def test_linear_xeloss_backward():
np.random.seed(11785)
mytorch_mlp = Sequential(Linear(10, 20))
mytorch_optimizer = SGD(mytorch_mlp.parameters())
mytorch_criterion = CrossEntropyLoss()
test_forward_backward(mytorch_mlp, mytorch_criterion=mytorch_criterion)
return True
def mnist(train_x, train_y, val_x, val_y):
"""Problem 3.1: Initialize objects and start training
You won't need to call this function yourself.
(Data is provided by autograder)
Args:
train_x (np.array): training data (55000, 784)
train_y (np.array): training labels (55000,)
val_x (np.array): validation data (5000, 784)
val_y (np.array): validation labels (5000,)
Returns:
val_accuracies (list(float)): List of accuracies per validation round
(num_epochs,)
"""
# TODO: Initialize an MLP, optimizer, and criterion
mnist_model = Sequential(Linear(784, 20), ReLU(), Linear(20, 10))
#mnist_model = Sequential(Linear(784,20),BatchNorm1d(20),ReLU(),Linear(20,10))
creterion = CrossEntropyLoss()
mdl_optimizer = SGD(mnist_model.parameters(), momentum=0.9, lr=0.1)
# TODO: Call training routine (make sure to write it below)
val_accuracies = train(mnist_model, mdl_optimizer, creterion, train_x,
train_y, val_x, val_y)
return val_accuracies
def mnist(train_x, train_y, val_x, val_y):
"""Problem 3.1: Initialize objects and start training
You won't need to call this function yourself.
(Data is provided by autograder)
Args:
train_x (np.array): training data (55000, 784)
train_y (np.array): training labels (55000,)
val_x (np.array): validation data (5000, 784)
val_y (np.array): validation labels (5000,)
Returns:
val_accuracies (list(float)): List of accuracies per validation round
(num_epochs,)
"""
# TODO: Initialize an MLP, optimizer, and criterion
model = sequential.Sequential(linear.Linear(784, 20), activations.ReLU(),
linear.Linear(20, 10))
optimizer = SGD(model.parameters(), lr=0.1)
criterion = loss.CrossEntropyLoss()
# TODO: Call training routine (make sure to write it below)
val_accuracies = train(model, optimizer, criterion, tensor.Tensor(train_x),
tensor.Tensor(train_y), tensor.Tensor(val_x),
tensor.Tensor(val_y))
return val_accuracies
def test_big_linear_relu_xeloss_momentum():
np.random.seed(11785)
mytorch_mlp = Sequential(Linear(10, 20), ReLU(), Linear(20, 30), ReLU())
mytorch_optimizer = SGD(mytorch_mlp.parameters(), momentum=0.9)
mytorch_criterion = CrossEntropyLoss()
test_step(mytorch_mlp,
mytorch_optimizer,
5,
5,
mytorch_criterion=mytorch_criterion)
return True
def test_big_linear_bn_relu_xeloss_train_eval():
np.random.seed(11785)
mytorch_mlp = Sequential(Linear(10, 20), BatchNorm1d(20), ReLU(),
Linear(20, 30), BatchNorm1d(30), ReLU())
mytorch_optimizer = SGD(mytorch_mlp.parameters())
mytorch_criterion = CrossEntropyLoss()
test_step(mytorch_mlp,
mytorch_optimizer,
5,
5,
mytorch_criterion=mytorch_criterion)
return True
def test_cnn_step():
np.random.seed(11785)
mytorch_cnn = CNN()
mytorch_optimizer = SGD(mytorch_cnn.parameters())
pytorch_cnn = nn.Sequential(nn.Conv1d(24, 56, 5, 1), nn.Tanh(),
nn.Conv1d(56, 28, 6, 2), nn.ReLU(),
nn.Conv1d(28, 14, 2, 2), nn.Sigmoid(),
nn.Flatten(), nn.Linear(13 * 14, 10))
assert len(pytorch_cnn) == len(
mytorch_cnn.layers.layers
), "Check number of modules in model; must be same as reference."
# check that all layers are the same
for idx, layer in enumerate(pytorch_cnn):
if isinstance(layer, nn.Conv1d):
assert isinstance(
mytorch_cnn.layers.layers[idx],
Conv1d), "Incorrect layer type at index " + str(idx)
if isinstance(layer, nn.Linear):
assert isinstance(
mytorch_cnn.layers.layers[idx],
Linear), "Incorrect layer type at index " + str(idx)
if isinstance(layer, nn.ReLU):
assert isinstance(
mytorch_cnn.layers.layers[idx],
ReLU), "Incorrect layer type at index " + str(idx)
if isinstance(layer, nn.Sigmoid):
assert isinstance(
mytorch_cnn.layers.layers[idx],
Sigmoid), "Incorrect layer type at index " + str(idx)
if isinstance(layer, nn.Flatten):
assert isinstance(
mytorch_cnn.layers.layers[idx],
Flatten), "Incorrect layer type at index " + str(idx)
if isinstance(layer, nn.Tanh):
assert isinstance(
mytorch_cnn.layers.layers[idx],
Tanh), "Incorrect layer type at index " + str(idx)
# Copy over weight data
if isinstance(layer, nn.Conv1d) or isinstance(layer, nn.Linear):
l = mytorch_cnn.layers.layers[idx]
layer.weight = nn.Parameter(torch.tensor(l.weight.data))
layer.bias = nn.Parameter(torch.tensor(l.bias.data))
return test_step(mytorch_cnn.layers,
mytorch_optimizer,
5,
pytorch_model=pytorch_cnn)
def test_linear_batchnorm_relu_train_eval():
np.random.seed(11785)
mytorch_mlp = Sequential(Linear(10, 20), BatchNorm1d(20), ReLU())
mytorch_optimizer = SGD(mytorch_mlp.parameters())
test_step(mytorch_mlp, mytorch_optimizer, 5, 5)
return True
def test_big_linear_relu_step():
np.random.seed(11785)
mytorch_mlp = Sequential(Linear(10, 20), ReLU(), Linear(20, 30), ReLU())
mytorch_optimizer = SGD(mytorch_mlp.parameters())
test_step(mytorch_mlp, mytorch_optimizer, 5, 5)
return True
def test_linear_momentum():
np.random.seed(11785)
mytorch_mlp = Sequential(Linear(10, 20), ReLU())
mytorch_optimizer = SGD(mytorch_mlp.parameters(), momentum=0.9)
test_step(mytorch_mlp, mytorch_optimizer, 5, 0)
return True
