def get_dataloader(self, df, features_df, bs, include_df=False, **kwargs):
text_values = df.preprocessed.values
label_ids = df.label.values
if features_df is not None:
print(features_df.columns)
extra_features = features_df.drop('Readable', axis=1).to_numpy()
extra_features = StandardScaler().fit_transform(extra_features)
assert (extra_features.shape[1] == self.config.extra_features_size)
else:
extra_features = None
# dataloader = tu.get_dataloader(self.config, self.tokenizer, text_values, label_ids, bs, extra_features=extra_features, **kwargs)
if self.config.sep_token:
# leave text section empty
dataloader = tu.get_dataloader(self.config,
self.tokenizer,
[''] * len(text_values),
label_ids,
bs,
text_pair_values=text_values,
**kwargs)
else:
dataloader = tu.get_dataloader(self.config, self.tokenizer,
text_values, label_ids, bs,
**kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, df, bs, include_df=False, fake=False, **kwargs):
text_values = df.methodText.values
label_ids = df.cyclomatic.values
if fake:
label_ids = label_ids.copy()[:128]
text_values = text_values[:128]
np.random.shuffle(label_ids)
if self.config.sep_token:
# leave text section empty
dataloader = tu.get_dataloader(self.config, self.tokenizer, [''] * len(text_values), label_ids, bs, text_pair_values=text_values, **kwargs)
else:
dataloader = tu.get_dataloader(self.config, self.tokenizer, text_values, label_ids, bs, **kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, df, bs, include_df=False, **kwargs):
text_values = df.preprocessed
label_ids = df['class'].values
dataloader = tu.get_dataloader(self.config, self.tokenizer,
text_values, label_ids, bs, **kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, df, bs, include_df=False, **kwargs):
text_values = df.text.values
label_ids = df.informative.astype(int).values
dataloader = tu.get_dataloader(self.config, self.tokenizer,
text_values, label_ids, bs, **kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, df, bs, text_col='text', **kwargs):
text_values = df[text_col].values
if self.config.soft_label:
label_ids = df[self.LABEL_NAMES].to_numpy()
else:
label_ids = df.label.values
dataloader = tu.get_dataloader(self.config, self.tokenizer, text_values, label_ids, bs, **kwargs)
return dataloader
def get_dataloader(self, df, bs, include_df=False, **kwargs):
text_values = df.code.values
label_ids = df.problematic.astype(int).values
print(label_ids)
if self.config.sep_token:
# leave text section empty
dataloader = tu.get_dataloader(self.config,
self.tokenizer,
[''] * len(text_values),
label_ids,
bs,
text_pair_values=text_values,
**kwargs)
else:
dataloader = tu.get_dataloader(self.config, self.tokenizer,
text_values, label_ids, bs,
**kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, df, bs, include_df=False, **kwargs):
text_values = df.code.values
label_ids = df.complexity.map(
{label: idx
for idx, label in enumerate(self.LABEL_NAMES)}).values
if self.config.sep_token:
# leave text section empty
dataloader = tu.get_dataloader(self.config,
self.tokenizer,
[''] * len(text_values),
label_ids,
bs,
text_pair_values=text_values,
**kwargs)
else:
dataloader = tu.get_dataloader(self.config, self.tokenizer,
text_values, label_ids, bs,
**kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, tokenizer, df, bs, include_df=False, **kwargs):
text_values = df.preprocessed.values
single_class = self.config.single_class
if self.config.single_class:
label_ids = df[single_class].to_numpy()
else:
label_ids = df[Dataset.LABELS].to_numpy()
dataloader = tu.get_dataloader(self.config, tokenizer, text_values,
label_ids, bs, **kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, tokenizer, df, bs, include_df=False, **kwargs):
