GLUE Benchmark, BPE Tokenizer, RoBERTa and WarmUp Scheduler Series

Updates:

• Add base_config for glue
• Add regression to model, reader and dataset
• Add glue data readers and metrics
• Fixed dependency pytorch-transformers==1.1.0
• Tokenizer: BPE(Byte-Pair Encoding) Tokenizer (name: roberta)
• Mode: roberta_for_seq_cls and roberta_for_reg (Sequence Classification and Regression)
• BaseConfig: add roberta base model base_configs for glue tasks
• refactor utils.make_bert_input with input_type (bert or roberta)

GLUE Report

Task (Metric)	Model	Result	Official Result	BaseConfig
CoLA (Matthew's Corr)	BERT-Base	59.393	52.1 (Test set)	glue/cola_bert_base.json
	BERT-Large	-	60.6	-
	RoBERTa-Base	64.828	63.6	glue/cola_roberta_base.json
	RoBERTa-Large	-	68.0	-
MNLI (Accuracy)	BERT-Base	83.923/84.306	84.6/83.4 (Test set)	glue/mnli{m/mm}_bert_base.json
	BERT-Large	-	86.6/-	-
	RoBERTa-Base	87.305/87.236	87.6/-	glue/mnli{m/mm}_roberta_base.json
	RoBERTa-Large	-	90.2/90.2	-
MRPC (Accuracy/F1)	BERT-Base	87.5/91.282	88.9 (Test set)	glue/mrpc_bert_base.json
	BERT-Large	-	88.0	-
	RoBERTa-Base	87.745/91.496	90.2	glue/mrpc_roberta_base.json
	RoBERTa-Large	-	90.9	-
QNLI (Accuracy)	BERT-Base	88.521	90.5 (Test set)	glue/qnli_bert_base.json
	BERT-Large	-	92.3	-
	RoBERTa-Base	90.597	92.8	glue/qnli_roberta_base.json
	RoBERTa-Large	-	94.7	-
QQP (Accuracy/F1)	BERT-Base	90.378/87.171	71.2 (Test set)	glue/qqp_bert_base.json
	BERT-Large	-	91.3	-
	RoBERTa-Base	91.541/88.768	91.9	glue/qqp_roberta_base.json
	RoBERTa-Large	-	92.2	-
RTE (Accuracy)	BERT-Base	69.314	66.4 (Test set)	glue/rte_bert_base.json
	BERT-Large	-	70.4	-
	RoBERTa-Base	73.646	78.7	glue/rte_roberta_base.json
	RoBERTa-Large	-	86.6	-
SST-2 (Accuracy)	BERT-Base	92.546	93.5 (Test set)	glue/sst_bert_base.json
	BERT-Large	-	93.2	-
	RoBERTa-Base	94.495	94.8	glue/sst_roberta_base.json
	RoBERTa-Large	-	96.4	-
STS-B (Pearson/Spearman)	BERT-Base	88.070/87.881	85.8 (Test set)	glue/stsb_bert_base.json
	BERT-Large	-	90.0	-
	RoBERTa-Base	87.033/86.904	91.2	glue/stsb_roberta_base.json
	RoBERTa-Large	-	92.4	-
WNLI (Accuracy)	BERT-Base	56.338	65.1 (Test set)	glue/wnli_bert_base.json
	BERT-Large	-	-	-
	RoBERTa-Base	60.563	-	glue/wnli_roberta_base.json
	RoBERTa-Large	-	91.3	-

Refactoring:

• Add claf.data.dto package
  • Since the structure is defined, providing specifications with DTO helps understand the code (logic).
  • Batch, BertFeature and Helper
• Apply dto object to all readers.
• Fixed typo (#19, #16)
• Beautifully redrawn the image (#14)