electra_large_squad2_512.sh
947 B · 37 lines · bash Raw
1 #!/bin/bash
2
3 export OMP_NUM_THREADS=12
4 EXAMPLES=/media/dn/work2/nlpu/transformers/examples/question-answering
5 SQUAD=${EXAMPLES}/scripts/squad2.0
6 MODEL_PATH=${EXAMPLES}/runs/electra_large_discriminator_squad2_512
7
8 python ${EXAMPLES}/run_squad.py \
9 --model_type electra \
10 --model_name_or_path google/electra-large-discriminator \
11 --do_train \
12 --do_eval \
13 --train_file ${SQUAD}/train-v2.0.json \
14 --predict_file ${SQUAD}/dev-v2.0.json \
15 --version_2_with_negative \
16 --do_lower_case \
17 --num_train_epochs 3 \
18 --warmup_steps 306 \
19 --weight_decay 0.01 \
20 --learning_rate 3e-5 \
21 --max_grad_norm 0.5 \
22 --adam_epsilon 1e-6 \
23 --max_seq_length 512 \
24 --doc_stride 128 \
25 --per_gpu_train_batch_size 8 \
26 --gradient_accumulation_steps 16 \
27 --per_gpu_eval_batch_size 128 \
28 --fp16 \
29 --fp16_opt_level O1 \
30 --threads 12 \
31 --logging_steps 50 \
32 --save_steps 1000 \
33 --overwrite_output_dir \
34 --output_dir ${MODEL_PATH}
35 $@
36
37