def check_regression(tracker):
import models
model = models.RegressionModel(1)
dataset = backend.RegressionDataset(model)
detected_parameters = None
for batch_size in (1, 2, 4):
inp_x = nn.Constant(dataset.x[:batch_size])
inp_y = nn.Constant(dataset.y[:batch_size])
output_node = model.run(inp_x)
verify_node(output_node, 'node', (batch_size, 1),
"RegressionModel.run()")
trace = trace_node(output_node)
assert inp_x in trace, "Node returned from RegressionModel.run() does not depend on the provided input (x)"
if detected_parameters is None:
detected_parameters = [
node for node in trace if isinstance(node, nn.Parameter)
]
for node in trace:
assert not isinstance(
node, nn.Parameter
) or node in detected_parameters, (
"Calling RegressionModel.run() multiple times should always re-use the same parameters, but a new nn.Parameter object was detected"
)
for batch_size in (1, 2, 4):
inp_x = nn.Constant(dataset.x[:batch_size])
inp_y = nn.Constant(dataset.y[:batch_size])
loss_node = model.get_loss(inp_x, inp_y)
verify_node(loss_node, 'loss', None, "RegressionModel.get_loss()")
trace = trace_node(loss_node)
assert inp_x in trace, "Node returned from RegressionModel.get_loss() does not depend on the provided input (x)"
assert inp_y in trace, "Node returned from RegressionModel.get_loss() does not depend on the provided labels (y)"
for node in trace:
assert not isinstance(
node, nn.Parameter
) or node in detected_parameters, (
"RegressionModel.get_loss() should not use additional parameters not used by RegressionModel.run()"
)
tracker.add_points(2) # Partial credit for passing sanity checks
model.train(dataset)
backend.maybe_sleep_and_close(1)
train_loss = model.get_loss(nn.Constant(dataset.x), nn.Constant(dataset.y))
verify_node(train_loss, 'loss', None, "RegressionModel.get_loss()")
train_loss = nn.as_scalar(train_loss)
# Re-compute the loss ourselves: otherwise get_loss() could be hard-coded
# to always return zero
train_predicted = model.run(nn.Constant(dataset.x))
verify_node(train_predicted, 'node', (dataset.x.shape[0], 1),
"RegressionModel.run()")
sanity_loss = 0.5 * np.mean((train_predicted.data - dataset.y)**2)
assert np.isclose(
train_loss,
sanity_loss), ("RegressionModel.get_loss() returned a loss of {:.4f}, "
"but the autograder computed a loss of {:.4f} "
"based on the output of RegressionModel.run()".format(
train_loss, sanity_loss))
loss_threshold = 0.15
if train_loss <= loss_threshold:
print("Your final loss is: {:f}".format(train_loss))
tracker.add_points(4)
else:
print(
"Your final loss ({:f}) must be no more than {:.4f} to receive full points for this question"
.format(train_loss, loss_threshold))
#print(list(dataset.x.flatten()))
b0, b1 = model.closedFormSolution(list(dataset.x.flatten()),
list(dataset.y.flatten()))
x = dataset.x.flatten()
y = dataset.y.flatten()
A = np.vstack([x, np.ones(len(x))]).T
m, c = np.linalg.lstsq(A, y, rcond=None)[0]
if (np.isclose(b0, c) and np.isclose(b1, m)):
tracker.add_points(2)
def check_regression(tracker):
import models
model = models.RegressionModel()
dataset = backend.RegressionDataset(model)
detected_parameters = None
for batch_size in (1, 2, 4):
inp_x = nn.Constant(dataset.x[:batch_size])
inp_y = nn.Constant(dataset.y[:batch_size])
output_node = model.run(inp_x)
verify_node(output_node, 'node', (batch_size, 1), "RegressionModel.run()")
trace = trace_node(output_node)
assert inp_x in trace, "Node returned from RegressionModel.run() does not depend on the provided input (x)"
if detected_parameters is None:
detected_parameters = [node for node in trace if isinstance(node, nn.Parameter)]
for node in trace:
assert not isinstance(node, nn.Parameter) or node in detected_parameters, (
"Calling RegressionModel.run() multiple times should always re-use the same parameters, but a new nn.Parameter object was detected")
for batch_size in (1, 2, 4):
inp_x = nn.Constant(dataset.x[:batch_size])
inp_y = nn.Constant(dataset.y[:batch_size])
loss_node = model.get_loss(inp_x, inp_y)
verify_node(loss_node, 'loss', None, "RegressionModel.get_loss()")
trace = trace_node(loss_node)
assert inp_x in trace, "Node returned from RegressionModel.get_loss() does not depend on the provided input (x)"
assert inp_y in trace, "Node returned from RegressionModel.get_loss() does not depend on the provided labels (y)"
for node in trace:
assert not isinstance(node, nn.Parameter) or node in detected_parameters, (
"RegressionModel.get_loss() should not use additional parameters not used by RegressionModel.run()")
tracker.add_points(2) # Partial credit for passing sanity checks
time_out = False
if platform == "linux" or platform == "linux2" or platform == "darwin":
# https://stackoverflow.com/questions/492519/timeout-on-a-function-call
# linux
# OS X
# print("Using MacOS")
signal.signal(signal.SIGALRM, handler)
signal.alarm(240)
try:
model.train(dataset)
except Exception as exc:
print(exc)
time_out = True
signal.alarm(0)
elif platform == "win32":
# Windows...
print("Using Windows no automatic timeout included")
model.train(dataset)
if time_out == False:
model.train(dataset)
backend.maybe_sleep_and_close(1)
train_loss = model.get_loss(nn.Constant(dataset.x), nn.Constant(dataset.y))
verify_node(train_loss, 'loss', None, "RegressionModel.get_loss()")
train_loss = nn.as_scalar(train_loss)
# Re-compute the loss ourselves: otherwise get_loss() could be hard-coded
# to always return zero
train_predicted = model.run(nn.Constant(dataset.x))
verify_node(train_predicted, 'node', (dataset.x.shape[0], 1), "RegressionModel.run()")
sanity_loss = 0.5 * np.mean((train_predicted.data - dataset.y)**2)
assert np.isclose(train_loss, sanity_loss), (
"RegressionModel.get_loss() returned a loss of {:.4f}, "
"but the autograder computed a loss of {:.4f} "
"based on the output of RegressionModel.run()".format(
train_loss, sanity_loss))
loss_threshold = 0.02
if train_loss <= loss_threshold:
print("Your final loss is: {:f}".format(train_loss))
tracker.add_points(4)
else:
print("Your final loss ({:f}) must be no more than {:.4f} to receive full points for this question".format(train_loss, loss_threshold))
else:
print("Your training timed out, therefore no final loss to report and test.")
