def test_col2im_consistency(self):
col = conv.im2col_cpu(self.x, 3, 3, 2, 2, 2, 2, dy=2, dx=2)
h, w = self.x.shape[2:]
im_cpu = conv.col2im_cpu(col, 2, 2, 2, 2, h, w, dy=2, dx=2)
im_gpu = conv.col2im_gpu(
cuda.to_gpu(col), 2, 2, 2, 2, h, w, dy=2, dx=2)
testing.assert_allclose(im_cpu, im_gpu.get())
def backward_cpu(self, x, gy):
h, w = x[0].shape[2:]
gcol = numpy.tile(gy[0][:, :, numpy.newaxis, numpy.newaxis],
(1, 1, self.kh, self.kw, 1, 1))
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
gx /= self.kh * self.kw
return gx,
def forward(self, x):
self.retain_inputs(())
h, w = x[0].shape[2:]
if self.outh is None:
self.outh = conv.get_deconv_outsize(h,
self.kh,
self.sy,
self.ph,
cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w,
self.kw,
self.sx,
self.pw,
cover_all=self.cover_all)
xp = cuda.get_array_module(*x)
col = xp.tile(x[0][:, :, None, None], (1, 1, self.kh, self.kw, 1, 1))
if xp is numpy:
y = conv.col2im_cpu(col, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
else:
y = conv.col2im_gpu(col, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
return y,
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
n, c, h, w = x.shape
out_c, input_c, kh, kw = W.shape
gn, gout_c, gout_h, gout_w = gy.shape
"""
For MKLDNN backward, only support float32
"""
if mkld.enable_convF(inputs):
gW = numpy.empty(shape=(out_c, input_c, kh, kw), dtype=W.dtype)
gx = numpy.empty(shape=(n, c, h, w), dtype=W.dtype)
if b is None:
mkldnn.Convolution2D_F32.do_backward(x, W, gy, gW, gx, kh, kw, self.sy, self.sx, self.ph, self.pw, self.pd, self.pr, self.mkldnn_opt)
return gx, gW
else:
gb = numpy.empty(shape=b.shape, dtype=W.dtype)
mkldnn.Convolution2D_F32.do_backward(x, W, b, gy, gW, gx, gb, kh, kw, self.sy, self.sx, self.ph, self.pw, self.pd, self.pr, self.mkldnn_opt)
return gx, gW, gb
else:
gW = numpy.tensordot(
gy, self.col, ((0, 2, 3), (0, 4, 5))).astype(W.dtype, copy=False)
gcol = numpy.tensordot(W, gy, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
if not type_check.same_types(*inputs):
if b is not None:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}, type(b): {2}'
.format(type(W), type(x), type(b)))
else:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}'
.format(type(W), type(x)))
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(
gy, self.col, ((0, 2, 3), (0, 4, 5))).astype(W.dtype, copy=False)
if not self.requires_x_grad:
gx = None
else:
gcol = numpy.tensordot(W, gy, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw,
h, w)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
h, w = x.shape[2:]
"""
N = numpy.random.uniform(-0.5, 0.5)
coef = numpy.max(gy, axis=tuple([i for i in xrange(1,gy.ndim)])).astype(numpy.float32)
coef = _as_mat(coef)
gy = _as_mat(gy)
coef_invert = 0.5*coef**(-1)
gy = gy*coef_invert+0.5+N/E_g
gy = quantize(gy, E_g)
gy = 2*coef*(gy-0.5)
gy = gy.reshape(grad_outputs[0].shape)
"""
gW = numpy.tensordot(gy, self.col,
((0, 2, 3), (0, 4, 5))).astype(W.dtype,
copy=False)
W = numpy.clip(W * 0.5 + 5, 0, 1)
Wq = 2 * quantize(W, E_w) - 1
gcol = numpy.tensordot(Wq, gy, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def forward_cpu(self, x):
self.retain_inputs(())
self._in_dtype = x[0].dtype
n, c, h, w = x[0].shape
if self.outh is None:
self.outh = conv.get_deconv_outsize(
h, self.kh, self.sy, self.ph, cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(
w, self.kw, self.sx, self.pw, cover_all=self.cover_all)
up_y = numpy.zeros((n, c, self.outh, self.outw), dtype=self._in_dtype)
up_y = conv.im2col_cpu(
up_y, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
for n in six.moves.range(up_y.shape[0]):
for c in six.moves.range(up_y.shape[1]):
