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('StackOBERTflow-comments-small-v1')
# config.train_head_only = True
ds = TDDataset(config, binary=True)
tokenizer = tu.load_tokenizer(config)
model_cls = tu.get_model_cls(config)
train_dataloader = ds.get_complete_train_dataloader(tokenizer)
model = tu.load_model(config)
model.to(config.device)
util.set_seed(config)
experiment = Experiment(config, model, tokenizer)
global_step, tr_loss = experiment.train(train_dataloader)
experiment.save_model(util.models_path('satd_complete_binary'))
def main(config, results):
model_config = tu.load_model_config(config)
tokenizer = tu.load_tokenizer(config, model_config)
ds = Dataset(config, tokenizer)
label_names = ds.label_names
train_dataloader, valid_dataloader = ds.get_train_valid_dataloaders()
model = tu.load_model(config, model_config)
model.to(config.device)
util.set_seed(config)
experiment = Experiment(config,
model,
tokenizer,
label_names=label_names,
results=results)
global_step, tr_loss = experiment.train(train_dataloader,
valid_dataloader=valid_dataloader)
results = experiment.results
experiment.save_model(util.models_path('comment_code_shuffle'))
return results
def set_defaults(config):
config.train_epochs = 8
config.lr = 4e-5
config.max_seq_len = 255
config.train_bs = 32
config.eval_bs = 128 * 2
config.num_labels = 2
config.hidden_dropout_prob = 0.1
config.logging_steps = 50
config.model_path = util.models_path('StackOBERTflow-comments-small-v1')
def set_defaults(config):
config.train_epochs = 3
config.lr = 1e-5
# config.model_path = util.models_path('android.stackexchange.com')
# config.model_path = util.models_path('android.stackexchange.com')#a
# config.model_path = util.models_path('stackoverflow_1M')
# config.model_path = 'bert-base-uncased'
# config.model_path = util.models_path('StackOBERTflow-comments-small-v1/')
# config.model_path = 'bert-large-uncased'
config.max_seq_len = 255
config.model_type = None
config.model_path = util.models_path('StackOBERTflow-comments-small-v1')
config.train_bs = 32
config.eval_bs = 256
#config.hidden_dropout_prob = 0.07
config.num_labels = 12
config.hidden_dropout_prob = 0.1
config.multi_label = True
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)
def init_config(config, parse_args_func=None, experiment_parse_args_func=None, after_parse_func=None):
parser = argparse.ArgumentParser()
parser.add_argument("--lr", default=None, type=float, required=False)
parser.add_argument("--eval_bs", default=None, type=int, required=False)
parser.add_argument("--train_bs", default=None, type=int, required=False)
parser.add_argument("--max_steps", default=None, type=int, required=False)
parser.add_argument("--train_epochs", default=None, type=int, required=False)
parser.add_argument("--logging_steps", default=None, type=int, required=False)
parser.add_argument('--loss_label_weights', nargs='+', type=float, required=False)
parser.add_argument('--loss_func', default='cross_entropy', type=str, required=False)
parser.add_argument('--seed', type=int, default=42, required=False)
parser.add_argument('--seeds', nargs='+', type=int, required=False)
parser.add_argument('--reinit_layers', nargs='+', type=int, default=[], required=False)
parser.add_argument("--reinit_pooler", action="store_true")
parser.add_argument("--no_cuda", action="store_true")
parser.add_argument("--fp16", action="store_true")
parser.add_argument("--n_gpu", default=1, type=int)
parser.add_argument("--out_file", default=None, type=str, required=False)
parser.add_argument("--no_pretrain", action="store_true")
default_output_model_path = util.models_path("{}-{}".format(
list(filter(None, config.model_path.split("/"))).pop(),
time.strftime("%Y_%m_%d-%H_%M_%S")
))
parser.add_argument("--output_model_path",
default=default_output_model_path,
type=str,
required=False)
parser.add_argument(
"--model_path",
default=None,
type=str,
required=False,
help="Path to pre-trained model or shortcut name"
)
parser.add_argument(
"--tokenizer_path",
default=None,
type=str,
required=False
)
if parse_args_func:
parse_args_func(parser)
if experiment_parse_args_func:
experiment_parse_args_func(parser)
args = parser.parse_args()
for var, value in vars(args).items():
if value or not var in config:
config[var] = value
config['device'] = "cuda" if not config.no_cuda else "cpu"
if after_parse_func:
after_parse_func(config)