# convert_label = {'DEFECT': 0, 'DESIGN': 1,
# 'IMPLEMENTATION': 2, 'TEST': 3,
# 'WITHOUT_CLASSIFICATION': 4, 'DOCUMENTATION': 5}
text_values = df.preprocessed.values
label_ids = df.label.values
print(label_ids.dtype)
print(np.unique(label_ids))
dataloader = tu.get_dataloader(self.config, tokenizer, text_values,
label_ids, bs, **kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, df, bs, backtrans_langs=[], include_df=False, **kwargs):
text_values = df.preprocessed.values
label_ids = df.label_id.astype(int).values
if backtrans_langs:
logger.info("Text values array shape before augmentation: %s", text_values.shape)
logger.info("Label ids array shape before augmentation: %s", label_ids.shape)
for l in backtrans_langs:
text_values = np.append(text_values, df[f'preprocessed_{l}'].values, axis=0)
label_ids = np.tile(label_ids, len(backtrans_langs) + 1)
logger.info("Text values array shape AFTER augmentation: %s", text_values.shape)
logger.info("Label ids array shape AFTER augmentation: %s", label_ids.shape)
dataloader = tu.get_dataloader(self.config, self.tokenizer, text_values, label_ids, bs, **kwargs)
if include_df:
return (dataloader, df)
return dataloader
def get_dataloader(self, df, bs, include_df=False, **kwargs):
if self.config.sep_token:
text_values1 = df.preprocessed_comment.values
text_values2 = df.preprocessed_code.values
else:
text_values1 = df.preprocessed.values
text_values2 = None
label_ids = df.label.values
dataloader = tu.get_dataloader(self.config,
self.tokenizer,
text_values1,
label_ids,
bs,
text_pair_values=text_values2,
**kwargs)
if include_df:
return (dataloader, df)
return dataloader
def main():
config = get_config()
with config:
config.logging_steps = 400
config.train_epochs = 2
config.lr = 4e-5
# config.lr = 1e-4
config.model_type = 'roberta'
config.model_path = util.models_path('satd_complete_binary')
# config.train_head_only = True
tokenizer = tu.load_tokenizer(config)
model_cls = tu.get_model_cls(config)
df = pd.read_csv(util.data_path('satd', 'unclassified.csv'))
# df = pd.read_csv(util.data_path('satd', 'dataset.csv'))
df.dropna(inplace=True)
# df.rename(columns={'classification': 'orig_classification'}, inplace=True)
print(df.dtypes)
print(df.head())
df['preprocessed'] = df.commenttext.map(TDDataset.preprocess)
df.dropna(inplace=True)
# df = df.head(100)
preprocessed = df.preprocessed.values
dummy_labels = np.zeros(preprocessed.shape[0])
dataloader = tu.get_dataloader(config,
tokenizer,
preprocessed,
dummy_labels,
bs=128,
shuffle=False)
model = tu.load_model(config)
model.to(config.device)
util.set_seed(config)
experiment = Experiment(config, model, tokenizer)
preds = experiment.predict(dataloader)
preds = torch.from_numpy(preds)
probs = F.softmax(preds, dim=1)
uncertainty = least_conf(probs).numpy()
labels = np.argmax(preds, axis=1)
df['uncertainty'] = uncertainty
df['probs0'] = probs[:, 0].numpy()
df['probs1'] = probs[:, 1].numpy()
df['classification'] = labels
df.drop('preprocessed', axis='columns', inplace=True)
label_name_map = {i: l for i, l in enumerate(TDDataset.BINARY_LABEL_NAMES)}
print(label_name_map)
# convert_label = {'DEFECT': 1, 'DESIGN': 1,
# 'IMPLEMENTATION': 1, 'TEST': 1,
# 'WITHOUT_CLASSIFICATION': 0, 'DOCUMENTATION': 1}
# df['correct'] = (df.orig_classification.map(convert_label) == df.classification)
# print(df.correct.value_counts(normalize=True))
df.classification = df.classification.map(label_name_map)
df.to_csv(util.data_path('satd', 'unclassified_evaled.csv'), index=False)
tech_debt_df = df[df.classification == 'TECHNICAL_DEBT']
print(tech_debt_df.shape)
tech_debt_df.to_csv(util.data_path('satd', 'unclassified_pos.csv'),
index=False)
