def _test_conv_forward(input_shape, out_channels, kernel_size, stride):
return
np.random.seed(0)
torch.manual_seed(0)
in_channels = input_shape[1]
padding = (kernel_size - 1) // 2
input = np.random.random(input_shape).astype(np.float32) * 20
original_input = input.copy()
layer = ConvLayer(in_channels, out_channels, kernel_size, stride)
torch_layer = nn.Conv2d(in_channels,
out_channels,
kernel_size,
stride,
padding,
bias=True)
utils.assign_conv_layer_weights(layer, torch_layer)
output = layer.forward(input)
torch_data = utils.from_numpy(input)
torch_out = torch_layer(torch_data)
assert np.all(input == original_input)
assert output.shape == torch_out.shape
try:
utils.assert_close(output, torch_out, atol=TOLERANCE)
except AssertionError:
import pdb
pdb.set_trace()
def _test_conv_backward(input_shape, out_channels, kernel_size, stride):
np.random.seed(0)
torch.manual_seed(0)
in_channels = input_shape[1]
padding = (kernel_size - 1) // 2
input = np.random.random(input_shape).astype(np.float32) * 20
layer = ConvLayer(in_channels, out_channels, kernel_size, stride)
torch_layer = nn.Conv2d(in_channels,
out_channels,
kernel_size,
stride,
padding,
bias=True)
utils.assign_conv_layer_weights(layer, torch_layer)
output = layer.forward(input)
out_grad = layer.backward(2 * np.ones_like(output) / output.size)
torch_input = utils.from_numpy(input).requires_grad_(True)
torch_out = torch_layer(torch_input)
(2 * torch_out.mean()).backward()
try:
utils.assert_close(out_grad, torch_input.grad, atol=TOLERANCE)
except ValueError:
import pdb
pdb.set_trace()
utils.check_conv_grad_match(layer, torch_layer)
def test_networks():
np.random.seed(0)
torch.manual_seed(0)
data = np.random.random((100, 1, 28, 28)).astype(np.float32) * 10 - 5
labels = np.random.randint(0, 10, 100).astype(np.int64)
net = MNISTResNetwork()
torch_net = TorchMNISTResNetwork()
utils.assign_conv_layer_weights(net.layers[0], torch_net.layers[0])
utils.assign_conv_layer_weights(net.layers[3], torch_net.layers[3])
utils.assign_conv_layer_weights(net.layers[4].conv_layers[0], torch_net.layers[4].conv1)
utils.assign_conv_layer_weights(net.layers[4].conv_layers[2], torch_net.layers[4].conv2)
utils.assign_conv_layer_weights(net.layers[5].conv_layers[0], torch_net.layers[5].conv1)
utils.assign_conv_layer_weights(net.layers[5].conv_layers[2], torch_net.layers[5].conv2)
utils.assign_linear_layer_weights(net.layers[9], torch_net.layers[9])
utils.assign_linear_layer_weights(net.layers[11], torch_net.layers[11])
utils.assign_linear_layer_weights(net.layers[13], torch_net.layers[13])
forward = net(data)
data_torch = utils.from_numpy(data).requires_grad_(True)
forward_torch = torch_net(data_torch)
utils.assert_close(forward, forward_torch)
loss = net.loss(forward, labels)
torch_loss = torch_net.loss(forward_torch, utils.from_numpy(labels))
utils.assert_close(loss, torch_loss)
out_grad = net.backward()
torch_loss.backward()
utils.assert_close(out_grad, data_torch.grad, atol=0.01)
tolerance = 1e-4
utils.check_linear_grad_match(net.layers[13], torch_net.layers[13], tolerance=tolerance)
utils.check_linear_grad_match(net.layers[11], torch_net.layers[11], tolerance=tolerance)
utils.check_linear_grad_match(net.layers[9], torch_net.layers[9], tolerance=tolerance)
utils.check_conv_grad_match(net.layers[5].conv_layers[2], torch_net.layers[5].conv2, tolerance=tolerance)
utils.check_conv_grad_match(net.layers[5].conv_layers[0], torch_net.layers[5].conv1, tolerance=tolerance)
utils.check_conv_grad_match(net.layers[4].conv_layers[2], torch_net.layers[4].conv2, tolerance=tolerance)
utils.check_conv_grad_match(net.layers[4].conv_layers[0], torch_net.layers[4].conv1, tolerance=tolerance)
utils.check_conv_grad_match(net.layers[3], torch_net.layers[3], tolerance=tolerance)
