run.sh

#!/usr/bin/bash

# API key should probably not be hard coded into version control =)
# Add your wandb api key and wandb username here
wandb login TODO_add_wandb_api_key_here
export WANDB_ENTITY=joeyohman
export WANDB_PROJECT=ner_kram
export WANDB_MODE=offline

which python
/bin/hostname -s

# DATA_VAR="original_tags/lower"
# DATA_VAR="original_tags"
DATA_VAR="original_tags/lower_mix"
TUNE_ALG="ASHA"  # BOHB, PBT, or ASHA
dv=$(echo ${DATA_VAR} | sed 's/\//_/g')
TUNE_NAME="${dv}_${TUNE_ALG}"

# DATA_SUFFIX=""
# --train_file ./data/martin_data/train.lower.only.jsonl \
  #                  --validation_file ./data/martin_data/dev.lower.only.jsonl \
  #                  --test_file ./data/martin_data/test.lower.only.jsonl \

# For HPO
# --model_name_or_path KB/bert-base-swedish-cased \
# --do_train \

# For Eval All
# --model_name_or_path ./best_tuned_models/${TUNE_NAME} \
# --eval_all \

run_cmd="python run_ner.py
                  --model_name_or_path ./best_tuned_models/${TUNE_NAME} \
                  --eval_all \
                  --train_file ./data/${DATA_VAR}/train.jsonl \
                  --validation_file ./data/${DATA_VAR}/dev.jsonl \
                  --test_file ./data/${DATA_VAR}/test.jsonl \
                  --output_dir KB-BERT-ner-regular-tune \
                  --do_eval \
                  --do_predict \
                  --task_name ner \
                  --cache_dir models \
                  --return_entity_level_metrics 0 \
                  --per_device_train_batch_size 64 \
                  --per_device_eval_batch_size 128 \
                  --overwrite_output_dir \
                  --gradient_accumulation_steps 1
                  --num_train_epochs 5 \
                  --evaluation_strategy steps \
                  --save_strategy steps \
                  --skip_memory_metrics \
                  --eval_steps 200 \
                  --fp16 \
                  --disable_tqdm 1 \
                  --tune ${TUNE_NAME} \
                  --tune_alg ${TUNE_ALG} \
                  --tune_trials 50 \
                  --tune_local_dir ./ray_results/
                  "
                  # --eval_steps 10000 \
                  # --save_steps 10000 \
                  # --max_train_samples 3000
                  # --max_val_samples 100
                  # --max_test_samples 100
                  # --do_eval \

                  #
                  # --tune regular_lower_only \
                  #--train_file ./data/suc_train.both.jsonl \
                  #--validation_file ./data/suc_dev.both.jsonl \
                  #--test_file ./data/suc_test.both.jsonl \
                  #--learning_rate 0.000836949 \
                  #--weight_decay 0.0727959 \

                  # --learning_rate 2.2199e-05
                  # --weight_decay 0.015191

                  # --learning_rate 9.808539615186473e-06
                  # --weight_decay 0.06887642560886498

                  # after crashed BOHB on VEGA
                  # --learning_rate 4.9515e-05
                  # --weight_decay 0.15764
                  # test f1: 0.8561

                  # ASHA test with 27 trials:
                  # Current best trial: dd39fe7e with eval_f1=0.871240345297592 and
                  # parameters={'weight_decay': 0.028092325160590393, 'learning_rate': 2.1576349294034158e-05}

echo $run_cmd
$run_cmd