SWAG transformers

This repository provides functionality for Stochastic Weight Averaging-Gaussian training for Transformer models. The implementation is tied into two libraries:

• transformers (maintained by Hugging Face)
• swa_gaussian (maintained by the Language Technology Research Group at the University of Helsinki, a fork of https://github.com/wjmaddox/swa_gaussian)

The goal is to make an implementation that works directly with the convenience tools in the transformers library (e.g. Pipeline and Trainer) as well as evaluator from the related evaluate library.

Usage

See also examples.

Fine-tuning with SWAG

BERT model, sequence classification task:

1. Load pretrained Bert model by base_model = AutoModelForSequenceClassification.from_pretrained(name_or_path)
2. Initialize SWAG model by swag_model = SwagBertForSequenceClassification.from_base(base_model, no_cov_mat=False)
3. Initialize SWAG callback object swag_callback = SwagUpdateCallback(swag_model)
4. Initialize transformers.Trainer with the base_model as model and swag_callback in callbacks.
5. Train the model (trainer.train())
6. Store the complete model using swag_model.save_pretrained(path)

Note that trainer.save_model(path) will save only the base model without the distribution parameters from SWAG.

For collecting the SWAG parameters, two possible schedules are supported:

• After the end of each training epoch (default, collect_steps = 0 for SwagUpdateCallback)
• After each N training steps (set collect_steps > 0 for SwagUpdateCallback)

SWA, SWAG-Diagonal, and SWAG

The library supports both SWA (stochastic weight averaging) and two variants of SWAG (SWA-Gaussian): SWAG-Diagonal that uses diagonal covariance and "full" SWAG that does low-rank covariance matrix estimation.

The method is selected by the no_cov_mat attribute when initializing the model (e.g. SwagModel.from_base(model, no_cov_mat=True)). The default value True works only with SWAG-Diagonal and SWA, and you need to explicitly set no_cov_mat=False to activate the low-rank covariance estimation of SWAG. Note that you can also test SWA and SWAG-Diagonal methods when the model is trained with no_cov_mat=False (see the next section).

With SWAG, the max_num_models option controls the maximum rank of the covariance matrix. The rank is increased by each parameter collection step until the maximum is reached. The current rank is stored in model.swag.cov_mat_rank and automatically updated to model.config.cov_mat_rank when using SwagUpdateCallback. If you call model.swag.collect_model() manually, you should also update the configuration accordingly before saving the model.

Restricting SWAG to certain parameters

For N original parameters, SWAG requires:

• N mean values (SWA, SWAG-Diag, SWAG)
• N squared mean values for variances (SWAG-Diag & SWAG)
• max_num_models x N parameters for covariance matrix estimation (SWAG)

This means that for full SWAG, the number of parameters may easily grow e.g. ten times larger than in the baseline model.

However, it does not always make sense to estimate the full (co)variance for all of the parameters. With the module_prefix_list option, variance estimation can be limited to certain modules of the model. The prefixes in the list are matched to full names of the parameters. For example, with BERT, embeddings.word_embeddings.weight would be matched by prefix embeddings.word_embeddings and encoder.layer.11.output.dense.weight by prefix encoder.layer.11. If module_prefix_list is provided, the mean (SWA method) is used for all parameters that do not match any of the prefixes.

For tied parameters, you should provide the name of module that actually stored the parameters.

Sampling model parameters

After swag_model is trained or fine-tuned as described above, swag_model.sample_parameters() should be called to sample new model parameters. After that, swag_model.forward() can be used to predict new output from classifiers and swag_model.generate() to generate new output from generative LMs. In order to get a proper distribution of outputs, sample_parameters() needs to be called each time before forward() or generate(). For classifiers, the SampleLogitsMixin class provides the convenience method get_logits() that samples the parameters and makes a new prediction num_predictions times, and returns the logit values in a tensor.

Note that both for sample_parameters() and get_logits() the default keyword arguments are suitable only for SWAG-Diagonal. For SWAG, you should use cov=True (required to use the covariance matrix) and scale=0.5 (recommended). For SWA, you should use cov=False and scale=0. To summarize:

• SWA: scale=0, cov=False
• SWAG-Diagonal: scale=1, cov=False (defaults)
• SWAG: scale=0.5, cov=True (no_cov_mat=False required for the model)

Currently supported models

• BERT (bidirectional encoder)
  • BertPreTrainedModel -> SwagBertPreTrainedModel
  • BertModel -> SwagBertModel
  • BertLMHeadModel -> SwagBertLMHeadModel
  • BertForSequenceClassification -> SwagBertForSequenceClassification
  • BertForQuestionAnswering -> SwagBertForQuestionAnswering
• BART (bidirectional encoder + causal decoder)
  • BartPreTrainedModel -> SwagBartPreTrainedModel
  • BartModel -> SwagBartModel
  • BartForConditionalGeneration -> SwagBartForConditionalGeneration
  • BartForSequenceClassification -> SwagBartForSequenceClassification
• MarianMT (bidirectional encoder + causal decoder)
  • MarianPreTrainedModel -> SwagMarianPreTrainedModel
  • MarianModel -> SwagMarianModel
  • MarianMTModel -> SwagMarianMTModel