def test_fit(self):
x = np.column_stack([np.ones(len(X_TRAIN)), X_TRAIN])
mlp = MultilayerPerceptron(num_inputs=4,
num_hidden_layers=1,
num_hidden_nodes=3)
mlp.fit(x, LABELS_TRAIN, epochnum=5, add_constant=False)
mlp.classify(x, add_constant=False)
def test_easy(self):
x = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([0, 1, 0, 1])
mlp = MultilayerPerceptron(
num_inputs=3, num_hidden_layers=1, num_hidden_nodes=6, seed=323440,
learn_rate = .8, learn_rate_evol='constant', momentum=.1
)
mlp.fit(x, y, add_constant=True, epochnum=100, verbose=True)
print(mlp.classify(x, add_constant=True))
assert np.allclose(mlp.classify(x, add_constant=True).reshape(len(y)),
y)
def test_easy(self):
x = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([0, 1, 0, 1])
mlp = MultilayerPerceptron(num_inputs=3,
num_hidden_layers=1,
num_hidden_nodes=6,
seed=323440,
learn_rate=.8,
learn_rate_evol='constant',
momentum=.1)
mlp.fit(x, y, add_constant=True, epochnum=100, verbose=True)
print(mlp.classify(x, add_constant=True))
assert np.allclose(
mlp.classify(x, add_constant=True).reshape(len(y)), y)
def test_easy_multidim_y(self):
x = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([[0, 1], [1, 0], [0, 1], [1, 0]])
mlp = MultilayerPerceptron(
num_inputs=3, num_outputs=2,
num_hidden_layers=1, num_hidden_nodes=6, seed=323440,
learn_rate = .8, learn_rate_evol='constant', momentum=.1
)
mlp.fit(x, y, epochnum=50)
results = mlp.classify(x, max_ind=True)
assert np.allclose(to_dummies(results), y)
def test_easy_multidim_y(self):
x = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([[0, 1], [1, 0], [0, 1], [1, 0]])
mlp = MultilayerPerceptron(num_inputs=3,
num_outputs=2,
num_hidden_layers=1,
num_hidden_nodes=6,
seed=323440,
learn_rate=.8,
learn_rate_evol='constant',
momentum=.1)
mlp.fit(x, y, epochnum=50)
results = mlp.classify(x, max_ind=True)
assert np.allclose(to_dummies(results), y)
# Setup estimator
mlp = MultilayerPerceptron(num_inputs=3,
num_outputs=1,
num_hidden_layers=1,
num_hidden_nodes=6,
learn_rate=1,
learn_rate_evol='sqrt',
momentum=.1,
seed=23456)
# Estimate and plot
start = time.perf_counter()
with PdfPages('ex_mlp_2d.pdf') as pdf:
for i in range(0, 25):
mlp.fit(x, y, epochnum=20, add_constant=True)
Z = mlp.classify(x_flatmesh, add_constant=True)
Z = Z.reshape(x0.shape)
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.contour(x0, x1, Z, cmap=plt.cm.Paired)
ax.scatter(x[:, 0], x[:, 1], c=y, cmap=plt.cm.Paired)
pdf.savefig(fig)
plt.close(fig)
if i % 5 == 4:
print(i + 1)
print('Time elapsed: {:6.3f}'.format(time.perf_counter() - start))
# Setup estimator
num_hidden_nodes = [101]
mlp = MultilayerPerceptron(
num_inputs=data['train'][0].shape[1]+1,
num_outputs=data['train'][2].shape[1],
num_hidden_layers=len(num_hidden_nodes), num_hidden_nodes=num_hidden_nodes,
learn_rate=.5, momentum=.1, seed=23456
)
# Estimate multilayer perceptron
start = time.perf_counter()
mlp.fit(data['train'][0], data['train'][2],
epochnum=10, verbose=1)
pred = mlp.classify(data['test'][0], max_ind=True)
print("Time: {:5.2f}, Error: {:5.4f}".format(
time.perf_counter() - start,
1 - np.mean(pred == data['test'][1])
))
# Visualize first hidden layer
fig1 = plt.figure(figsize=(10, 10))
for i in range(num_hidden_nodes[0]-1):
side = np.sqrt(num_hidden_nodes[0]-1)
ax = fig1.add_subplot(side, side, i+1)
ax.imshow(mlp.layers[0].weights[1:, i].reshape([28, 28]), cmap=cm.Greys_r)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
fig1.savefig('ex_mlp_mnist.pdf')
x0, x1 = np.meshgrid(np.linspace(x0_min, x0_max, 500),
np.linspace(x1_min, x1_max, 500))
x_flatmesh = np.column_stack([x0.ravel(), x1.ravel()])
# Setup estimator
mlp = MultilayerPerceptron(
num_inputs=3, num_outputs=1, num_hidden_layers=1, num_hidden_nodes=6,
learn_rate=1, learn_rate_evol='sqrt', momentum=.1, seed=23456
)
# Estimate and plot
start = time.perf_counter()
with PdfPages('ex_mlp_2d.pdf') as pdf:
for i in range(0, 25):
mlp.fit(x, y, epochnum=20, add_constant=True)
Z = mlp.classify(x_flatmesh, add_constant=True)
Z = Z.reshape(x0.shape)
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.contour(x0, x1, Z, cmap=plt.cm.Paired)
ax.scatter(x[:,0], x[:,1], c=y, cmap=plt.cm.Paired)
pdf.savefig(fig)
plt.close(fig)
if i % 5 == 4:
print(i+1)
print('Time elapsed: {:6.3f}'.format(time.perf_counter() - start))
data = pickle.load(f)
data = {key: (val[0], val[1], to_dummies(val[1])) for key, val in data.items()}
# Setup estimator
num_hidden_nodes = [101]
mlp = MultilayerPerceptron(num_inputs=data['train'][0].shape[1] + 1,
num_outputs=data['train'][2].shape[1],
num_hidden_layers=len(num_hidden_nodes),
num_hidden_nodes=num_hidden_nodes,
learn_rate=.5,
momentum=.1,
seed=23456)
# Estimate multilayer perceptron
start = time.perf_counter()
mlp.fit(data['train'][0], data['train'][2], epochnum=10, verbose=1)
pred = mlp.classify(data['test'][0], max_ind=True)
print("Time: {:5.2f}, Error: {:5.4f}".format(
time.perf_counter() - start, 1 - np.mean(pred == data['test'][1])))
# Visualize first hidden layer
fig1 = plt.figure(figsize=(10, 10))
for i in range(num_hidden_nodes[0] - 1):
side = np.sqrt(num_hidden_nodes[0] - 1)
ax = fig1.add_subplot(side, side, i + 1)
ax.imshow(mlp.layers[0].weights[1:, i].reshape([28, 28]), cmap=cm.Greys_r)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
fig1.savefig('ex_mlp_mnist.pdf')
def test_fit(self):
x = np.column_stack([np.ones(len(X_TRAIN)), X_TRAIN])
mlp = MultilayerPerceptron(
num_inputs=4, num_hidden_layers=1, num_hidden_nodes=3)
mlp.fit(x, LABELS_TRAIN, epochnum=5, add_constant=False)
mlp.classify(x, add_constant=False)
