gobbli uses some heavy-duty abstractions to hide a lot of complexity, so there’s a lot of potential for complex errors along the way. Here are some things you might run into:

My model is predicting only one class

This indicates something’s wrong with the training process. Some things to check:

  • Do you have a very imbalanced dataset? You may need to raise the batch size, if possible, or try downsampling/upsampling to ensure you don’t get many batches composed of only one class.

  • Is your dataset ordered by label? The model needs to have a mix of classes in each batch to learn effectively. Ensure your training dataset is shuffled. gobbli Datasets take care of this with the train_input method.

I’m running out of memory

Your dataset might be too big to fit in memory. gobbli currently doesn’t support lazily loading datasets from disk, so anything you use to train has to fit in memory. You can try:

  • Sampling from your dataset.

  • If you’re running an experiment in parallel, try reducing the number of models training at a time to reduce the number of copies of your dataset in memory.

  • If you’re running a distributed experiment, the ray object store has to be large enough to fit weights for all models trained during the experiment at the same time. You may need to train fewer models or increase the size of the object store.

I’m running out of GPU memory

Some models are larger than others. You can try:

  • Decreasing the train_batch_size if you’re training; this is the biggest driver of GPU memory usage. Beware of making the batch size so small that the model can’t update gradients accurately, though. The gobbli.model.transformer.Transformer model supports gradient accumulation, which can be used to counteract the detrimental effect of a smaller batch size.

  • Decreasing the max_seq_len parameter of your model, if it has one. Consider using Document Windowing if you do this to account for the truncation of your texts.

  • Using a smaller set of pretrained weights (ex. instead of bert-large-uncased, try bert-base-uncased).

My disk is filling up

gobbli saves a ton of data to disk, especially if you have a large dataset, and won’t automatically remove any of it in case you need to access it later. If you need to free up some space, see Housekeeping.

A gobbli function appears to be hanging

gobbli sometimes needs to download very large files in the background – for example, the pretrained weights for BERT Large are over 1GB. It can look like nothing is happening while this is going on. You can try enabling debug logs to see (much) more detailed information about what gobbli is doing behind the scenes:

import logging

We have noticed that sometimes a download can time out repeatedly, depending on the quality of your internet connection. If this keeps happening, you can try manually downloading the file gobbli is trying to download by pasting the URL in your browser and moving the resulting file to the path where gobbli is trying to save it. gobbli should detect the file and not attempt to redownload the file the next time it tries to access the file.