def check_node_grads(node, in_arrays, grad_arrays, rtol=None, atol=None):
for in_port, in_array in in_arrays.items():
for out_port, out_array in grad_arrays.items():
# print(in_port, out_port)
layer = LayerWrap(node, in_port, out_port, in_arrays, grad_arrays)
layer._setup(in_arrays[in_port].shape)
check_grad(layer, in_arrays[in_port], rtol=rtol, atol=atol)
def test_convolution():
n_filters = [1, 5, 10]
win_shapes = [(1, 1), (3, 3), (5, 5)]
strides = [(1, 1), (2, 2), (3, 3)]
border_modes = ['valid', 'same', 'full']
confs = itertools.product(batch_sizes, n_channels, img_shapes, n_filters,
win_shapes, strides, border_modes)
# Sample random parameter configurations to reduce workload.
confs = shuffled(confs)[:100]
for (batch_size, n_channel, img_shape, n_filter, win_shape, stride,
border_mode) in confs:
if img_shape[0] < win_shape[0] or img_shape[1] < win_shape[1]:
continue
print('Convolution: batch_size=%i, n_channel=%i, img_shape=%s, '
'n_filter=%i, win_shape=%s, stride=%s, border_mode=%s'
% (batch_size, n_channel, str(img_shape), n_filter,
str(win_shape), str(stride), border_mode))
x_shape = (batch_size, n_channel) + img_shape
w_shape = (n_filter, n_channel) + win_shape
x = np.random.normal(size=x_shape).astype(ca.float_)
w = np.random.normal(size=w_shape).astype(ca.float_)*1e-4
b = np.random.normal(size=(1, n_filter, 1, 1)).astype(ca.float_)*1e-4
layer = dp.Convolution(n_filter, win_shape, weights=w, bias=b,
strides=stride, border_mode=border_mode)
layer._setup(x_shape)
y_img_shape = img_out_shape(img_shape, win_shape, stride, border_mode)
assert layer.y_shape(x_shape) == (batch_size, n_filter) + y_img_shape
check_grad(layer, x)
check_params(layer)
def check_node_grads(node, in_arrays, grad_arrays, rtol=None, atol=None):
for in_port, in_array in in_arrays.items():
for out_port, out_array in grad_arrays.items():
# print(in_port, out_port)
layer = LayerWrap(node, in_port, out_port, in_arrays, grad_arrays)
layer._setup(in_arrays[in_port].shape)
check_grad(layer, in_arrays[in_port], rtol=rtol, atol=atol)
def test_convolution():
n_filters = [1, 5, 10]
win_shapes = [(1, 1), (3, 3), (5, 5)]
strides = [(1, 1), (2, 2), (3, 3)]
border_modes = ['valid', 'same', 'full']
confs = itertools.product(batch_sizes, n_channels, img_shapes, n_filters,
win_shapes, strides, border_modes)
# Sample random parameter configurations to reduce workload.
confs = shuffled(confs)[:100]
for (batch_size, n_channel, img_shape, n_filter, win_shape, stride,
border_mode) in confs:
if img_shape[0] < win_shape[0] or img_shape[1] < win_shape[1]:
continue
print('Convolution: batch_size=%i, n_channel=%i, img_shape=%s, '
'n_filter=%i, win_shape=%s, stride=%s, border_mode=%s'
% (batch_size, n_channel, str(img_shape), n_filter,
str(win_shape), str(stride), border_mode))
x_shape = (batch_size, n_channel) + img_shape
w_shape = (n_filter, n_channel) + win_shape
x = np.random.normal(size=x_shape).astype(ca.float_)
w = np.random.normal(size=w_shape).astype(ca.float_)*1e-4
b = np.random.normal(size=(1, n_filter, 1, 1)).astype(ca.float_)*1e-4
layer = dp.Convolution(n_filter, win_shape, weights=w, bias=b,
strides=stride, border_mode=border_mode)
layer.setup(x_shape)
y_img_shape = img_out_shape(img_shape, win_shape, stride, border_mode)
assert layer.y_shape(x_shape) == (batch_size, n_filter) + y_img_shape
check_grad(layer, x)
check_params(layer)
def test_activation():
confs = itertools.product(batch_sizes, n_ins, activations)
for batch_size, n_in, activation in confs:
print('Activation: batch_size=%i, n_in=%i, method=%s'
% (batch_size, n_in, str(activation)))
x_shape = (batch_size, n_in)
x = np.random.normal(size=x_shape).astype(dp.float_)
layer = dp.Activation.from_any(activation)
layer._setup(x_shape)
assert layer.y_shape(x_shape) == x_shape
if activation == 'sigmoid':
rtol = 1e-05
else:
rtol = None
check_grad(layer, x, rtol=rtol)
def test_pool():
win_shapes = [(1, 1), (2, 2), (3, 3)]
strides = [(1, 1), (2, 2), (3, 3)]
border_modes = ['valid', 'same', 'full']
methods = ['max', 'avg']
confs = itertools.product(batch_sizes, n_channels, img_shapes,
win_shapes, strides, border_modes, methods)
