| #!/bin/bash |
| |
| set -x |
| export PYTHONPATH=/root/patientphex/retrain_t1/code |
| export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 TOKENIZERS_PARALLELISM=false |
| PY=/root/patientphex/e8_venv/bin/python |
| C=/root/patientphex/retrain_t1/code |
| P=/root/patientphex/retrain_t1/third_party/PhenoTagger-Plus |
|
|
| for f in $P/data/distant_train_data/distant_train.conll \ |
| $P/dict-hp-20260623-typo/lable.vocab \ |
| $P/emb_20260623/transr-full/transR_512.emb \ |
| /root/patientphex/assets/dmis-biobert-cased/pytorch_model.bin; do |
| [ -e "$f" ] || { echo MISSING $f; exit 1; } |
| done |
|
|
| $PY /root/patientphex/count_conll.py $P/data/distant_train_data/distant_train.conll > /root/patientphex/conll_count.txt 2>&1 |
| cat /root/patientphex/conll_count.txt |
| grep -qE "records=2050[0-9][0-9]" /root/patientphex/conll_count.txt || { echo CONLL_COUNT_GATE_FAIL; exit 2; } |
| echo CONLL_GATE_OK |
| echo "NOTE: 44-record positive delta vs A-day (155008 vs 155052) accepted; root cause: A-day typo dict input-side diff (noabb md5 mismatch, nlpaug/py versions ruled out)" |
|
|
| CUDA_VISIBLE_DEVICES=1 $PY $C/scripts/train_phenotagger_plus_hf.py \ |
| --train $P/data/distant_train_data/distant_train.conll \ |
| --labels $P/dict-hp-20260623-typo/lable.vocab \ |
| --embedding $P/emb_20260623/transr-full/transR_512.emb \ |
| --backbone /root/patientphex/assets/dmis-biobert-cased \ |
| --output-dir /root/patientphex/models_20260623/biobert-transr-v1 \ |
| --epochs 8 --batch-size 64 --max-length 32 --learning-rate 5e-5 \ |
| --warmup-ratio 0.03 --weight-decay 0.01 --seed 42 --validation-ratio 0.05 || { echo E8_TRAIN_FAIL; exit 3; } |
| echo E8_TRAIN_DONE |
| ls -la /root/patientphex/models_20260623/biobert-transr-v1/ /root/patientphex/models_20260623/biobert-transr-v1/best/ 2>/dev/null |
|
|