for oy in six.moves.range(up_y.shape[4]):
for ox in six.moves.range(up_y.shape[5]):
ky = self.indexes[n, c, oy, ox] // up_y.shape[3]
kx = self.indexes[n, c, oy, ox] % up_y.shape[3]
up_y[n, c, ky, kx, oy, ox] = x[0][n, c, oy, ox]
up_y = conv.col2im_cpu(up_y, self.sy, self.sx, self.ph,
self.pw, self.outh, self.outw)
return up_y,
def backward_cpu(self, x, gy):
if mkld.enable_max_poolingF((x, gy)):
n, c, h, w = x[0].shape
gx = numpy.empty((n, c, h, w), dtype=x[0].dtype)
mkldnn.MaxPooling_F32.do_backward(gy[0], x[0], gx, self.indexes,
self.sy, self.sx, self.ph,
self.pd, self.pw, self.pr,
self.kh, self.kw)
return gx,
else:
n, c, out_h, out_w = gy[0].shape
h, w = x[0].shape[2:]
kh, kw = self.kh, self.kw
gcol = numpy.zeros((n * c * out_h * out_w * kh * kw),
dtype=x[0].dtype)
indexes = self.indexes.flatten()
indexes += numpy.arange(0, indexes.size * kh * kw, kh * kw)
gcol[indexes] = gy[0].ravel()
gcol = gcol.reshape(n, c, out_h, out_w, kh, kw)
gcol = numpy.swapaxes(gcol, 2, 4)
gcol = numpy.swapaxes(gcol, 3, 5)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h,
w)
return gx,
def test_col2im_consistency(self):
col = conv.im2col_cpu(self.x, 3, 3, 2, 2, 2, 2, dy=2, dx=2)
h, w = self.x.shape[2:]
im_cpu = conv.col2im_cpu(col, 2, 2, 2, 2, h, w, dy=2, dx=2)
im_gpu = conv.col2im_gpu(
cuda.to_gpu(col), 2, 2, 2, 2, h, w, dy=2, dx=2)
testing.assert_allclose(im_cpu, im_gpu.get())
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
if not all([isinstance(i, numpy.ndarray) for i in inputs]):
if b is not None:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}, type(b): {2}'
.format(type(W), type(x), type(b)))
else:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}'
.format(type(W), type(x)))
kh, kw = W.shape[2:]
_, _, h, w = x.shape
gcol = numpy.tensordot(W, x, (0, 1)).astype(x.dtype, copy=False)
# - k, m, n: shape of out_channel
# - b: number of inputs
# - h, w: height and width of kernels
# k, m, n, b, h, w -> b, k, m, n, h, w
gcol = numpy.rollaxis(gcol, 3)
if self.outh is None:
self.outh = conv.get_deconv_outsize(h, kh, self.sy, self.ph)
assert self.outh > 0, 'Height in the output should be positive.'
if self.outw is None:
self.outw = conv.get_deconv_outsize(w, kw, self.sx, self.pw)
assert self.outw > 0, 'Width in the output should be positive.'
y = conv.col2im_cpu(
gcol, self.sy, self.sx, self.ph, self.pw, self.outh, self.outw)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return y,
def forward_cpu(self, x):
self._in_dtype = x[0].dtype
n, c, h, w = x[0].shape
if self.outh is None:
self.outh = conv.get_deconv_outsize(h,
self.kh,
self.sy,
self.ph,
cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w,
self.kw,
self.sx,
self.pw,
cover_all=self.cover_all)
up_y = numpy.zeros((n, c, self.outh, self.outw), dtype=self._in_dtype)
up_y = conv.im2col_cpu(up_y,
self.kh,
self.kw,
self.sy,
self.sx,
self.ph,
self.pw,
cover_all=self.cover_all).transpose(
0, 1, 4, 5, 2, 3)
colh, colw = up_y.shape[2:4]
up_y = up_y.reshape(-1, self.kh * self.kw)
indexes = self.indexes.ravel()
up_y[numpy.arange(len(indexes)), indexes] = x[0].ravel()
up_y = up_y.reshape(n, c, colh, colw, self.kh, self.kw)
up_y = conv.col2im_cpu(up_y.transpose(0, 1, 4, 5, 2, 3), self.sy,
self.sx, self.ph, self.pw, self.outh, self.outw)
return up_y,
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
if not type_check.same_types(*inputs):
if b is not None:
raise ValueError(
'numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}, type(b): {2}'.format(
type(W), type(x), type(b)))
else:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}'.format(
type(W), type(x)))
