def test_forward_consistency(self):
x_cpu = chainer.Variable(self.x)
W_cpu = chainer.Variable(self.W)
b_cpu = None if self.nobias else chainer.Variable(self.b)
y_cpu = F.deconvolution_2d(x_cpu,
W_cpu,
b_cpu,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
use_cudnn=self.use_cudnn)
x_gpu = chainer.Variable(cuda.to_gpu(self.x))
W_gpu = chainer.Variable(cuda.to_gpu(self.W))
b_gpu = None if self.nobias else chainer.Variable(cuda.to_gpu(self.b))
y_gpu = F.deconvolution_2d(x_gpu,
W_gpu,
b_gpu,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
use_cudnn=self.use_cudnn)
self.assertEqual(y_cpu.data.dtype, self.x_dtype)
self.assertEqual(y_gpu.data.dtype, self.x_dtype)
testing.assert_allclose(y_cpu.data, y_gpu.data.get(),
**self.test_forward_options)
def test_forward_consistency(self):
x_cpu = chainer.Variable(self.x)
W_cpu = chainer.Variable(self.W)
b_cpu = None if self.nobias else chainer.Variable(self.b)
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
y_cpu = F.deconvolution_2d(
x_cpu, W_cpu, b_cpu, stride=self.stride, pad=self.pad,
outsize=self.outsize)
x_gpu = chainer.Variable(cuda.to_gpu(self.x))
W_gpu = chainer.Variable(cuda.to_gpu(self.W))
b_gpu = None if self.nobias else chainer.Variable(
cuda.to_gpu(self.b))
with chainer.using_config('use_cudnn', self.use_cudnn):
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
y_gpu = F.deconvolution_2d(
x_gpu, W_gpu, b_gpu, stride=self.stride, pad=self.pad,
outsize=self.outsize)
self.assertEqual(y_cpu.data.dtype, self.x_dtype)
self.assertEqual(y_gpu.data.dtype, self.x_dtype)
testing.assert_allclose(
y_cpu.data, y_gpu.data.get(), **self.test_forward_options)
def test_forward_consistency(self):
x_cpu = chainer.Variable(self.x)
W_cpu = chainer.Variable(self.W)
b_cpu = None if self.nobias else chainer.Variable(self.b)
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
y_cpu = F.deconvolution_2d(x_cpu,
W_cpu,
b_cpu,
stride=self.stride,
pad=self.pad,
outsize=self.outsize)
x_gpu = chainer.Variable(cuda.to_gpu(self.x))
W_gpu = chainer.Variable(cuda.to_gpu(self.W))
b_gpu = None if self.nobias else chainer.Variable(cuda.to_gpu(self.b))
with chainer.using_config('use_cudnn', self.use_cudnn):
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
with chainer.using_config('autotune', self.autotune):
y_gpu = F.deconvolution_2d(x_gpu,
W_gpu,
b_gpu,
stride=self.stride,
pad=self.pad,
outsize=self.outsize)
self.assertEqual(y_cpu.data.dtype, self.x_dtype)
self.assertEqual(y_gpu.data.dtype, self.x_dtype)
testing.assert_allclose(y_cpu.data, y_gpu.data.get(),
**self.test_forward_options)
def decode(self, c):
h = compose(c, [
lambda x: F.deconvolution_2d(x, self.conv4.W), self.bnd4, F.relu,
lambda x: F.deconvolution_2d(x, self.conv3.W, stride=2), self.bnd3,
F.relu, lambda x: F.deconvolution_2d(x, self.conv2.W), self.bnd2,
F.relu,
lambda x: F.deconvolution_2d(x, self.conv1.W, stride=2, pad=2),
F.sigmoid
])
return h
def forward_expected(self, link, inputs):
x, = inputs
W = link.W
if self.nobias:
y = F.deconvolution_2d(
x, W,
stride=self.stride, pad=self.pad,
dilate=self.dilate, groups=self.groups)
else:
b = link.b
y = F.deconvolution_2d(
x, W, b,
stride=self.stride, pad=self.pad,
dilate=self.dilate, groups=self.groups)
return y.array,
def _run_forward(self, x_data, W_data, b_data):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
b = None if self.nobias else chainer.Variable(b_data)
with chainer.using_config('use_cudnn', 'always'):
y = F.deconvolution_2d(x, W, b, stride=self.stride, pad=self.pad)
return x, W, b, y
def check_forward_consistency_regression(self,
