def train():
mnn = MultilayerNN()
mnn.add(FullConnectLayer([784, 300], 0.003))
mnn.add(FullConnectLayer([300, 100], 0.003))
mnn.add(FullConnectLayer([100, 10], 0.003))
mnn.add(SoftmaxLayer())
from util.mnist import mnist_train_data, \
mnist_train_label, image_to_binary, \
label_to_one_hot, mnist_test_data, mnist_test_label
test_data = image_to_binary(mnist_test_data)
test_label = label_to_one_hot(mnist_test_label)
batch_size = 32
for i in range(1000):
for j in range(60000 // batch_size):
train_data = image_to_binary(mnist_train_data)
train_label = label_to_one_hot(mnist_train_label)
input_tensor = train_data[j * batch_size:(j + 1) * batch_size].reshape(-1, 1, 28, 28)
label_tensor = train_label[j * batch_size:(j + 1) * batch_size]
input_tensor = input_tensor.reshape([-1, 1, 28, 28])
mnn.train(input_tensor, label_tensor)
if j % 10 == 0:
pred = mnn.predict(test_data)
accuracy = np.mean(np.equal(np.argmax(pred, axis=1), np.argmax(test_label, axis=1)))
print(mnn.loss(), mnn.accuracy(), accuracy)
if j % 100 == 0:
mnn.save_model_as_json()
def train_from_exist():
mnn = MultilayerNN()
mnn.load_model_from_json("./model-1513582360.model")
from util.mnist import mnist_train_data, \
mnist_train_label, image_to_binary, \
label_to_one_hot, mnist_test_data, mnist_test_label
test_data = image_to_binary(mnist_test_data)
test_label = label_to_one_hot(mnist_test_label)
batch_size = 32
for i in range(1000):
for j in range(60000 // batch_size):
train_data = image_to_binary(mnist_train_data)
train_label = label_to_one_hot(mnist_train_label)
input_tensor = train_data[j * batch_size:(j + 1) * batch_size].reshape(-1, 1, 28, 28)
label_tensor = train_label[j * batch_size:(j + 1) * batch_size]
input_tensor = input_tensor.reshape([-1, 1, 28, 28])
mnn.train(input_tensor, label_tensor)
if j % 10 == 0:
pred = mnn.predict(test_data)
accuracy = np.mean(np.equal(np.argmax(pred, axis=1), np.argmax(test_label, axis=1)))
print(mnn.loss(), mnn.accuracy(), accuracy)
if j % 100 == 0:
mnn.save_model_as_json()
def mnist_test():
from util.mnist import mnist_train_data, \
mnist_train_label, mnist_test_data, mnist_test_label, image_to_binary, \
label_to_one_hot, one_hot_to_label
train_data = image_to_binary(mnist_train_data)
train_label = label_to_one_hot(mnist_train_label)
test_data = image_to_binary(mnist_test_data)
test_label = mnist_test_label
sm = Softmax(train_data, train_label)
sm.train()
pred = one_hot_to_label(sm.predict(test_data)).reshape(-1)
print(np.sum(pred == test_label))
def test():
mnn = MultilayerNN()
mnn.load_model_from_json("./model-1514368830.model")
from util.mnist import mnist_test_data, mnist_test_label, image_to_binary, \
label_to_one_hot
test_data = image_to_binary(mnist_test_data)
test_label = label_to_one_hot(mnist_test_label)
pred = mnn.predict(test_data)
accuracy = np.mean(np.equal(np.argmax(pred, axis=1), np.argmax(test_label, axis=1)))
print(accuracy)
from sklearn.neighbors import NearestNeighbors
from util.mnist import mnist_train_data, \
mnist_train_label, mnist_test_data, mnist_test_label, image_to_binary, \
label_to_one_hot, one_hot_to_label
nn = NearestNeighbors(n_neighbors=3,
metric="l2",
n_jobs=-1,
algorithm='kd_tree')
train_data = image_to_binary(mnist_train_data)
train_label = label_to_one_hot(mnist_train_label)
test_data = image_to_binary(mnist_test_data)
test_label = mnist_test_label
nn.fit(train_data)
print(nn.kneighbors(test_data))