This document contains notes on the COMET codebase that the CEDModel inherits from.
COMETKiwi (the model we finetune) is an instance of the UnifiedMetrics class, which in turn inherits from the base CometModel class, which is an instance of a Pytorch Lightning module.
There are parameters in the UnifiedMetrics class which are leftovers from the base class and not actually used. A number of parameters are also passed when the class is instantiated but then hardcoded to some value within the class itself with the input ignored.
The below are notes on the hyperparameters listed in the hparams.yaml file that is downloaded from HuggingFace with the COMETKiwi DA 2022 model that are not otherwise described in the CEDModel docstring:
| Parameter | Default | Notes |
|---|---|---|
| class_identifier | unified_metric | We override this when creating our model. |
| layer | mix | This is a parameter in the base class that is not actually used. UnifiedMetrics has the sent_layer parameter instead. |
| pool | avg | When a UnifiedMetric object is created this isn't a parameter that can be set on initialisation. There is a default value in the base CometMetric class, which is avg, but this is only used in methods that the UnifiedMetric ignores. Instead, they hardcode the sentence embedding to be the CLS token.] |
| word_weights | 0.15, 0.85 | Can't actually find where this is used, doesn't seem to be an argument that the UnifiedMetric class is expecting. |
The COMET repository also contains a Pytorch Lightning training config.
Values for most of the parameters are just the PyTorch Lightning Trainer defaults. Specifically: benchmark, check_val_every_n_epoch, fast_dev_run, max_steps, min_steps, max_time, limit_train_batches, limit_val_batches, limit_test_batches, limit_predict_batches, overfit_batches, val_check_interval, log_every_n_steps, enable_progress_bar, enable_model_summary, gradient_clip_algorithm, benchmark, inference_mode, use_distributed_sampler, profiler, detect_anomaly, barebones, sync_batchnorm, reload_dataloaders_every_n_epochs.