def test_regular_encode(self):
"""Test that encoding is done properly."""
test_input = ["this is a test", "so is this"]
len_encoding = 20
encoded_input = regular_encode(test_input, self.test_tokenizer, len_encoding)
expected_encoded_input = np.array([[0, 9226, 16, 10, 1296, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
[0, 2527, 16, 42, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])
self.assertTrue((encoded_input == expected_encoded_input).all())
def make_predictions(df: pd.DataFrame,
model,
model_name: str,
max_len: int = 512,
method: str = "multiclass"):
"""
Make predictions using trained model and data to predict on.
:param df: Pandas DataFrame containing data to predict on
:param model: end-to-end trained Transformer model
:param model_name: name of model to be loaded by Transformer to get proper tokenizer
:param max_len: max length of string to be encoded
:param method: "multiclass" or "binary"--describes setting for prediction outputs
:return: Pandas DataFrame augmented with predictions made using trained model
"""
# First insert the CLS and SEP tokens
inputs = []
for i in range(len(df)):
# NOTE: this expects columns named "text1" and "text2" for the two claims
inputs.append(
str('[CLS]' + df.loc[i, 'text1'] + '[SEP]' + df.loc[i, 'text2']))
# Then make predictions
tokenizer = AutoTokenizer.from_pretrained(model_name)
encoded_inputs = regular_encode(inputs, tokenizer, maxlen=max_len)
predictions = model.predict(encoded_inputs)
if method == "multiclass":
# NEED TO CHECK THIS!!!
df['predicted_con'] = predictions[:, 2]
df['predicted_ent'] = predictions[:, 1]
df['predicted_neu'] = predictions[:, 0]
# Calculate predicted class as the max predicted label
df['predicted_class'] = df[[
'predicted_con', 'predicted_ent', 'predicted_neu'
]].idxmax(axis=1)
df.predicted_class.replace(to_replace={
'predicted_con': 'contradiction',
'predicted_ent': 'entailment',
'predicted_neu': 'neutral'
},
inplace=True)
elif method == "binary":
df.predicted_con = predictions[:, 0]
else:
raise ValueError(
f"{method} not a valid method type. Must be \"multiclass\" or \"binary\""
)
return df
def make_predictions(df: pd.DataFrame,
model,
model_name: str,
max_len: int = 512,
multi_class: bool = True):
"""
Make predictions using trained model and data to predict on.
:param df: Pandas DataFrame containing data to predict on
:param model: end-to-end trained Transformer model
:param model_name: name of model to be loaded by Transformer to get proper tokenizer
:param max_len: max length of string to be encoded
:param multi_class: "multiclass" or "binary"--describes setting for prediction outputs
:return: Pandas DataFrame augmented with predictions made using trained model
"""
# First insert the CLS and SEP tokens
inputs = []
# NOTE: this expects columns named "text1" and "text2" for the two claims
if multi_class:
for i in range(len(df)):
inputs.append(
str('[CLS]' + df.loc[i, 'text1'] + '[SEP]' +
df.loc[i, 'text2']))
else:
# Add the category info (CON, ENT, NEU) as auxillary text at the end
for i in range(len(df)):
inputs.append(
str('[CLS]' + df.loc[i, 'text1'] + '[SEP]' +
df.loc[i, 'text2'] + '[SEP]' + 'CON')) # noqa:
for i in range(len(df)):
inputs.append(
str('[CLS]' + df.loc[i, 'text1'] + '[SEP]' +
df.loc[i, 'text2'] + '[SEP]' + 'ENT')) # noqa: W503
for i in range(len(df)):
inputs.append(
str('[CLS]' + df.loc[i, 'text1'] + '[SEP]' +
df.loc[i, 'text2'] + '[SEP]' + 'NEU')) # noqa: W503
# Then make predictions
tokenizer = AutoTokenizer.from_pretrained(model_name)
encoded_inputs = regular_encode(inputs,
tokenizer,
maxlen=max_len,
multi_class=multi_class)
predictions = model.predict(encoded_inputs)
if multi_class:
df['predicted_con'] = predictions[:, 2]
df['predicted_ent'] = predictions[:, 1]
df['predicted_neu'] = predictions[:, 0]
else:
# Note: For the binary method using auxillary input, after retrieving the prediction probability
# for each class, we structure the prediction output dataframe in the same format
# as the multiclass method.
df['predicted_con'] = predictions[0:len(df)]
df['predicted_ent'] = predictions[len(df):(2 * len(df))]
df['predicted_neu'] = predictions[(2 * len(df)):]
# Calculate predicted class as the max predicted label
df['predicted_class'] = df[[
'predicted_con', 'predicted_ent', 'predicted_neu'
]].idxmax(axis=1)
df.predicted_class.replace(to_replace={
'predicted_con': 'contradiction',
'predicted_ent': 'entailment',
'predicted_neu': 'neutral'
},
inplace=True)
return df