utils.check_conv_grad_match(net.layers[0], torch_net.layers[0], tolerance=tolerance)
def test_sgd_update():
net = nn.layers.LinearLayer(100, 10)
learning_rate = 1
optimizer = sgd_optimizer.SGDOptimizer(net.parameters(), learning_rate)
data = np.random.random((20, 100)).astype(np.float32) * 2 - 1
initial_weight = net.weight.data.copy()
initial_bias = net.bias.data.copy()
torch_net = TorchNet()
with torch.no_grad():
torch_net.layer.weight[:] = utils.from_numpy(net.weight.data.T)
torch_net.layer.bias[:] = utils.from_numpy(net.bias.data)
torch_optimizer = torch.optim.SGD(torch_net.parameters(), learning_rate)
optimizer.zero_grad()
out = net(data)
loss = out.sum()
net.backward(np.ones_like(out))
torch_optimizer.zero_grad()
torch_out = torch_net(utils.from_numpy(data))
assert np.allclose(out,
utils.to_numpy(torch_out.clone().detach()),
atol=0.001)
torch_loss = torch_out.sum()
assert np.allclose(loss, torch_loss.item(), atol=0.001)
torch_loss.backward()
assert np.allclose(net.weight.grad.T,
utils.to_numpy(torch_net.layer.weight.grad))
assert np.allclose(net.bias.grad,
utils.to_numpy(torch_net.layer.bias.grad))
optimizer.step()
torch_optimizer.step()
assert np.allclose(net.weight.data.T,
utils.to_numpy(torch_net.layer.weight))
assert np.allclose(net.bias.data, utils.to_numpy(torch_net.layer.bias))
assert not np.allclose(net.weight.data, initial_weight)
assert not np.allclose(net.bias.data, initial_bias)
def _test_forward_overflow(input_shape, reduction, axis):
layer = SoftmaxCrossEntropyLossLayer(reduction=reduction)
data = np.random.random(input_shape) * 10000 - 1
labels_shape = list(data.shape)
labels_shape.pop(axis)
labels = np.random.randint(0, data.shape[axis], labels_shape)
loss = layer(data, labels, axis=axis)
if axis == 1:
pytorch_loss = F.cross_entropy(utils.from_numpy(data),
utils.from_numpy(labels),
reduction=reduction)
else:
pytorch_loss = F.cross_entropy(utils.from_numpy(data.swapaxes(1,
axis)),
utils.from_numpy(labels),
reduction=reduction)
pytorch_loss = utils.to_numpy(pytorch_loss)
assert np.allclose(loss, pytorch_loss, atol=0.001)
def _test_backward(input_shape, reduction, axis):
#np.random.Seed(0)
layer = SoftmaxCrossEntropyLossLayer(reduction=reduction)
data = np.random.random(input_shape) * 2 - 1
labels_shape = list(data.shape)
labels_shape.pop(axis)
labels = np.random.randint(0, data.shape[axis], labels_shape)
loss = layer(data, labels, axis=axis)
if axis == 1:
torch_input = utils.from_numpy(data).requires_grad_(True)
else:
torch_input = utils.from_numpy(np.moveaxis(data, axis,
1)).requires_grad_(True)
pytorch_loss = F.cross_entropy(torch_input,
utils.from_numpy(labels),
reduction=reduction)
if len(pytorch_loss.shape) > 0:
pytorch_loss.sum().backward()
else:
pytorch_loss.backward()
assert np.allclose(loss, utils.to_numpy(pytorch_loss))
grad = layer.backward()
torch_grad = utils.to_numpy(torch_input.grad)
if axis != 1:
torch_grad = np.moveaxis(torch_grad, 1, axis)
print(torch_grad.shape)
print(grad.shape)
print("grad - torch_grad < .000001")
print(np.absolute(grad - torch_grad) < .9)
assert np.allclose(grad, torch_grad, atol=0.001)
def _test_relu_forward(input_shape, out_channels):
in_channels = input_shape[1]
input = np.random.random(input_shape).astype(np.float32) * 20
original_input = input.copy()
layer = ReLULayer()
torch_layer = nn.ReLU()
output = layer.forward(input)
torch_data = utils.from_numpy(input)
torch_out = utils.to_numpy(torch_layer(torch_data))
output[np.abs(output) < 1e-4] = 0
torch_out[np.abs(torch_out) < 1e-4] = 0
assert np.all(input == original_input)
assert output.shape == torch_out.shape
assert np.allclose(output, torch_out, atol=TOLERANCE)
def _test_max_pool_backward(input_shape, kernel_size, stride):
np.random.seed(0)
torch.manual_seed(0)
padding = (kernel_size - 1) // 2
input = np.random.random(input_shape).astype(np.float32) * 20