x_data,
W_data,
b_data,
use_cudnn='always'):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
b = None if self.nobias else chainer.Variable(b_data)
with chainer.using_config('use_cudnn', use_cudnn):
y_nd = F.deconvolution_nd(x,
W,
b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate)
y_2d = F.deconvolution_2d(x,
W,
b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate)
testing.assert_allclose(y_nd.data, y_2d.data,
**self.test_forward_options)
def check_forward_consistency_regression(self, backend_config):
inputs = self.generate_inputs()
if self.nobias:
x, W = inputs
b = None
else:
x, W, b = inputs
x = chainer.Variable(backend_config.get_array(x))
W = chainer.Variable(backend_config.get_array(W))
if b is not None:
b = chainer.Variable(backend_config.get_array(b))
use_cudnn = backend_config.use_cudnn
with chainer.using_config('use_cudnn', use_cudnn):
y_nd = F.deconvolution_nd(x,
W,
b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate)
y_2d = F.deconvolution_2d(x,
W,
b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate)
testing.assert_allclose(y_nd.array, y_2d.array,
**self.check_forward_options)
def _run_forward(self, x_data, W_data, b_data):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
b = None if self.nobias else chainer.Variable(b_data)
y = F.deconvolution_2d(x, W, b, stride=self.stride, pad=self.pad,
use_cudnn=True, deterministic=True)
return x, W, b, y
def _run_forward(self, x_data, W_data, b_data):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
b = None if self.nobias else chainer.Variable(b_data)
with chainer.using_config('use_cudnn', 'always'):
y = F.deconvolution_2d(x, W, b, stride=self.stride, pad=self.pad)
return x, W, b, y
def check_forward(self, inputs, backend_config):
y_expected, = self.forward_cpu(inputs)
if backend_config.use_cuda:
inputs = cuda.to_gpu(inputs)
x, W, b = inputs
x = chainer.Variable(x)
W = chainer.Variable(W)
b = None if b is None else chainer.Variable(b)
with backend_config:
y_actual = F.deconvolution_2d(x,
W,
b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate,
groups=self.groups)
assert y_expected.data.dtype == self.x_dtype
assert y_actual.data.dtype == self.x_dtype
testing.assert_allclose(y_expected.data, y_actual.data,
**self.test_forward_options)
def gen_convtranspose_bn(test_name):
gb = onnx_script.GraphBuilder(test_name)
bsize = 2
ichan = 3
ochan = 4
ksize = 3
isize = 7
x = aranges(bsize, ochan, isize, isize)
w = aranges(ochan, ichan, ksize, ksize) * 0.01
scale = aranges(ichan) * 0.1 + 1
bias = aranges(ichan) * 0.1 + 2
mean = aranges(ichan) * 0.1 + 3
var = aranges(ichan) * 0.1 + 4
conv = F.deconvolution_2d(x, w, pad=1, outsize=(isize, isize))
y = F.fixed_batch_normalization(conv, scale, bias, mean, var)
x_v = gb.input('x', x)
w_v = gb.param('w', w)
scale_v = gb.param('scale', scale)
bias_v = gb.param('bias', bias)
mean_v = gb.param('mean', mean)
var_v = gb.param('var', var)
conv_v = gb.ConvTranspose([x_v, w_v],
kernel_shape=[ksize, ksize],
pads=[1, 1, 1, 1],
output_shape=[isize, isize])
y_v = gb.BatchNormalization([conv_v, scale_v, bias_v, mean_v, var_v])
gb.output(y_v, y)
gb.gen_test()
def __call__(self, x):
return F.deconvolution_2d(x,
spectral_normalize(self.W, self.init_u),
b=self.b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize)
def f(*args):
return F.deconvolution_2d(*args,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate,
groups=self.groups)
def conv_point_to_circle(in_imgs):
""" 点画像 ⇒ circleを描画、in_imgs = [N_pic][1][imgH][imgW] np, fp32 (0-1)
