def test_matrix_elementwise_multiply():
ctx = ndarray.gpu(0)
shape = (500, 200)
x = np.random.uniform(0, 10, size=shape).astype(np.float32)
y = np.random.uniform(0, 10, size=shape).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.array(y, ctx=ctx)
arr_z = ndarray.empty(shape, ctx=ctx)
gpu_op.matrix_elementwise_multiply(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(x * y, z, rtol=1e-5)
def test_relu_gradient():
shape = (2000, 2500)
ctx = ndarray.gpu(0)
x = np.random.uniform(-1, 1, shape).astype(np.float32)
grad_x = np.random.uniform(-5, 5, shape).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_grad_x = ndarray.array(grad_x, ctx=ctx)
arr_y = ndarray.empty(shape, ctx=ctx)
gpu_op.relu_gradient(arr_x, arr_grad_x, arr_y)
y = arr_y.asnumpy()
np.testing.assert_allclose(((x > 0) * grad_x).astype(np.float32), y)
def test_softmax_cross_entropy():
ctx = ndarray.gpu(0)
shape = (400, 1000)
y = np.random.uniform(-5, 5, shape).astype(np.float32)
y_ = np.random.uniform(-5, 5, shape).astype(np.float32)
arr_y = ndarray.array(y, ctx=ctx)
arr_y_ = ndarray.array(y_, ctx=ctx)
arr_out = ndarray.empty((1, ), ctx=ctx)
gpu_op.softmax_cross_entropy(arr_y, arr_y_, arr_out)
out = arr_out.asnumpy()
# numpy calculation
cross_entropy = np.mean(
-np.sum(y_ * np.log(autodiff.softmax_func(y)), axis=1), keepdims=True)
np.testing.assert_allclose(cross_entropy, out, rtol=1e-5)
def test_softmax():
ctx = ndarray.gpu(0)
shape = (400, 1000)
x = np.random.uniform(-5, 5, shape).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.empty(shape, ctx=ctx)
gpu_op.softmax(arr_x, arr_y)
y = arr_y.asnumpy()
np.testing.assert_allclose(autodiff.softmax_func(x), y, rtol=1e-5)
def test_relu():
shape = (2000, 2500)
ctx = ndarray.gpu(0)
x = np.random.uniform(-1, 1, shape).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.empty(shape, ctx=ctx)
gpu_op.relu(arr_x, arr_y)
y = arr_y.asnumpy()
np.testing.assert_allclose(np.maximum(x, 0).astype(np.float32), y)
def test_matrix_elementwise_multiply_by_const():
shape = (2000, 3000)
ctx = ndarray.gpu(0)
x = np.random.uniform(0, 10, size=shape).astype(np.float32)
val = np.random.uniform(-5, 5)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.empty(shape, ctx=ctx)
gpu_op.matrix_elementwise_multiply_by_const(arr_x, val, arr_y)
y = arr_y.asnumpy()
np.testing.assert_allclose(x * val, y, rtol=1e-5)
def test_broadcast_to():
ctx = ndarray.gpu(0)
shape = (200, 300)
to_shape = (130, 200, 300)
x = np.random.uniform(-1, 1, shape).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.empty(to_shape, ctx=ctx)
gpu_op.broadcast_to(arr_x, arr_y)
y = arr_y.asnumpy()
np.testing.assert_allclose(np.broadcast_to(x, to_shape), y)
def test_matrix_multiply():
ctx = ndarray.gpu(0)
x = np.random.uniform(0, 10, size=(500, 700)).astype(np.float32)
y = np.random.uniform(0, 10, size=(700, 1000)).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.array(y, ctx=ctx)
arr_z = ndarray.empty((500, 1000), ctx=ctx)
gpu_op.matrix_multiply(arr_x, False, arr_y, False, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(x, y), z, rtol=1e-5)
x = np.random.uniform(0, 10, size=(1000, 500)).astype(np.float32)
y = np.random.uniform(0, 10, size=(2000, 500)).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.array(y, ctx=ctx)
arr_z = ndarray.empty((1000, 2000), ctx=ctx)
gpu_op.matrix_multiply(arr_x, False, arr_y, True, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(x, np.transpose(y)), z, rtol=1e-5)
x = np.random.uniform(0, 10, size=(500, 1000)).astype(np.float32)
y = np.random.uniform(0, 10, size=(2000, 500)).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.array(y, ctx=ctx)
arr_z = ndarray.empty((1000, 2000), ctx=ctx)
gpu_op.matrix_multiply(arr_x, True, arr_y, True, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(np.transpose(x), np.transpose(y)),
z,
rtol=1e-5)
def test_convolution_forward():
ctx = ndarray.gpu(0)
in_shape = (1, 1, 8)
filter_shape = (1, 1, 5)
out_shape = (1, 1, 8)
