understanding-search / maze-transformer

This repo is built to facilitate the training and analysis of autoregressive transformers on maze-solving tasks.
24 stars 6 forks source link

Don't log dataset_seq_length_summary to w&b summary #128

Open rusheb opened 1 year ago

rusheb commented 1 year ago

In get_dataloader in training.py we are adding dataset_seq_length_stats to the w&b summary which produces a lot of clutter.

We should just log this to the console and remove it from the summary.

The dataset_seq_length_stats_summary is also being logged twice

    logger.summary({"dataset_seq_len_stats_summary": length_stats.summary()})
    logger.summary(
        {"dataset_seq_len_stats": length_stats.serialize(typecast=lambda x: str(x))}
    )

this is because the call to lengh_stats.serialize also includes the summary. We should remove the redundant first line.

traeuker commented 1 year ago

Seems to not happen anymore in the code; unsure, low-confidence, likely wrong, but also probably worth simply de-prioing