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(gy, self.col,
((0, 2, 3), (0, 4, 5))).astype(W.dtype,
copy=False)
gcol = numpy.tensordot(W, gy, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def template(col_shape=(2, 5, 3, 3, 3, 4),
col_order=col_chainer_order,
im_order=OrderNHWC,
ksize=(3, 3),
padding=(1, 1),
stride=(1, 1),
description: str = ""):
col = Variable(col_shape, col_order)
op = Col2Im(None, ksize, stride, padding)
im, = op(col)
im = im.change_order(im_order)
vcol = np.random.rand(*(col.shape_dict[a]
for a in col_chainer_order.axes)).astype(
np.float32)
h1 = get_deconv_outsize(col.shape_dict[Axis.H], op.KH, op.SH, op.PH)
w1 = get_deconv_outsize(col.shape_dict[Axis.W], op.KW, op.SW, op.PW)
vim = col2im_cpu(vcol, op.SH, op.SW, op.PH, op.PW, h1, w1)
vcol = vcol.transpose(
[col_chainer_order.axes_dict[a] for a in col_order.axes])
vim = vim.transpose([OrderNCHW.axes_dict[a] for a in im_order.axes])
generate_kernel_test_case(
description=f"Col2Im {description}",
backend=["webgpu", "webgl", "webassembly"],
graph=Graph([col], [im]),
inputs={col: vcol},
expected={im: vim},
)
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(gy, self.col,
((0, 2, 3), (0, 4, 5))).astype(W.dtype,
copy=False)
if not self.requires_x_grad:
gx = None
else:
gcol = numpy.tensordot(W, gy, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol,
self.sy,
self.sx,
self.ph,
self.pw,
h,
w,
dy=self.dy,
dx=self.dx)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def forward_cpu(self, x):
n, c, h, w = x[0].shape
if self.outh is None:
self.outh = conv.get_deconv_outsize(h,
self.kh,
self.sy,
self.ph,
cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w,
self.kw,
self.sx,
self.pw,
cover_all=self.cover_all)
up_y = numpy.zeros((n, c, self.outh, self.outw), dtype=numpy.float32)
up_y = conv.im2col_cpu(up_y,
self.kh,
self.kw,
self.sy,
self.sx,
self.ph,
self.pw,
cover_all=self.cover_all)
for n in six.moves.range(up_y.shape[0]):
for c in six.moves.range(up_y.shape[1]):
for oy in six.moves.range(up_y.shape[4]):
for ox in six.moves.range(up_y.shape[5]):
ky = self.indexes[n, c, oy, ox] // up_y.shape[3]
kx = self.indexes[n, c, oy, ox] % up_y.shape[3]
up_y[n, c, ky, kx, oy, ox] = x[0][n, c, oy, ox]
up_y = conv.col2im_cpu(up_y, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
return up_y,
def backward_cpu(self, x, gy):
h, w = self._in_shape[2:]
gcol = numpy.tile(gy[0][:, :, None, None],
(1, 1, self.kh, self.kw, 1, 1))
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
gx /= self.kh * self.kw
return gx,
def backward_cpu(self, x, gy):
h, w = x[0].shape[2:]
gcol = numpy.tile(gy[0][:, :, numpy.newaxis, numpy.newaxis],
(1, 1, self.kh, self.kw, 1, 1))
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
gx /= self.kh * self.kw
return gx,
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
if not type_check.same_types(*inputs):
if b is not None:
raise ValueError(
'numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}, type(b): {2}'.format(
type(W), type(x), type(b)))
else:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}'.format(
type(W), type(x)))
kh, kw = W.shape[2:]
_, _, h, w = x.shape
gcol = numpy.tensordot(W, x, (0, 1)).astype(x.dtype, copy=False)
# - k, m, n: shape of out_channel
# - b: number of inputs
# - h, w: height and width of kernels
# k, m, n, b, h, w -> b, k, m, n, h, w
gcol = numpy.rollaxis(gcol, 3)
if self.outh is None:
self.outh = conv.get_deconv_outsize(h, kh, self.sy, self.ph)
assert self.outh > 0, 'Height in the output should be positive.'
if self.outw is None:
self.outw = conv.get_deconv_outsize(w, kw, self.sx, self.pw)
assert self.outw > 0, 'Width in the output should be positive.'
y = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return y,
def backward_cpu(self, x, gy):
h, w = self._in_shape[2:]
gcol = numpy.tile(gy[0][:, :, None, None],
(1, 1, self.kh, self.kw, 1, 1))
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
gx /= self.kh * self.kw
return gx,
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
if self.bcoeffs is not None:
olen, ilen, hlen, wlen = W.shape
if self.coeffs is None:
self.coeffs = numpy.ones(ilen)
coeffs = numpy.copy(self.bcoeffs)
coeffs = numpy.expand_dims(coeffs, 1)
coeffs = numpy.expand_dims(coeffs, 1)
coeffs = numpy.expand_dims(coeffs, 0)
coeffs = numpy.broadcast_to(coeffs, W.shape)
self.mW = numpy.asarray(coeffs,numpy.float32).reshape(W.shape)
if self.ocoeffs is not None:
coeffs = numpy.copy(self.ocoeffs)
self.mb = numpy.asarray(coeffs,numpy.float32)
W = self.M*W
b = inputs[2] if len(inputs) == 3 else None
if not type_check.same_types(*inputs):
if b is not None:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}, type(b): {2}'
.format(type(W), type(x), type(b)))
else:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}'
.format(type(W), type(x)))
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(
gy, self.col, ((0, 2, 3), (0, 4, 5))).astype(W.dtype, copy=False)
if not self.requires_x_grad:
gx = None
else:
gcol = numpy.tensordot(W, gy, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw,
h, w)
if hasattr(self,'mW'):
gW = self.mW * gW
if hasattr(self,'mb'):
xp = cuda.get_array_module(*x)
gW = xp.broadcast_to(
xp.expand_dims(xp.expand_dims(xp.expand_dims(self.mb,1),1),1)
,gW.shape) * gW
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
if hasattr(self,'mb'):
gb = self.mb * gb
return gx, gW, gb
def backward_cpu(self, x, gy):
if self.gb is not None:
self.gb += gy[0].sum(axis=(0, 2, 3))
self.gW += numpy.tensordot(gy[0], self.col, ([0, 2, 3], [0, 4, 5]))
gcol = numpy.tensordot(self.W, gy[0], (0, 1))
gcol = numpy.rollaxis(gcol, 3)
h, w = x[0].shape[2:]
return conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w),
def backward_cpu(self, x, gy):
if self.gb is not None:
self.gb += gy[0].sum(axis=(0, 2, 3))
self.gW += numpy.tensordot(gy[0], self.col, ([0, 2, 3], [0, 4, 5]))
gcol = numpy.tensordot(self.W, gy[0], (0, 1))
gcol = numpy.rollaxis(gcol, 3)
h, w = x[0].shape[2:]
return conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w),
def _forward_cpu_core(self, x, W, b):
gcol = numpy.tensordot(W, x, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
y = conv.col2im_cpu(
gcol, self.sy, self.sx, self.ph, self.pw, self.outh, self.outw,
dy=self.dy, dx=self.dx)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return y,
def _forward_cpu_core(self, x, W, b):
gcol = numpy.tensordot(W, x, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
y = conv.col2im_cpu(
gcol, self.sy, self.sx, self.ph, self.pw, self.outh, self.outw,
dy=self.dy, dx=self.dx)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return y,
def forward_cpu(self, gy):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(gy)):
return self._forward_ideep(gy)
h, w = self._in_shape[2:]
gcol = numpy.tile(gy[0][:, :, None, None],
(1, 1, self.kh, self.kw, 1, 1))
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
gx /= self.kh * self.kw
return gx,
def forward_cpu(self, gy):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(gy)):
return self._forward_ideep(gy)
h, w = self._in_shape[2:]
gcol = numpy.tile(gy[0][:, :, None, None],
(1, 1, self.kh, self.kw, 1, 1))
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
gx /= self.kh * self.kw
return gx,
def backward_cpu(self, x, gy):
n, c, out_h, out_w = gy[0].shape
h, w = x[0].shape[2:]
gcol = numpy.zeros(
(n, c, self.kh, self.kw, out_h, out_w), dtype=numpy.float32)
# TODO(beam2d): Make it fast
gcol_r = numpy.rollaxis(gcol.reshape(n, c, -1, out_h, out_w), 2)
for i in numpy.ndindex(n, c, out_h, out_w):
gcol_r[self.indexes[i]][i] = gy[0][i]
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