layer = MaxPoolLayer(kernel_size, stride)
torch_layer = nn.MaxPool2d(kernel_size, stride, padding)
output = layer.forward(input)
out_grad = layer.backward(2 * np.ones_like(output) / output.size)
torch_input = utils.from_numpy(input).requires_grad_(True)
torch_out = torch_layer(torch_input)
(2 * torch_out.mean()).backward()
torch_out_grad = utils.to_numpy(torch_input.grad)
utils.assert_close(out_grad, torch_out_grad, atol=TOLERANCE)
def _test_max_pool_forward(input_shape, kernel_size, stride):
return
np.random.seed(0)
torch.manual_seed(0)
padding = (kernel_size - 1) // 2
input = np.random.random(input_shape).astype(np.float32) * 20
original_input = input.copy()
layer = MaxPoolLayer(kernel_size, stride)
torch_layer = nn.MaxPool2d(kernel_size, stride, padding)
output = layer.forward(input)
torch_data = utils.from_numpy(input)
torch_out = utils.to_numpy(torch_layer(torch_data))
output[np.abs(output) < 1e-4] = 0
torch_out[np.abs(torch_out) < 1e-4] = 0
assert np.all(input == original_input)
assert output.shape == torch_out.shape
utils.assert_close(output, torch_out, atol=TOLERANCE)
def _test_linear_backward(input_shape, out_channels,):
in_channels = input_shape[1]
input = np.random.random(input_shape).astype(np.float32) * 20
layer = LinearLayer(in_channels, out_channels)
torch_layer = nn.Linear(in_channels, out_channels, bias=True)
with torch.no_grad():
torch_layer.weight[:] = torch.from_numpy(layer.weight.data).transpose(0, 1)
torch_layer.bias[:] = torch.from_numpy(layer.bias.data)
output = layer.forward(input)
out_grad = layer.backward(np.ones_like(output))
torch_input = utils.from_numpy(input).requires_grad_(True)
torch_out = torch_layer(torch_input)
torch_out.sum().backward()
torch_out_grad = utils.to_numpy(torch_input.grad)
out_grad[np.abs(out_grad) < 1e-4] = 0
torch_out_grad[np.abs(torch_out_grad) < 1e-4] = 0
assert np.allclose(out_grad, torch_out_grad, atol=TOLERANCE)
w_grad = layer.weight.grad
w_grad[np.abs(w_grad) < 1e-4] = 0
torch_w_grad = utils.to_numpy(torch_layer.weight.grad.transpose(0, 1))
torch_w_grad[np.abs(torch_w_grad) < 1e-4] = 0
print(w_grad)
print(torch_w_grad)
print()
print("--------------")
print()
assert np.allclose(w_grad, torch_w_grad, atol=TOLERANCE)
b_grad = layer.bias.grad
b_grad[np.abs(b_grad) < 1e-4] = 0
torch_b_grad = utils.to_numpy(torch_layer.bias.grad)
torch_b_grad[np.abs(torch_b_grad) < 1e-4] = 0
assert np.allclose(b_grad, torch_b_grad, atol=TOLERANCE)
def _test_linear_forward(input_shape, out_channels):
in_channels = input_shape[1]
input = np.random.random(input_shape).astype(np.float32) * 20
original_input = input.copy()
layer = LinearLayer(in_channels, out_channels)
torch_layer = nn.Linear(in_channels, out_channels, bias=True)
with torch.no_grad():
torch_layer.weight[:] = torch.from_numpy(layer.weight.data).transpose(0, 1)
torch_layer.bias[:] = torch.from_numpy(layer.bias.data)
output = layer.forward(input)
torch_data = utils.from_numpy(input)
torch_out = utils.to_numpy(torch_layer(torch_data))
output[np.abs(output) < 1e-4] = 0
torch_out[np.abs(torch_out) < 1e-4] = 0
assert np.all(input == original_input)
assert output.shape == torch_out.shape
assert np.allclose(output, torch_out, atol=TOLERANCE)
def _test_relu_backward(
input_shape,
out_channels,
):
in_channels = input_shape[1]
input = np.random.random(input_shape).astype(np.float32) * 20
layer = ReLULayer()
torch_layer = nn.ReLU()
output = layer.forward(input)
out_grad = layer.backward(np.ones_like(output))
torch_input = utils.from_numpy(input).requires_grad_(True)
torch_out = torch_layer(torch_input)
torch_out.sum().backward()
torch_out_grad = utils.to_numpy(torch_input.grad)
out_grad[np.abs(out_grad) < 1e-4] = 0
torch_out_grad[np.abs(torch_out_grad) < 1e-4] = 0
assert np.allclose(out_grad, torch_out_grad, atol=TOLERANCE)