General API Reference

This page contains the API reference for modules that contain code used for both the extractive and abstractive summarization components.

Helpers

class helpers.LabelSmoothingLoss(label_smoothing, tgt_vocab_size, ignore_index=-100)

CrossEntropyLoss with label smoothing, KL-divergence between q_{smoothed ground truth prob.}(w) and p_{prob. computed by model}(w) is minimized. From OpenNMT with modifications: https://github.com/OpenNMT/OpenNMT-py/blob/e8622eb5c6117269bb3accd8eb6f66282b5e67d9/onmt/utils/loss.py#L186

forward(output, target)

output (FloatTensor): batch_size x n_classes target (LongTensor): batch_size

class helpers.SortishSampler(data, batch_size, pad_token_id)

Go through the text data by order of src length with a bit of randomness. From fastai repo with modifications.

key(i)

class helpers.StepCheckpointCallback(step_interval=1000, save_name='model', save_path='.', num_saves_to_keep=5)

on_batch_end(trainer, pl_module)

Deprecated since version v1.6: This callback hook was deprecated in v1.6 and will be removed in v1.8. Use on_train_batch_end instead.

Called when the training batch ends.

helpers.block_trigrams(candidate, prediction)

Decrease repetition in summaries by checking if a trigram from prediction exists in candidate

Parameters:

• candidate (str) – The string to check for trigrams from prediction

• prediction (list) – A list of strings to extract trigrams from

Returns:

True if overlapping trigrams detected, False otherwise.

Return type:

bool

helpers.generic_configure_optimizers(hparams, train_dataloader, params_to_update)

Configure the optimizers. Returns the optimizer and scheduler specified by the values in hparams. This is a generic function that both the extractive and abstractive scripts use.

helpers.get_optimizer(hparams, optimizer_grouped_parameters)

helpers.load_json(json_file)

Load a json file even if it is compressed with gzip.

Parameters:

json_file (str) – Path to json file

Returns:

(documents, file_path), loaded json and path to file

Return type:

tuple

helpers.lr_lambda_func(current_step, num_warmup_steps, num_training_steps)

helpers.pad(data, pad_id, width=None, pad_on_left=False, nearest_multiple_of=False)

Pad data with pad_id to width on the right by default but if pad_on_left then left.

helpers.pad_tensors(tensors, pad_id=0, width=None, pad_on_left=False, nearest_multiple_of=False)

Pad tensors with pad_id to width on the right by default but if pad_on_left then left.

helpers.test_rouge(temp_dir, cand, ref)

Compute ROUGE scores using the official ROUGE 1.5.5 package. This function uses the pyrouge python module to interface with the office ROUGE script. There should be a “<q>” token between each sentence in the cand and ref files. pyrouge splits sentences based on newlines but we cannot store all the summaries easily in a single text file if there is a newline between each sentence since newlines mark new summaries. Thus, the “<q>” token is used in the text files and is converted to a newline in this function. Using “<q>” instead of \\n also makes it easier to store the ground-truth summaries in the convert_to_extractive.py script.

Parameters:

• temp_dir (str) – A temporary folder to store files for input to the ROUGE script.

• cand (str) – The path to the file containing one candidate summary per line with “<q>” tokens in between each sentence.

• ref (str) – The path to the file containing one ground-truth/gold summary per line with “<q>” tokens in between each sentence.

Returns:

Results from the ROUGE script as a python dictionary.

Return type:

dict