Slow tokenization

[forrestbao]

PrefScore/model.py

Lines 10 to 25 in 05e2ffb

	import config as CFG
	class Scorer(nn.Module):
	def __init__(self):
	super(Scorer, self).__init__()
	self.tokenizer = BertTokenizer.from_pretrained(CFG.BERT_MODEL)
	self.model = BertModel.from_pretrained(CFG.BERT_MODEL)
	self.fc = nn.Linear(self.model.config.hidden_size, 1)
	def forward(self, article, summary):
	inputs = self.tokenizer(article, summary, padding='longest', truncation="longest_first" , return_tensors='pt').to(CFG.DEVICE)
	outputs = self.model(**inputs)
	x = self.fc(outputs.pooler_output)
	return x
	class Siamese(nn.Module):

Text pieces or pairs are tokenized individually and move to GPU. This is very slow.

At least we should tokenize them after a batch is loaded.