def eval_(self, test_loader, device):
"""Evaluate the model
Parameters
----------
test_loader: generator
Test data generator
device: str
cuda or cpu
Returns
-------
test_loss: Mean
Class object that collects the test loss
test_accuracy: Mean
Class object that collects the test accuracy
"""
# Test loss for this epoch
test_loss = Mean()
# Test accuracy for this epoch
test_accuracy = Mean()
for batch_x, batch_y, batch_classes in test_loader:
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
prediction = self(batch_x)
loss = F.cross_entropy(prediction, batch_y)
acc = self.accuracy_(prediction, batch_y)
test_loss.update(loss.item(), n=len(batch_x))
test_accuracy.update(acc.item(), n=len(batch_x))
return test_loss, test_accuracy
def train_(self, training_loader, device, optimizer):
"""Train the model
Parameters
----------
training_loader: generator
Training data generator
device: str
cuda or cpu
optimizer: callable
One of the PyTorch optimizers
Returns
-------
train_loss: Mean
Class object that collects the train loss
train_accuracy: Mean
Class object that collects the train accuracy
"""
# Train loss for this epoch
train_loss = Mean()
# Train accuracy for this epoch
train_accuracy = Mean()
for batch_x, batch_y, batch_classes in training_loader:
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
batch_classes1 = batch_classes[:, 0].to(device)
batch_classes2 = batch_classes[:, 1].to(device)
# Set gradients to zero and Compute gradients for the batch
optimizer.zero_grad()
# Calculate loss and accuracy
predict_digit1, predict_digit2, predict_y = self(batch_x)
# prediction size: 1000, 10;
# batch_y size: 1000
loss = F.cross_entropy(predict_y, batch_y) + \
F.cross_entropy(predict_digit1, batch_classes1) + \
F.cross_entropy(predict_digit2, batch_classes2)
acc = self.accuracy_(predict_y, batch_y)
# Backward propagation of gradients
loss.backward()
# Do an optimizer step (updating parameters)
optimizer.step()
# Store the statistics
train_loss.update(loss.item(), n=len(batch_x))
train_accuracy.update(acc.item(), n=len(batch_x))
return train_loss, train_accuracy
def eval_(self, test_loader, device):
"""Evaluate the model
Parameters
----------
test_loader: generator
Test data generator
device: str
cuda or cpu
Returns
-------
test_loss: Mean
Class object that collects the test loss
test_accuracy: Mean
Class object that collects the test accuracy
"""
# Test loss for this epoch
test_loss = Mean()
# Test accuracy for this epoch
test_accuracy = Mean()
for batch_x, batch_y, batch_classes in test_loader:
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
batch_classes1 = batch_classes[:, 0].to(device)
batch_classes2 = batch_classes[:, 1].to(device)
# Calculate loss and accuracy
predict_digit1, predict_digit2, predict_y = self(batch_x)
loss = F.cross_entropy(predict_y, batch_y) + \
F.cross_entropy(predict_digit1, batch_classes1) + \
F.cross_entropy(predict_digit2, batch_classes2)
acc = self.accuracy_(predict_y, batch_y)
# Store the statistics
test_loss.update(loss.item(), n=len(batch_x))
test_accuracy.update(acc.item(), n=len(batch_x))
return test_loss, test_accuracy