def test_mlp(self):
"""
Train MLP on dataset 1, which is not linearly separable. Tests will try multiple random initializations to
reduce chance of bad initialization.
"""
i = 1
num_hidden_units = [4, 5]
params = {
'max_iter': 400,
'squash_function': logistic,
'loss_function': nll,
'num_hidden_units': num_hidden_units,
'lambda': 0.01
}
input_dim = self.train_data[i].shape[0] + 1
total_weight_length = input_dim * num_hidden_units[0]
for j in range(len(num_hidden_units) - 1):
total_weight_length += num_hidden_units[j] * num_hidden_units[j +
1]
total_weight_length += num_hidden_units[-1]
# try at most 10 random initializations
for trial in range(10):
mlp_model = mlp_train(self.train_data[i], self.train_labels[i],
params)
predictions, _, _, _ = mlp_predict(self.test_data[i], mlp_model)
accuracy = np.mean(predictions == self.test_labels[i])
print(
"On trial %d, 3-layer MLP had test accuracy %2.3f "
"(should be around 0.95, depending on random initialization)" %
(trial, accuracy))
if accuracy > 0.9:
return
assert False, "Accuracy was never above 0.9. Could be a bug, or could be bad luck. " \
"Try running again to check."
def test_mlp(self):
"""
Train 3-layer MLP on datasets 0 to 9. Tests will try multiple random initialization trials to
reduce chance of bad initialization.
"""
threshold_accuracy = [0.95,0.9,0.85,0.9,0.9,0.875,0.9,0.875,0.85,0.85] # threshold accuracy values to beat for every dataset
pass_vec = np.zeros(10, dtype =bool) # stores whether test passed on a dataset or not
for i in range(10):
num_hidden_units = [4, 5]
params = {
'max_iter': 400,
'activation_function': logistic,
'loss_function': nll,
'num_hidden_units': num_hidden_units,
'lambda': 0.01
}
input_dim = self.train_data[i].shape[0] + 1
total_weight_length = input_dim * num_hidden_units[0]
for j in range(len(num_hidden_units) - 1):
total_weight_length += num_hidden_units[j] * num_hidden_units[j + 1]
total_weight_length += num_hidden_units[-1]
# try at most 10 random initializations
best_accuracy = 0
for trial in range(10):
mlp_model = mlp_train(self.train_data[i], self.train_labels[i], params)
predictions, _, _, _ = mlp_predict(self.test_data[i], mlp_model)
accuracy = np.mean(predictions == self.test_labels[i])
if accuracy > best_accuracy:
best_accuracy = accuracy
print("On dataset %d, 3-layer MLP had test accuracy %2.3f (should be greater than %2.3f)" %
(i, best_accuracy,threshold_accuracy[i]))
pass_vec = best_accuracy - threshold_accuracy[i] < 0
assert np.sum(pass_vec) == 0, "3-layer MLP accuracy was less than threshold for one of the datasets. Could be a bug, or could be bad luck. " \
"Try running again to check."