE_WANDB module var (which turns on all WANDB data uploading) to be set by the ENABLE_WANDB environment variable. This means w-and-b integration is off by default unless the env var is set. I think this might be helpful for debugging scenarios where we dont want to create a bunch of extra wandb runs.