def persist(self, filepath):
checkpoint_callback = self.callbacks_config.get('model_checkpoint')
if checkpoint_callback:
checkpoint_filepath = checkpoint_callback['filepath']
if os.path.exists(checkpoint_filepath):
shutil.copyfile(checkpoint_filepath, filepath)
else:
persist_torch_model(self.model, filepath)
else:
persist_torch_model(self.model, filepath)
def on_epoch_end(self, *args, **kwargs):
if self.epoch_every and ((self.epoch_id % self.epoch_every) == 0):
self.model.eval()
val_loss = self.get_validation_loss()
loss_sum = val_loss[self.metric_name]
loss_sum = loss_sum.data.cpu().numpy()[0]
self.model.train()
if self.best_score is None:
self.best_score = loss_sum
if (self.minimize and loss_sum < self.best_score) or (not self.minimize and loss_sum > self.best_score) or (
self.epoch_id == 0):
self.best_score = loss_sum
persist_torch_model(self.model, self.filepath)
logger.info('epoch {0} model saved to {1}'.format(self.epoch_id, self.filepath))
self.epoch_id += 1
def on_epoch_end(self, *args, **kwargs):
if self.epoch_every and ((self.epoch_id % self.epoch_every) == 0):
self.model.eval()
val_loss = self.get_validation_loss()
loss_sum = val_loss[self.metric_name]
loss_sum = loss_sum.data.cpu().numpy()[0]
self.model.train()
if self.best_score is None:
self.best_score = loss_sum
if (self.minimize and loss_sum < self.best_score) or (not self.minimize and loss_sum > self.best_score) or (
self.epoch_id == 0):
self.best_score = loss_sum
persist_torch_model(self.model, self.filepath)
logger.info('epoch {0} model saved to {1}'.format(self.epoch_id, self.filepath))
self.epoch_id += 1
