def _specific_train_val(self, balanced_train, balanced_val, neural_net, optimizer, current_cm, fold=0):
train_time_s = 0
val_time_s = 0
all_epoch_train_acc = []
all_epoch_val_acc = []
all_epoch_train_loss = []
all_epoch_val_loss = []
# self.__msd_train_dict = self.compute_label_msd_dict(balanced_train, fold+1)
for epoch in range(1, self.learning_def.num_epochs + 1):
train_start = time.time()
train_loss, train_accuracy = self.train(balanced_train, neural_net, optimizer)
train_time_diff = utils.time_diff(train_start)
train_time_s += train_time_diff
# after each epoch the new trained model
torch.save(neural_net, self.model_path)
# start validating the learning
val_start = time.time()
val_loss, val_accuracy = self.val(balanced_val, current_cm)
val_time_diff = utils.time_diff(val_start)
val_time_s += val_time_diff
all_epoch_train_acc.append(train_accuracy)
all_epoch_val_acc.append(val_accuracy)
all_epoch_train_loss.append(train_loss)
all_epoch_val_loss.append(val_loss)
# Print epoch number, loss, accuracy, name and guess
print_every = 10
if epoch % print_every == 0:
self.result_logger.info(
f"Epoch {epoch}: Train Loss: {train_loss :.5f} Train Accuracy:"
f" {train_accuracy:.3f} Val Loss: {val_loss :.5f} Val Accuracy: {val_accuracy:.3f}")
train_time = utils.time_s_to_str(train_time_s)
val_time = utils.time_s_to_str(val_time_s)
self.result_logger.info(f"\nTrain time (over last cross-validation pass): {train_time}")
self.result_logger.info(f"Test time (over last cross-validation pass): {val_time}")
num_last_epochs = 50
# calculate averages over the last 50 epochs
avg_train_loss = sum(all_epoch_train_loss[-num_last_epochs:]) / len(all_epoch_train_loss[-num_last_epochs:])
avg_train_accuracy = sum(all_epoch_train_acc[-num_last_epochs:]) / len(all_epoch_train_acc[-num_last_epochs:])
avg_val_loss = sum(all_epoch_val_loss[-num_last_epochs:]) / len(all_epoch_val_loss[-num_last_epochs:])
avg_val_accuracy = sum(all_epoch_val_acc[-num_last_epochs:]) / len(all_epoch_val_acc[-num_last_epochs:])
return avg_train_loss, avg_train_accuracy, avg_val_loss, avg_val_accuracy
def _specific_eval_only(self, balanced_val, confusion_matrix, model_path=None):
val_time_s = 0
# start validating the learning
val_start = time.time()
val_loss, val_accuracy = self.val(balanced_val, confusion_matrix, model_path)
val_time_diff = utils.time_diff(val_start)
val_time_s += val_time_diff
val_time = utils.time_s_to_str(val_time_s)
self.result_logger.info(f"Test Avg Loss: {val_loss:.5f} Test Accuracy: {val_accuracy:.3f}")
self.result_logger.info(f"\nTest time (over last cross-validation pass): {val_time}\n")
return val_accuracy
def _specific_train_only(self, balanced_train, neural_net, optimizer):
epoch_train_row_losses = []
epoch_train_row_accuracies = []
epoch_train_column_losses = []
epoch_train_column_accuracies = []
epoch_train_button_accuracies = []
train_time_s = 0
print(f"\nNetwork Architecture: {neural_net}\n")
for epoch in range(1, self.learning_def.num_epochs + 1):
train_start = time.time()
row_loss, row_accuracy, column_loss, column_accuracy, button_accuracy = self.train(
balanced_train, neural_net, optimizer)
train_time_diff = utils.time_diff(train_start)
train_time_s += train_time_diff
print_every = 10
if epoch % print_every == 0:
print(
f"Epoch {epoch}\n: Train Row Loss: {(row_loss / epoch):.5f} Train Row Accuracy:"
f" {row_accuracy:.3f} Train Column Loss: {(column_loss / epoch):.5f}"
f" Train Row Accuracy: {column_accuracy:.3f} Train Button Accuracy: {button_accuracy:.3f}\n"
)
# Add current loss avg to list of losses
epoch_train_row_losses.append(row_loss / epoch)
epoch_train_row_accuracies.append(row_accuracy)
epoch_train_column_losses.append(column_loss / epoch)
epoch_train_column_accuracies.append(column_accuracy / epoch)
epoch_train_button_accuracies.append(button_accuracy)
self.__all_epoch_train_row_accuracies.append(
epoch_train_row_accuracies)
self.__all_epoch_train_row_losses.append(epoch_train_row_losses)
self.__all_epoch_train_col_accuracies.append(
epoch_train_column_accuracies)
self.__all_epoch_train_col_losses.append(epoch_train_column_losses)
self.__all_epoch_train_button_accuracies.append(
epoch_train_button_accuracies)