input_arr_np = np.arange(8).reshape(in_shape)
dinput_arr_np = np.arange(8).reshape(in_shape)
filter_arr_np = np.arange(5).reshape(filter_shape)
dfilter_arr_np = np.arange(5).reshape(filter_shape)
dinput = ndarray.array(dinput_arr_np, ctx=ctx)
dfilter = ndarray.array(dfilter_arr_np, ctx=ctx)
arr_in = ndarray.array(input_arr_np, ctx=ctx)
arr_filter = ndarray.array(filter_arr_np, ctx=ctx)
arr_out = ndarray.empty(out_shape, ctx=ctx)
gpu_op.convolution_1d_forward(arr_in, arr_filter, arr_out, "NCHW", "SAME",
1)
gpu_op.convolution_1d_backward(arr_in, arr_out, arr_filter, dfilter,
dinput, "NCHW", "SAME", 1)
print(arr_out.asnumpy())
print(dfilter.asnumpy())
print(dinput.asnumpy())
print(arr_in.asnumpy())
def test_reduce_sum_axis_zero():
ctx = ndarray.gpu(0)
shape = (500, 200, 100)
to_shape = (200, 100)
x = np.random.uniform(0, 20, shape).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
arr_y = ndarray.empty(to_shape, ctx=ctx)
gpu_op.reduce_sum_axis_zero(arr_x, arr_y)
y = arr_y.asnumpy()
y_ = np.sum(x, axis=0)
for index, _ in np.ndenumerate(y):
v = y[index]
v_ = y_[index]
if abs((v - v_) / v_) > 1e-4:
print(index, v, v_)
np.testing.assert_allclose(np.sum(x, axis=0), y, rtol=1e-5)
def mnist_logreg(executor_ctx=None, num_epochs=10, print_loss_val_each_epoch=False):
print("Build logistic regression model...")
W1 = ad.Variable(name="W1")
b1 = ad.Variable(name="b1")
X = ad.Variable(name="X")
y_ = ad.Variable(name="y_")
z1 = ad.matmul_op(X, W1)
y = z1 + ad.broadcastto_op(b1, z1)
loss = ad.softmaxcrossentropy_op(y, y_)
grad_W1, grad_b1 = ad.gradients(loss, [W1, b1])
executor = ad.Executor([loss, grad_W1, grad_b1, y], ctx=executor_ctx)
# Read input data
datasets = load_mnist_data("mnist.pkl.gz")
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# Set up minibatch
batch_size = 1000
n_train_batches = train_set_x.shape[0] // batch_size
n_valid_batches = valid_set_x.shape[0] // batch_size
print("Start training loop...")
# Initialize parameters
W1_val = np.zeros((784, 10))
b1_val = np.zeros((10))
X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
valid_X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
valid_y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
if ndarray.is_gpu_ctx(executor_ctx):
W1_val = ndarray.array(W1_val, ctx=executor_ctx)
b1_val = ndarray.array(b1_val, ctx=executor_ctx)
X_val = ndarray.array(X_val, ctx=executor_ctx)
y_val = ndarray.array(y_val, ctx=executor_ctx)
lr = 1e-3
for i in range(num_epochs):
print("epoch %d" % i)
for minibatch_index in range(n_train_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
X_val[:] = train_set_x[minibatch_start:minibatch_end]
y_val[:] = convert_to_one_hot(
train_set_y[minibatch_start:minibatch_end])
loss_val, grad_W1_val, grad_b1_val, _ = executor.run(
feed_dict = {X: X_val, y_: y_val, W1: W1_val, b1: b1_val})
# SGD update
if (executor_ctx is None):
W1_val = W1_val - lr * grad_W1_val
b1_val = b1_val - lr * grad_b1_val
else:
sgd_update_gpu(W1_val, grad_W1_val, lr)
sgd_update_gpu(b1_val, grad_b1_val, lr)
if print_loss_val_each_epoch:
if isinstance(loss_val, ndarray.NDArray):
print(loss_val.asnumpy())
else:
print(loss_val)
correct_predictions = []
for minibatch_index in range(n_valid_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
valid_X_val[:] = valid_set_x[minibatch_start:minibatch_end]
valid_y_val[:] = convert_to_one_hot(
valid_set_y[minibatch_start:minibatch_end])
_, _, _, valid_y_predicted = executor.run(
feed_dict={
X: valid_X_val,
y_: valid_y_val,
W1: W1_val,
b1: b1_val},
convert_to_numpy_ret_vals=True)
correct_prediction = np.equal(
np.argmax(valid_y_val, 1),
np.argmax(valid_y_predicted, 1)).astype(np.float)
correct_predictions.extend(correct_prediction)
accuracy = np.mean(correct_predictions)
# validation set accuracy=0.928200
print("validation set accuracy=%f" % accuracy)
def mnist_mlp(executor_ctx=None, num_epochs=10, print_loss_val_each_epoch=False):
print("Build 3-layer MLP model...")
