w4 = net.variable(guass(0., std, (1024, 10)))
b4 = net.variable(np.ones((10, )) * .1)
conv1 = net.conv2d(image, k1, pad=(2, 2), stride=(1, 1))
conv1 = net.batch_norm(conv1, net.variable(guass(0., std, (32, ))),
is_training)
conv1 = net.plus_b(conv1, b1)
conv1 = net.relu(conv1)
pool1 = net.maxpool2(conv1)
conv2 = net.conv2d(pool1, k2, (2, 2), (1, 1))
conv2 = net.plus_b(conv2, b2)
conv2 = net.relu(conv2)
pool2 = net.maxpool2(conv2)
flat = net.reshape(pool2, (7 * 7 * 64, ))
fc1 = net.plus_b(net.matmul(flat, w3), b3)
fc1 = net.relu(fc1)
fc2 = net.plus_b(net.matmul(fc1, w4), b4)
loss = net.softmax_crossent(fc2, label)
net.optimize(loss, 'adam', 1e-3)
mnist_data = read_mnist()
batch = 128
for count in range(5):
batch_num = int(mnist_data.train.num_examples / batch)
for i in range(batch_num):
feed, target = mnist_data.train.next_batch(batch)
feed = feed.reshape(batch, 28, 28, 1).astype(np.float64)
batch = 8
out_dim = inp_dim = 10
hid_dim = 300
mem_size = 128
vec_size = 30
max_seq_len = 20
start_symbol = np.zeros([batch, 1, inp_dim])
stop_symbol = np.zeros([batch, 1, inp_dim])
start_symbol[:, 0, 0] = np.ones([batch])
stop_symbol[:, 0, 1] = np.ones([batch])
net = Net()
x = net.portal((inp_dim, ))
y = net.portal((inp_dim, ))
y = net.reshape(y, [-1, inp_dim], over_batch=True)
ntm_out = net.turing(x, out_dim, mem_size, vec_size, hid_dim, shift=1)
start, end = net.portal(), net.portal()
copy = net.dynamic_slice(ntm_out, start=start, end=end, axis=0)
logits = net.reshape(copy, [-1, inp_dim], over_batch=True)
loss = net.logistic(logits, y)
net.optimize(loss, 'adam', 1e-2)
def generate_random_input(batch, seq_length, inp_dim):
x = np.random.rand(batch, seq_length, inp_dim).round()
x[:, :, :2] = np.zeros(x[:, :, :2].shape)
return x, np.zeros(x.shape)