train_distilbart_xsum.sh
No OneTemporary
Actions

Subscribers

None

File Metadata

	#!/usr/bin/env bash
	export PYTHONPATH="../":"${PYTHONPATH}"
	python distillation.py \
	--teacher facebook/bart-large-xsum --data_dir xsum \
	--tokenizer_name facebook/bart-large-xsum \
	--student_decoder_layers 6 --student_encoder_layers 12 \
	--freeze_encoder --freeze_embeds \
	--learning_rate=3e-4 \
	--do_train \
	--do_predict \
	--fp16 --fp16_opt_level=O1 \
	--val_check_interval 0.1 --n_val 1000 --eval_beams 2 --length_penalty=0.5 \
	--max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 \
	--model_name_or_path IGNORED \
	--alpha_hid=3. \
	--train_batch_size=16 --eval_batch_size=16 --gradient_accumulation_steps=2 \
	--sortish_sampler \
	--num_train_epochs=6 \
	--warmup_steps 500 \
	--output_dir distilbart_xsum_12_6 \
	"$@"