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_nonbinary_classify():
"""Test non-binary classification with different inputs"""
train = pd.read_csv(DATADIR + "/mnli/train.csv")
X_train = train[['text_a', 'text_b']]
y_train = train['label']
#X_train = list(X_train.values)
#y_train = list(y_train.values)
model = BertClassifier()
model.validation_fraction = 0.0
model.max_seq_length = 64
model.train_batch_size = 16
model.eval_batch_size = 8
model.epochs = 1
model.fit(X_train, y_train)
accy = model.score(X_train, y_train)
# pandas df input
X = X_train[:5]
print("testing %s input" % (type(X)))
y1 = model.predict(X)
# numpy array input
X = X_train[:5].values
print("testing %s input" % (type(X)))
y2 = model.predict(X)
assert list(y2) == list(y1)
# list input
X = list(X_train[:5].values)
print("testing %s input" % (type(X)))
y3 = model.predict(X)
assert list(y3) == list(y1)
## 去除噪声标签,获得训练数据
data_df_not_na_label = data_df_not_na[data_df_not_na['label'].isin(
['0', '1', '-1'])]
data_df_not_na_label['label'].value_counts()
train_df, dev_df = train_test_split(data_df_not_na_label,
test_size=0.2,
shuffle=True)
## 准备模型的数据
X_train, y_train = train_df['content'], train_df['label']
X_dev, y_dev = dev_df['content'], dev_df['label']
# define model
model = BertClassifier('bert-base-uncased')
model.validation_fraction = 0.0
model.learning_rate = 3e-5
model.gradient_accumulation_steps = 1
model.max_seq_length = 64
model.train_batch_size = 1
model.eval_batch_size = 1
model.epochs = 1
# fit
model.fit(X_train, y_train)
# score
accy = model.score(X_dev, y_dev)
test_df = pd.read_csv(
'data/nCov_10k_test.csv',