def load_base_model_if_needed(learner: Learner,
lm_training_config: LMTrainingConfig,
model_file='best') -> None:
if lm_training_config.base_model:
model = os.path.join(lm_training_config.base_model, model_file)
logger.info(f"Using pretrained model: {model}.pth")
# not setting purge to True raises a pickle serialization error
learner.load(model, purge=False)
else:
logger.info("Training form scratch")
bert_train.iloc[val_idx, :],
bert_test,
tokenizer=fastai_tokenizer,
vocab=fastai_bert_vocab,
include_bos=False,
include_eos=False,
text_cols='comment_text',
label_cols=label_cols,
bs=BATCH_SIZE,
collate_fn=partial(pad_collate, pad_first=False, pad_idx=0),
)
learner = Learner(databunch, bert_model, loss_func=bert_custom_loss)
if CUR_STEP != 1:
learner.load('/kaggle/input/freeze-bert-1-s-uc-260ml-3e-8f-s-' +
str(CUR_STEP - 1) + '-f-' + str(MAKE_FOLD) +
'/models/' + FILE_NAME)
learner.fit_one_cycle(N_EPOCH, max_lr=MAX_LR)
oof[val_idx] = get_preds_as_nparray(DatasetType.Valid).astype(
np.float32)
predictions += get_preds_as_nparray(DatasetType.Test).astype(
np.float32) / NFOLDS
validate_df(train.iloc[val_idx], oof[val_idx, 0], verbose=True)
learner.save(FILE_NAME)
print('CV BIASED AUC:')
validate_df(train, oof[:, 0], verbose=True)