def test_save_load_model():
"""Test saving/loading a fitted model to disk"""
X_train, y_train, X_dev, y_dev, label_list = toxic_test_data()
model = BertClassifier()
model.max_seq_length = 64
model.train_batch_size = 8
model.epochs = 1
model.multilabel = True
model.label_list = label_list
model.fit(X_train, y_train)
accy1 = model.score(X_dev, y_dev)
savefile = './test_model_save.bin'
print("\nSaving model to ", savefile)
model.save(savefile)
# load model from disk
new_model = load_model(savefile)
# predict with new model
accy2 = new_model.score(X_dev, y_dev)
# clean up
print("Cleaning up model file: test_model_save.bin ")
os.remove(savefile)
assert accy1 == accy2
def test_bert_sklearn_accy():
"""
Test bert_sklearn accuracy
compare against huggingface run_classifier.py
on 200 rows of SST-2 data.
"""
print("Running bert-sklearn...")
X_train, y_train, X_dev, y_dev, label_list = toxic_test_data()
# define model
model = BertClassifier()
model.validation_fraction = 0.0
model.learning_rate = 5e-5
model.gradient_accumulation_steps = 2
model.max_seq_length = 64
model.train_batch_size = 16
model.eval_batch_size = 8
model.epochs = 2
model.multilabel = True # for multi-label classification
model.label_list = label_list
model.fit(X_train, y_train)
bert_sklearn_accy = model.score(X_dev, y_dev)
bert_sklearn_accy /= 100
# run huggingface BERT run_classifier and check we get the same accuracy
cmd = r"python tests/run_classifier.py --task_name sst-2 \
--data_dir ./tests/data/sst2 \
--do_train --do_eval \
--output_dir ./comptest \
--bert_model bert-base-uncased \
--do_lower_case \
--learning_rate 5e-5 \
--gradient_accumulation_steps 2 \
--max_seq_length 64 \
--train_batch_size 16 \
--eval_batch_size 8 \
--num_train_epochs 2"
print("\nRunning huggingface run_classifier.py...\n")
os.system(cmd)
print("...finished run_classifier.py\n")
# parse run_classifier.py output file and find the accy
accy = open("comptest/eval_results.txt").read().split("\n")[
0] # 'acc = 0.76'
accy = accy.split("=")[1]
accy = float(accy)
print("bert_sklearn accy: %.02f, run_classifier.py accy : %0.02f" %
(bert_sklearn_accy, accy))
# clean up
print("\nCleaning up eval file: eval_results.txt")
#os.remove("eval_results.txt")
shutil.rmtree("comptest")
assert bert_sklearn_accy == accy
def test_not_fitted_exception():
"""Test predicting with a model that has not been fitted"""
X_train, y_train, X_dev, y_dev, label_list = toxic_test_data()
model = BertClassifier()
model.max_seq_length = 64
model.train_batch_size = 8
model.epochs = 1
model.multilabel = True
model.label_list = label_list
# model has not been fitted: model.fit(X_train, y_train)
with pytest.raises(Exception):
model.score(X_dev, y_dev)