⇒return: [N_pic][1][imgH][imgW] np, fp32 0-255"""
cir = np.zeros((1, 1, 15, 15), dtype="float32")
rr, cc = draw.circle_perimeter(7, 7, 5)
cir[0][0][rr, cc] = 255
out_imgs = F.deconvolution_2d(in_imgs, W=cir, b=None, stride=1, pad=7)
return out_imgs.data
def f(*args):
y = F.deconvolution_2d(*args,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate,
groups=self.groups)
return y * y # make the function nonlinear
def test_forward_consistency(self):
x_cpu = chainer.Variable(self.x)
W_cpu = chainer.Variable(self.W)
b_cpu = None if self.nobias else chainer.Variable(self.b)
y_cpu = F.deconvolution_2d(
x_cpu, W_cpu, b_cpu, stride=self.stride, pad=self.pad,
outsize=self.outsize, use_cudnn=self.use_cudnn)
x_gpu = chainer.Variable(cuda.to_gpu(self.x))
W_gpu = chainer.Variable(cuda.to_gpu(self.W))
b_gpu = None if self.nobias else chainer.Variable(
cuda.to_gpu(self.b))
y_gpu = F.deconvolution_2d(
x_gpu, W_gpu, b_gpu, stride=self.stride, pad=self.pad,
outsize=self.outsize, use_cudnn=self.use_cudnn)
gradient_check.assert_allclose(y_cpu.data, y_gpu.data.get())
def test_forward_consistency(self):
x_cpu = chainer.Variable(self.x)
W_cpu = chainer.Variable(self.W)
b_cpu = None if self.nobias else chainer.Variable(self.b)
y_cpu = F.deconvolution_2d(
x_cpu, W_cpu, b_cpu, stride=self.stride, pad=self.pad,
outsize=self.outsize, use_cudnn=self.use_cudnn)
x_gpu = chainer.Variable(cuda.to_gpu(self.x))
W_gpu = chainer.Variable(cuda.to_gpu(self.W))
b_gpu = None if self.nobias else chainer.Variable(
cuda.to_gpu(self.b))
y_gpu = F.deconvolution_2d(
x_gpu, W_gpu, b_gpu, stride=self.stride, pad=self.pad,
outsize=self.outsize, use_cudnn=self.use_cudnn)
gradient_check.assert_allclose(y_cpu.data, y_gpu.data.get())
def forward(self):
x = chainer.Variable(self.x)
W = chainer.Variable(self.W)
return F.deconvolution_2d(x,
W,
None,
stride=1,
pad=1,
use_cudnn=self.use_cudnn)
def forward(self):
x = chainer.Variable(self.x)
W = chainer.Variable(self.W)
return F.deconvolution_2d(x,
W,
None,
stride=1,
pad=1,
groups=self.groups)
def forward_cpu(self, inputs):
x, W, b = inputs
x_cpu = chainer.Variable(x)
W_cpu = chainer.Variable(W)
b_cpu = None if b is None else chainer.Variable(b)
y_cpu = F.deconvolution_2d(
x_cpu, W_cpu, b_cpu, stride=self.stride, pad=self.pad,
outsize=self.outsize)
return y_cpu,
def visual_mask(self):
z = self.averageL4 * self.averageL3
z = z * self.averageL2
z = F.deconvolution_2d(Variable(z),self.initialW,stride = 2,pad=1).data * self.averageL1
z = F.deconvolution_2d(Variable(z),self.initialW,stride = 2,pad=1).data
return z
def forward_cpu(self, inputs):
x, W, b = inputs
x_cpu = chainer.Variable(x)
W_cpu = chainer.Variable(W)
b_cpu = None if b is None else chainer.Variable(b)
y_cpu = F.deconvolution_2d(
x_cpu, W_cpu, b_cpu, stride=self.stride, pad=self.pad,
outsize=self.outsize, dilate=self.dilate, group=self.group)
return y_cpu,
def forward_cpu(self, inputs):
x, W, b = inputs
x_cpu = chainer.Variable(x)
W_cpu = chainer.Variable(W)
b_cpu = None if b is None else chainer.Variable(b)
with chainer.using_config('use_ideep', 'never'):
y_cpu = F.deconvolution_2d(
x_cpu, W_cpu, b_cpu, stride=self.stride, pad=self.pad,
outsize=self.outsize, dilate=self.dilate, groups=self.groups)
return y_cpu,
def conv_point_to_core(in_imgs, sig=3.0, max_xy=15, c_xy=7):
""" 点画像 ⇒ コアを描画、in_imgs = [1][1][imgH][imgW] np, fp32 (0-1)
⇒return: [N_pic][1][imgH][imgW] np, fp32 (0-1)"""
sig2 = sig * sig
core = np.zeros((max_xy, max_xy), dtype="float32")
for px in range(0, max_xy):
for py in range(0, max_xy):
r2 = float((px - c_xy) * (px - c_xy) + (py - c_xy) * (py - c_xy))