# Sample random parameter configurations to reduce workload.
confs = shuffled(confs)[:100]
for (batch_size, n_channel, img_shape, win_shape, stride, border_mode,
method) in confs:
if img_shape[0] < win_shape[0] or img_shape[1] < win_shape[1]:
continue
if border_mode != 'valid' and \
(win_shape[0] != 1 or win_shape[1] != 1) and \
ca._backend == 'cuda' and \
ca.nnet.pool._default_impl == 'cudnn':
# Bug: I think CUDArray/DeepPy calculates the padding in a manner
# inconsistent with cuDNN
continue
if method == 'avg' and \
ca._backend == 'cuda' and \
ca.nnet.pool._default_impl == 'masked':
# Not implemented yet
continue
print('Pool: batch_size=%i, n_channel=%i, img_shape=%s, win_shape=%s, '
'stride=%s, border_mode=%s, method=%s'
% (batch_size, n_channel, str(img_shape), str(win_shape),
str(stride), border_mode, method))
x_shape = (batch_size, n_channel) + img_shape
x = np.random.normal(size=x_shape).astype(ca.float_)
layer = dp.Pool(win_shape=win_shape, method=method, strides=stride,
border_mode=border_mode)
layer._setup(x_shape)
y_img_shape = img_out_shape(img_shape, win_shape, stride, border_mode)
assert layer.y_shape(x_shape) == (batch_size, n_channel) + y_img_shape
check_grad(layer, x)
def test_pool():
win_shapes = [(1, 1), (2, 2), (3, 3)]
strides = [(1, 1), (2, 2), (3, 3)]
border_modes = ['valid', 'same', 'full']
methods = ['max', 'avg']
confs = itertools.product(batch_sizes, n_channels, img_shapes,
win_shapes, strides, border_modes, methods)
# Sample random parameter configurations to reduce workload.
confs = shuffled(confs)[:100]
for (batch_size, n_channel, img_shape, win_shape, stride, border_mode,
method) in confs:
if img_shape[0] < win_shape[0] or img_shape[1] < win_shape[1]:
continue
if border_mode != 'valid' and \
(win_shape[0] != 1 or win_shape[1] != 1):
# Bug: I think CUDArray/DeepPy calculates the padding in a manner
# inconsistent with cuDNN
continue
print('Pool: batch_size=%i, n_channel=%i, img_shape=%s, win_shape=%s, '
'stride=%s, border_mode=%s, method=%s'
% (batch_size, n_channel, str(img_shape), str(win_shape),
str(stride), border_mode, method))
x_shape = (batch_size, n_channel) + img_shape
x = np.random.normal(size=x_shape).astype(ca.float_)
layer = dp.Pool(win_shape=win_shape, method=method, strides=stride,
border_mode=border_mode)
layer.setup(x_shape)
y_img_shape = img_out_shape(img_shape, win_shape, stride, border_mode)
assert layer.y_shape(x_shape) == (batch_size, n_channel) + y_img_shape
rtol = 1e-5
atol = 1e-7
check_grad(layer, x, rtol=rtol, atol=atol)