def collate(data, tokenizer, block_size):
""" List of tuple as an input. """
# remove the files with empty an story/summary, encode and fit to block
data = filter(lambda x: not (len(x[0]) == 0 or len(x[1]) == 0), data)
data = [
encode_for_summarization(story, summary, tokenizer)
for story, summary in data
]
data = [(
fit_to_block_size(story, block_size, tokenizer.pad_token_id),
fit_to_block_size(summary, block_size, tokenizer.pad_token_id),
) for story, summary in data]
stories = torch.tensor([story for story, summary in data])
summaries = torch.tensor([summary for story, summary in data])
encoder_token_type_ids = compute_token_type_ids(stories,
tokenizer.cls_token_id)
encoder_mask = build_mask(stories, tokenizer.pad_token_id)
decoder_mask = build_mask(summaries, tokenizer.pad_token_id)
lm_labels = build_lm_labels(summaries, tokenizer.pad_token_id)
return (
stories,
summaries,
encoder_token_type_ids,
encoder_mask,
decoder_mask,
lm_labels,
)
def collate(data, tokenizer, block_size, device):
""" Collate formats the data passed to the data loader.
In particular we tokenize the data batch after batch to avoid keeping them
all in memory. We output the data as a namedtuple to fit the original BertAbs's
API.
"""
data = [x for x in data if not len(x[1]) == 0] # remove empty_files
names = [name for name, _, _ in data]
summaries = [" ".join(summary_list) for _, _, summary_list in data]
encoded_text = [encode_for_summarization(story, summary, tokenizer) for _, story, summary in data]
encoded_stories = torch.tensor(
[truncate_or_pad(story, block_size, tokenizer.pad_token_id) for story, _ in encoded_text]
)
encoder_token_type_ids = compute_token_type_ids(encoded_stories, tokenizer.cls_token_id)
encoder_mask = build_mask(encoded_stories, tokenizer.pad_token_id)
batch = Batch(
document_names=names,
batch_size=len(encoded_stories),
src=encoded_stories.to(device),
segs=encoder_token_type_ids.to(device),
mask_src=encoder_mask.to(device),
tgt_str=summaries,
)
return batch