def sample_2():
# create random input and output data
x = Variable(X.copy())
y = Variable(Y.copy())
# randomly initialize weights
w1 = Variable(W1.copy())
w2 = Variable(W2.copy())
for t in range(EPOCHS):
# forward pass: compute predicted y
h = F.matmul(x, w1)
h_relu = F.relu(h)
y_pred = F.matmul(h_relu, w2)
# compute and print loss
loss = F.mean_squared_error(y_pred, y)
print(loss.data)
# manually zero the gradients
w1.zerograd()
w2.zerograd()
# backward pass
# loss.grad = np.ones(loss.shape, dtype=np.float32)
loss.backward()
# update weights
w1.data -= LEARNING_RATE * w1.grad
w2.data -= LEARNING_RATE * w2.grad
def sample_2():
# create random input and output data
x = Variable(np.random.randn(DATA_SIZE, N_I).astype(np.float32))
y = Variable(np.random.randn(DATA_SIZE, N_O).astype(np.float32))
# randomly initialize weights
w1 = Variable(np.random.randn(N_I, HIDDEN_SIZE).astype(np.float32))
w2 = Variable(np.random.randn(HIDDEN_SIZE, N_O).astype(np.float32))
for t in range(EPOCHS):
# forward pass: compute predicted y
h = F.matmul(x, w1)
h_r = F.relu(h)
y_p = F.matmul(h_r, w2)
# compute and print loss
loss = F.mean_squared_error(y_p, y)
print(loss.data)
# manually zero the gradients
w1.zerograd()
w2.zerograd()
# backward pass
# loss.grad = np.ones(loss.shape, dtype=np.float32)
loss.backward()
# update weights
w1.data -= LEARNING_RATE * w1.grad
w2.data -= LEARNING_RATE * w2.grad
def _apply_backward(self, x, grid, grads, use_cudnn):
x = Variable(x)
grid = Variable(grid)
y = functions.spatial_transformer_sampler(x, grid, use_cudnn=use_cudnn)
x.zerograd()
grid.zerograd()
y.grad = grads
y.backward()
return x, grid, y
def _apply_backward(self, x, grid, grads, use_cudnn):
x = Variable(x)
grid = Variable(grid)
y = functions.spatial_transformer_sampler(
x, grid, use_cudnn=use_cudnn)
x.zerograd()
grid.zerograd()
y.grad = grads
y.backward()
return x, grid, y
def update_step(net, images, step_size=1.5, end='inception_4c/output', jitter=32, clip=True):
offset_x, offset_y = np.random.randint(-jitter, jitter + 1, 2)
data = np.roll(np.roll(images, offset_x, -1), offset_y, -2)
x = Variable(xp.asarray(data))
x.zerograd()
dest, = net(x, outputs=[end])
objective(dest).backward()
g = cuda.to_cpu(x.grad)
data[:] += step_size / np.abs(g).mean() * g
data = np.roll(np.roll(data, -offset_x, -1), -offset_y, -2)
if clip:
bias = net.mean.reshape((1, 3, 1, 1))
data[:] = np.clip(data, -bias, 255 - bias)
return data
# create random input and output data
x = Variable(X)
y = Variable(Y)
# randomly initialize weights
w1 = Variable(W1)
w2 = Variable(W2)
for t in range(EPOCHS):
# forward pass: compute predicted y
h = F.matmul(x, w1)
h_r = F.relu(h)
y_p = F.matmul(h_r, w2)
# compute and print loss
loss = F.mean_squared_error(y_p, y)
print(loss.data)
# manually zero the gradients
w1.zerograd()
w2.zerograd()
# backward pass
# loss.grad = np.ones(loss.shape, dtype=np.float32)
loss.backward()
# update weights
w1.data -= LEARNING_RATE * w1.grad
w2.data -= LEARNING_RATE * w2.grad
for batch_indexes in np.array_split(perm, num_batches):
x_batch = x_train[batch_indexes]
t_batch = t_train[batch_indexes]
x = Variable(x_batch)
t = Variable(t_batch)
# 順伝播
a_z = F.linear(x, w_1, b_1)
z = F.tanh(a_z)
a_y = F.linear(z, w_2, b_2)
loss = F.softmax_cross_entropy(a_y, t)
# 逆伝播
w_1.zerograd()
w_2.zerograd()
b_1.zerograd()
b_2.zerograd()
loss.backward(retain_grad=True)
grad_w_1 = w_1.grad
grad_w_2 = w_2.grad
grad_b_1 = b_1.grad
grad_b_2 = b_2.grad
w_1.data = w_1.data - learning_rate * grad_w_1
w_2.data = w_2.data - learning_rate * grad_w_2
b_1.data = b_1.data - learning_rate * grad_b_1
b_2.data = b_2.data - learning_rate * grad_b_2
for batch_indexes in np.array_split(perm, num_batches):
x_batch = x_train[batch_indexes]
t_batch = t_train[batch_indexes]
x = Variable(x_batch)
t = Variable(t_batch)
# 順伝播
a_z = F.linear(x, w_1, b_1)
z = F.tanh(a_z)
a_y = F.linear(z, w_2, b_2)
loss = F.softmax_cross_entropy(a_y, t)
# 逆伝播
w_1.zerograd()
w_2.zerograd()
b_1.zerograd()
b_2.zerograd()
loss.backward(retain_grad=True)
grad_w_1 = w_1.grad
grad_w_2 = w_2.grad
grad_b_1 = b_1.grad
grad_b_2 = b_2.grad
w_1.data = w_1.data - learning_rate * grad_w_1
w_2.data = w_2.data - learning_rate * grad_w_2
b_1.data = b_1.data - learning_rate * grad_b_1
b_2.data = b_2.data - learning_rate * grad_b_2