W1 = ad.Variable(name="W1")
W2 = ad.Variable(name="W2")
W3 = ad.Variable(name="W3")
b1 = ad.Variable(name="b1")
b2 = ad.Variable(name="b2")
b3 = ad.Variable(name="b3")
X = ad.Variable(name="X")
y_ = ad.Variable(name="y_")
# relu(X W1+b1)
z1 = ad.matmul_op(X, W1)
z2 = z1 + ad.broadcastto_op(b1, z1)
z3 = ad.relu_op(z2)
# relu(z3 W2+b2)
z4 = ad.matmul_op(z3, W2)
z5 = z4 + ad.broadcastto_op(b2, z4)
z6 = ad.relu_op(z5)
# softmax(z5 W2+b2)
z7 = ad.matmul_op(z6, W3)
y = z7 + ad.broadcastto_op(b3, z7)
loss = ad.softmaxcrossentropy_op(y, y_)
grad_W1, grad_W2, grad_W3, grad_b1, grad_b2, grad_b3 = ad.gradients(
loss, [W1, W2, W3, b1, b2, b3])
executor = ad.Executor(
[loss, grad_W1, grad_W2, grad_W3, grad_b1, grad_b2, grad_b3, y],
ctx=executor_ctx)
# Read input data
datasets = load_mnist_data("mnist.pkl.gz")
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# Set up minibatch
batch_size = 1000
n_train_batches = train_set_x.shape[0] // batch_size
n_valid_batches = valid_set_x.shape[0] // batch_size
print("Start training loop...")
# Initialize parameters
rand = np.random.RandomState(seed=123)
W1_val = rand.normal(scale=0.1, size=(784, 256))
W2_val = rand.normal(scale=0.1, size=(256, 100))
W3_val = rand.normal(scale=0.1, size=(100, 10))
b1_val = rand.normal(scale=0.1, size=(256))
b2_val = rand.normal(scale=0.1, size=(100))
b3_val = rand.normal(scale=0.1, size=(10))
X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
valid_X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
valid_y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
if ndarray.is_gpu_ctx(executor_ctx):
W1_val = ndarray.array(W1_val, ctx=executor_ctx)
W2_val = ndarray.array(W2_val, ctx=executor_ctx)
W3_val = ndarray.array(W3_val, ctx=executor_ctx)
b1_val = ndarray.array(b1_val, ctx=executor_ctx)
b2_val = ndarray.array(b2_val, ctx=executor_ctx)
b3_val = ndarray.array(b3_val, ctx=executor_ctx)
X_val = ndarray.array(X_val, ctx=executor_ctx)
y_val = ndarray.array(y_val, ctx=executor_ctx)
lr = 1.0e-3
for i in range(num_epochs):
print("epoch %d" % i)
for minibatch_index in range(n_train_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
X_val[:] = train_set_x[minibatch_start:minibatch_end]
y_val[:] = convert_to_one_hot(
train_set_y[minibatch_start:minibatch_end])
loss_val, grad_W1_val, grad_W2_val, grad_W3_val, \
grad_b1_val, grad_b2_val, grad_b3_val, _ = executor.run(
feed_dict={
X: X_val,
y_: y_val,
W1: W1_val,
W2: W2_val,
W3: W3_val,
b1: b1_val,
b2: b2_val,
b3: b3_val})
# SGD update
if (executor_ctx is None):
W1_val = W1_val - lr * grad_W1_val
W2_val = W2_val - lr * grad_W2_val
W3_val = W3_val - lr * grad_W3_val
b1_val = b1_val - lr * grad_b1_val
b2_val = b2_val - lr * grad_b2_val
b3_val = b3_val - lr * grad_b3_val
else:
sgd_update_gpu(W1_val, grad_W1_val, lr)
sgd_update_gpu(W2_val, grad_W2_val, lr)
sgd_update_gpu(W3_val, grad_W3_val, lr)
sgd_update_gpu(b1_val, grad_b1_val, lr)
sgd_update_gpu(b2_val, grad_b2_val, lr)
sgd_update_gpu(b3_val, grad_b3_val, lr)
if print_loss_val_each_epoch:
if isinstance(loss_val, ndarray.NDArray):
print(loss_val.asnumpy())
else:
print(loss_val)
correct_predictions = []
for minibatch_index in range(n_valid_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
valid_X_val[:] = valid_set_x[minibatch_start:minibatch_end]
valid_y_val[:] = convert_to_one_hot(
valid_set_y[minibatch_start:minibatch_end])
_, _, _, _, _, _, _, valid_y_predicted = executor.run(
feed_dict={
X: valid_X_val,
y_: valid_y_val,
W1: W1_val,
W2: W2_val,
W3: W3_val,
b1: b1_val,
b2: b2_val,
b3: b3_val},
convert_to_numpy_ret_vals=True)
correct_prediction = np.equal(
np.argmax(valid_y_val, 1),
np.argmax(valid_y_predicted, 1)).astype(np.float)
correct_predictions.extend(correct_prediction)
accuracy = np.mean(correct_predictions)
# validation set accuracy=0.970800
print("validation set accuracy=%f" % accuracy)