core[py][px] = math.exp(-r2 / sig2) * 1
core = core.reshape((1, 1, core.shape[0], core.shape[1]))
out_imgs = F.deconvolution_2d(in_imgs, W=core, b=None, stride=1, pad=c_xy)
return out_imgs.data
def check_forward_consistency_regression(self, x_data, W_data, b_data,
use_cudnn=True):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
b = None if self.nobias else chainer.Variable(b_data)
y_nd = F.deconvolution_nd(x, W, b, stride=self.stride, pad=self.pad,
outsize=self.outsize, use_cudnn=use_cudnn)
y_2d = F.deconvolution_2d(x, W, b, stride=self.stride, pad=self.pad,
outsize=self.outsize, use_cudnn=use_cudnn)
testing.assert_allclose(
y_nd.data, y_2d.data, **self.test_forward_options)
def forward(self, inputs, device):
if self.nobias:
x, W = inputs
b = None
else:
x, W, b = inputs
y = F.deconvolution_2d(x,
W,
b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate,
groups=self.groups)
return y,
def check_forward_consistency_regression(self, x_data, W_data, b_data,
use_cudnn='always'):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
b = None if self.nobias else chainer.Variable(b_data)
with chainer.using_config('use_cudnn', use_cudnn):
y_nd = F.deconvolution_nd(x, W, b, stride=self.stride,
pad=self.pad, outsize=self.outsize,
dilate=self.dilate)
y_2d = F.deconvolution_2d(x, W, b, stride=self.stride,
pad=self.pad, outsize=self.outsize,
dilate=self.dilate)
testing.assert_allclose(
y_nd.data, y_2d.data, **self.test_forward_options)
def test_forward2(self):
n, c_i, c_o = 10, 1, 3
h_i, w_i = 5, 10
h_k, w_k = 10, 10
h_p, w_p = 5, 5
s_y, s_x = 5, 5
x = np.random.uniform(0, 1, (n, c_i, h_i, w_i)).astype(np.float32)
W = np.random.uniform(0, 1, (c_i, c_o, h_k, w_k)).astype(np.float32)
b = None
expected = CF.deconvolution_2d(x,
W,
b,
stride=(s_y, s_x),
pad=(h_p, w_p))
y = F.deconv2d(x, W, b, stride=(s_y, s_x), pad=(h_p, w_p))
self.assertTrue(np.array_equal(expected.data, y.data))
def scale_layer(self, feature_map, node):
input_data = node.inputs[0].data
_, _, in_height, in_width = input_data.shape
_, _, feature_height, feature_width = feature_map.shape
kernel_height = in_height + 2 * node.ph - node.sy * (feature_height -
1)
kernel_width = in_width + 2 * node.pw - node.sx * (feature_width - 1)
scaled_feature = F.deconvolution_2d(
feature_map,
self.xp.ones((1, 1, kernel_height, kernel_width)),
stride=(node.sy, node.sx),
pad=(node.ph, node.pw),
outsize=(in_height, in_width),
)
averaged_feature_map = F.average(input_data, axis=1, keepdims=True)
feature_map = scaled_feature * averaged_feature_map
return feature_map
def check_forward(self, inputs, backend_config):
y_expected, = self.forward_cpu(inputs)
x, W, b = backend_config.get_array(inputs)
x = chainer.Variable(x)
W = chainer.Variable(W)
b = None if b is None else chainer.Variable(b)
with backend_config:
y_actual = F.deconvolution_2d(
x, W, b, stride=self.stride, pad=self.pad,
outsize=self.outsize, dilate=self.dilate, groups=self.groups)
assert y_expected.data.dtype == self.x_dtype
assert y_actual.data.dtype == self.x_dtype
testing.assert_allclose(
y_expected.data, y_actual.data, **self.test_forward_options)
def conv_spectral_norm_exact(kernel, shape, stride, pad):
xp = get_array_module(kernel)
shape = (128, ) + shape[1:]
x = xp.random.normal(size=shape).astype(xp.float32)
normalize(x, (1, 2, 3))
prev = None
eps = 1e20
with chainer.no_backprop_mode():
for i in range(500):
x = convolution_2d(x, kernel, stride=stride, pad=pad).array
x = deconvolution_2d(x, kernel, stride=stride, pad=pad).array
norm = normalize(x, (1, 2, 3))
