def train(self,
trained_classifier_save_path,
learning_rates=np.array([1e-4, 1e-4, 1e-4, 1e-3, 1e-2]),
weight_decay=1e-6,
cycle_length=14):
self.learner.freeze_to(-1)
self.learner.fit(learning_rates,
n_cycle=1,
wds=weight_decay,
cycle_len=1,
use_clr=(8, 3))
self.learner.freeze_to(-2)
self.learner.fit(learning_rates,
n_cycle=1,
wds=weight_decay,
cycle_len=1,
use_clr=(8, 3))
self.learner.unfreeze()
self.learner.fit(learning_rates,
n_cycle=1,
wds=weight_decay,
cycle_len=cycle_length,
use_clr=(32, 10))
save_model(self.learner.model, trained_classifier_save_path)
def train(self, finetuned_language_model_encoder_save_path, learning_rate=1e-3, weight_decay=1e-7, cycle_length=15):
print('language model training starts')
print(f'Loaded torch version: {torch.__version__}')
self.learner.freeze_to(-1)
self.learner.fit(learning_rate / 2, n_cycle=1, wds=weight_decay, use_clr=(32, 2), cycle_len=1)
self.learner.unfreeze()
self.learner.fit(learning_rate, n_cycle=1, wds=weight_decay, use_clr=(20, 10), cycle_len=cycle_length)
save_model(self.learner.model[0], finetuned_language_model_encoder_save_path)
def train(self, trained_classifier_save_path, learning_rates=np.array([1e-4, 1e-4, 1e-4, 1e-3, 1e-2]),
weight_decay=1e-6, cycle_length=14):
print('Core classifier model training starts')
print(f'Loaded torch version: {torch.__version__}')
self.learner.freeze_to(-1)
self.learner.fit(learning_rates, n_cycle=1, wds=weight_decay, cycle_len=1, use_clr=(8, 3))
self.learner.freeze_to(-2)
self.learner.fit(learning_rates, n_cycle=1, wds=weight_decay, cycle_len=1, use_clr=(8, 3))
self.learner.unfreeze()
self.learner.fit(learning_rates, n_cycle=1, wds=weight_decay, cycle_len=cycle_length, use_clr=(32, 10))
save_model(self.learner.model, trained_classifier_save_path)
self.calculate_f1_for_validation_dataset()
def train(self,
finetuned_language_model_encoder_save_path,
learning_rate=1e-3,
weight_decay=1e-7,
cycle_length=15):
self.learner.freeze_to(-1)
self.learner.fit(learning_rate / 2,
n_cycle=1,
wds=weight_decay,
use_clr=(32, 2),
cycle_len=1)
self.learner.unfreeze()
self.learner.fit(learning_rate,
n_cycle=1,
wds=weight_decay,
use_clr=(20, 10),
cycle_len=cycle_length)
save_model(self.learner.model[0],
finetuned_language_model_encoder_save_path)