self.train_time = utils.time_s_to_str(train_time_s)
def _specific_train_only(self, balanced_train, neural_net, optimizer):
train_time_s = 0
for epoch in range(1, self.learning_def.num_epochs + 1):
train_start = time.time()
current_train_loss, current_train_accuracy = self.train(balanced_train, neural_net, optimizer)
train_time_diff = utils.time_diff(train_start)
train_time_s += train_time_diff
# Print epoch number, loss, accuracy, name and guess
print_every = 10
if epoch % print_every == 0:
self.result_logger.info(
f"Epoch {epoch}: Train Loss: {(current_train_loss / epoch):.5f} Train Accuracy:"
f" {current_train_accuracy:.3f}")
# self.writer.add_scalar('Train Avg Loss', current_train_loss / epoch, epoch)
train_time = utils.time_s_to_str(train_time_s)
self.result_logger.info(f"\nTrain time (over last cross-validation pass): {train_time}")
def _specific_train_val(self, balanced_train, balanced_val, neural_net,
optimizer):
epoch_train_row_losses = []
epoch_train_row_accuracies = []
epoch_train_column_losses = []
epoch_train_column_accuracies = []
epoch_train_button_accuracies = []
epoch_val_row_losses = []
epoch_val_row_accuracies = []
epoch_val_column_losses = []
epoch_val_column_accuracies = []
epoch_val_button_accuracies = []
train_time_s = 0
val_time_s = 0
for epoch in range(1, self.learning_def.num_epochs + 1):
train_start = time.time()
train_row_loss, train_row_accuracy, train_column_loss, \
train_column_accuracy, train_button_accuracy = self.train(balanced_train, neural_net, optimizer)
train_time_diff = utils.time_diff(train_start)
train_time_s += train_time_diff
# start validating the learning
val_start = time.time()
val_row_loss, val_row_accuracy, val_column_loss, val_column_accuracy, val_button_accuracy = \
self.val(balanced_val)
val_time_diff = utils.time_diff(val_start)
val_time_s += val_time_diff
# Print epoch number, loss, accuracy, name and guess
print_every = 10
if epoch % print_every == 0:
print(
f"Epoch {epoch}: \n"
f"Train Row Loss: {(train_row_loss / epoch):.5f} Train Row Accuracy:"
f" {train_row_accuracy:.3f} Train Column Loss: {(train_column_loss / epoch):.5f} Train Column "
f"Accuracy: {train_column_accuracy:.3f} Train Button Accuracy: {train_button_accuracy:.3f} \n"
f"Val Row Loss: {(val_row_loss / epoch):.5f} Val Row Accuracy: {val_row_accuracy: .3f} "
f"Val Column Loss: {(val_column_loss / epoch):.5f} Val Column Accuracy: "
f"{val_column_accuracy:.3f} Val Button Accuracy: {val_button_accuracy:.3f} \n"
)
# print(f"Train Row Accuracies: {self._all_epoch_train_row_accuracies}\n")
# print(f"Train Button Accuracies: {self._all_epoch_train_button_accuracies}\n")
# print(f"Val Row Accuracies: {self._all_epoch_val_row_accuracies}\n")
# print(f"Val Button Accuracies: {self._all_epoch_val_button_accuracies}\n\n")
# Add current loss avg to list of losses
epoch_train_row_losses.append(train_row_loss / epoch)
epoch_val_row_losses.append(val_row_loss / epoch)
epoch_train_column_losses.append(train_column_loss / epoch)
epoch_val_column_losses.append(val_column_loss / epoch)
epoch_train_row_accuracies.append(train_row_accuracy)
epoch_val_row_accuracies.append(val_row_accuracy)
epoch_train_column_accuracies.append(train_column_accuracy)
epoch_val_column_accuracies.append(val_column_accuracy)
epoch_train_button_accuracies.append(train_button_accuracy)
epoch_val_button_accuracies.append(val_button_accuracy)
self.__all_epoch_train_row_accuracies.append(
epoch_train_row_accuracies)
self.__all_epoch_train_row_losses.append(epoch_train_row_losses)
self.__all_epoch_train_col_accuracies.append(
epoch_train_column_accuracies)
self.__all_epoch_train_col_losses.append(epoch_train_column_losses)
self.__all_epoch_train_button_accuracies.append(
epoch_train_button_accuracies)
self.__all_epoch_val_row_accuracies.append(epoch_val_row_accuracies)
self.__all_epoch_val_row_losses.append(epoch_val_row_losses)
self.__all_epoch_val_col_accuracies.append(epoch_val_column_accuracies)
self.__all_epoch_val_col_losses.append(epoch_val_column_losses)
self.__all_epoch_val_button_accuracies.append(
epoch_val_button_accuracies)
self.train_time = utils.time_s_to_str(train_time_s)
self.val_time = utils.time_s_to_str(val_time_s)