def main():
Xtrain, Ytrain = pvml.load_dataset("mnist_train")
Xtest, Ytest = pvml.load_dataset("mnist_test")
plt.ion()
batch_sz = 100
epochs = 250
network = pvml.MLP([Xtrain.shape[1], 128, 64, 10])
errors = [[], []]
for epoch in range(1, epochs + 1):
steps = Xtrain.shape[0] // batch_sz
network.train(Xtrain,
Ytrain,
lr=1e-4,
lambda_=1e-5,
steps=steps,
batch=batch_sz)
predictions = network.inference(Xtrain)[0]
training_error = (predictions != Ytrain).mean()
predictions = network.inference(Xtest)[0]
test_error = (predictions != Ytest).mean()
errors[0].append(100 * training_error)
errors[1].append(100 * test_error)
msg = "Epoch {}, Training err. {:.2f}, Test err. {:.2f}"
print(msg.format(epoch, 100 * training_error, 100 * test_error))
plot_errors(errors)
show_weights(network.weights)
show_confusion_matrix(Ytest, predictions)
plt.pause(0.05)
network.save("mnist_network.npz")
np.savetxt("mnist_errors.txt", np.array(errors).T, fmt="%.2f")
plt.ioff()
plt.show()
def test_saveload(self):
neurons = [2, 3, 4, 5]
net1 = pvml.MLP(neurons)
f = io.BytesIO()
net1.save(f)
f.seek(0)
net2 = pvml.MLP.load(f)
for w1, w2 in zip(net1.weights, net2.weights):
self.assertListEqual(list(w1.ravel()), list(w2.ravel()))
for b1, b2 in zip(net1.biases, net2.biases):
self.assertListEqual(list(b1), list(b2))
def train_step(self, X, Y, steps):
if self.net is None:
counts = [X.shape[1]] + self.hidden + [Y.max() + 1]
self.net = pvml.MLP(counts)
self.net.train(X,
Y,
lr=self.lr,
lambda_=self.lambda_,
momentum=self.momentum,
steps=steps,
batch=self.batch)
def test_batch(self):
X, Y = test_data.separable_circle_data_set(50, 2)
net = pvml.MLP([X.shape[1], 2, 2])
net.train(X, Y, lr=1e-1, steps=1000, batch=10)
Yhat, P = net.inference(X)
self.assertListEqual(Y.tolist(), Yhat.tolist())