if prev is not None:
eps = norm - prev
prev = norm
f = xp.abs(eps) * np.prod(shape[1:])
error = (f + xp.sqrt(f * (4 * prev + f))) / 2
return xp.sqrt(xp.max(prev + error))
def deconv(self, x):
wdec = self.c.W
wdec = F.transpose(wdec, (1, 0, 2, 3)) #[ch_in, ch_out, kh, kw]
wdec = F.pad(wdec, [[0, 0], [0, 0], [1, 1], [1, 1]],
'constant',
constant_values=0)
wdec = wdec[:, :, 1:,
1:] + wdec[:, :, :-1,
1:] + wdec[:, :, 1:, :-1] + wdec[:, :, :-1, :-1]
out_shape = [x.shape[2] * 2, x.shape[3] * 2]
return F.deconvolution_2d(x,
wdec,
b=self.c.b,
pad=self.c.pad,
stride=2,
outsize=out_shape)
def forward(self, x):
if self.W.array is None:
self._initialize_params(x.shape[1])
pad_width = [(0, 0), (0, 0)] + list(map(lambda x: (x, x), self.pad))
x = F.pad(x, pad_width, self.pad_mode)
y = F.deconvolution_2d(x,
self.W,
self.b,
self.stride,
0,
self.outsize,
dilate=self.dilate,
groups=self.groups)
return y
def test_forward_consistency(self):
x_cpu = chainer.Variable(self.x)
W_cpu = chainer.Variable(self.W)
b_cpu = None if self.nobias else chainer.Variable(self.b)
y_cpu = F.deconvolution_2d(
x_cpu, W_cpu, b_cpu, stride=self.stride, pad=self.pad,
outsize=self.outsize)
x_mkl = chainer.Variable(self.x)
W_mkl = chainer.Variable(self.W)
b_mkl = None if self.nobias else chainer.Variable(self.b)
y_mkl = E.deconvolution_2d(
x_mkl, W_mkl, b_mkl, stride=self.stride, pad=self.pad,
outsize=self.outsize)
self.assertEqual(y_cpu.data.dtype, self.x_dtype)
self.assertEqual(y_mkl.data.dtype, self.x_dtype)
testing.assert_allclose(
y_cpu.data, numpy.array(y_mkl.data, copy=False),
**self.test_forward_options)
def forward_expected(self, inputs):
"""
Current forward_expected implementation depends on
F.deconvolution_2d itself and thus it's only capable
of checking consistency between backends, not absolute
correctness of computations
"""
if self.nobias:
x, W = inputs
b = None
else:
x, W, b = inputs
y_expected = F.deconvolution_2d(x,
W,
b,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
dilate=self.dilate,
groups=self.groups)
return y_expected.array,
def check_backward(self, x_data, W_data, b_data, y_grad):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
b = None if b_data is None else chainer.Variable(b_data)
y = F.deconvolution_2d(x, W, b, stride=self.stride, pad=self.pad,
use_cudnn=self.use_cudnn)
y.grad = y_grad
y.backward()
func = y.creator
if b is None:
f = lambda: func.forward((x.data, W.data))
gx, gW = gradient_check.numerical_grad(
f, (x.data, W.data), (y.grad,), eps=1e-2)
else:
f = lambda: func.forward((x.data, W.data, b.data))
gx, gW, gb = gradient_check.numerical_grad(
f, (x.data, W.data, b.data), (y.grad,), eps=1e-2)
gradient_check.assert_allclose(gx, x.grad)
gradient_check.assert_allclose(gW, W.grad)
if b is not None:
gradient_check.assert_allclose(gb, b.grad)
def f(*args):
return F.deconvolution_2d(
*args, stride=self.stride, pad=self.pad, outsize=self.outsize)
def f(*args):
y = F.deconvolution_2d(
*args, stride=self.stride, pad=self.pad, outsize=self.outsize,
dilate=self.dilate, group=self.group)
return y * y # make the function nonlinear
def f(*args):
return F.deconvolution_2d(
*args, stride=self.stride, pad=self.pad, outsize=self.outsize,
dilate=self.dilate, groups=self.groups)
def forward(self):
x = chainer.Variable(self.x)
W = chainer.Variable(self.W)
return F.deconvolution_2d(x, W, None, stride=1, pad=1,
groups=self.groups)
def forward(self):
x = chainer.Variable(self.x)
W = chainer.Variable(self.W)
return F.deconvolution_2d(
x, W, None, stride=1, pad=1, use_cudnn=self.use_cudnn)