return gx,
def forward(self, x):
h, w = x[0].shape[2:]
if self.outh is None:
self.outh = conv.get_deconv_outsize(h, self.kh, self.sy, self.ph, cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w, self.kw, self.sx, self.pw, cover_all=self.cover_all)
xp = cuda.get_array_module(*x)
col = xp.tile(x[0][:, :, xp.newaxis, xp.newaxis], (1, 1, self.kh, self.kw, 1, 1))
if isinstance(x[0], cuda.ndarray):
y = conv.col2im_gpu(col, self.sy, self.sx, self.ph, self.pw, self.outh, self.outw)
else:
y = conv.col2im_cpu(col, self.sy, self.sx, self.ph, self.pw, self.outh, self.outw)
return (y,)
def backward_cpu(self, x, gy):
n, c, out_h, out_w = gy[0].shape
h, w = x[0].shape[2:]
gcol = numpy.zeros((n, c, self.kh, self.kw, out_h, out_w),
dtype=numpy.float32)
# TODO(beam2d): Make it fast
gcol_r = numpy.rollaxis(gcol.reshape(n, c, -1, out_h, out_w), 2)
for i in numpy.ndindex(n, c, out_h, out_w):
gcol_r[self.indexes[i]][i] = gy[0][i]
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
return gx,
def forward(self, x):
h, w = x[0].shape[2:]
n = x[0].shape[0]
c = x[0].shape[1]
indexes = x[1]
if self.outh is None:
self.outh = conv.get_deconv_outsize(h,
self.kh,
self.sy,
self.ph,
cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w,
self.kw,
self.sx,
self.pw,
cover_all=self.cover_all)
xp = cuda.get_array_module(*x)
col = xp.tile(x[0][:, :, xp.newaxis, xp.newaxis],
(1, 1, self.kh, self.kw, 1, 1))
# NOTE(hvy): Take indexes(Switches) into account
# TODO(hvy): Remove the loops and make it efficient
y = xp.zeros_like(col)
if isinstance(x[0], cuda.ndarray):
indexes = cuda.cupy.asnumpy(indexes)
for n_i in range(n):
for c_i in range(c):
for r in range(h):
for c in range(w):
index = indexes[n_i][c_i][r][c]
if index < self.kw:
y[n_i][c_i].T[c][r][index][0] = col[n_i][c_i].T[c][
r][index][0]
else:
y[n_i][c_i].T[c][r][
index %
self.kw][1] = col[n_i][c_i].T[c][r][index %
self.kw][1]
if isinstance(x[0], cuda.ndarray):
y = conv.col2im_gpu(y, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
else:
y = conv.col2im_cpu(y, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
return y,
def forward_cpu(self, x):
n, c, h, w = x[0].shape
gcol = numpy.tensordot(self.W, x[0], (0, 1))
# k, m, n, b, h, w
gcol = numpy.rollaxis(gcol, 3)
# b, k, m, n, h, w
h_ = get_deconv_outsize(h, self.kh, self.sy, self.ph)
w_ = get_deconv_outsize(w, self.kw, self.sx, self.pw)
y = conv.col2im_cpu(
gcol, self.sy, self.sx, self.ph, self.pw, h_, w_)
# b, k, h, w
if self.b is not None:
y += self.b.reshape(1, self.b.size, 1, 1)
return y,
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(gy, self.col, ((0, 2, 3), (0, 4, 5)))
gcol = numpy.tensordot(W, gy, (0, 1))
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
if len(inputs) == 3:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
else:
return gx, gW
def backward(self, inputs, grad_outputs):
x, = inputs
xp = cuda.get_array_module(x)
gy, = grad_outputs
n, _, out_h, out_w = gy.shape
_, c, h, w = x.shape
gy = gy.reshape(n, c, self.kh, self.kw, out_h, out_w)
if xp == numpy:
gx = col2im_cpu(
gy, self.sy, self.sx, self.ph, self.pw, h, w, self.dy, self.dx)
else:
gx = col2im_gpu(
gy, self.sy, self.sx, self.ph, self.pw, h, w, self.dy, self.dx)
return gx,
def backward(self, inputs, grad_outputs):
x, = inputs
xp = cuda.get_array_module(x)
gy, = grad_outputs
n, _, out_h, out_w = gy.shape
_, c, h, w = x.shape
gy = gy.reshape(n, c, self.kh, self.kw, out_h, out_w)
if xp == numpy:
gx = col2im_cpu(gy, self.sy, self.sx, self.ph, self.pw, h, w,
self.dy, self.dx)
else:
gx = col2im_gpu(gy, self.sy, self.sx, self.ph, self.pw, h, w,
self.dy, self.dx)
return gx,
def forward_cpu(self, x):
n, c, out_h, out_w = x[0].shape
gcol = numpy.zeros((n, c, self.kh, self.kw, out_h, out_w),
dtype=numpy.float32)
# TODO(beam2d): Make it fast
gcol_r = numpy.rollaxis(gcol.reshape(n, c, -1, out_h, out_w), 2)
for i in numpy.ndindex(n, c, out_h, out_w):
# gcol_r[self.indexes[i]][i] = x[0][i]
for j in xrange(gcol_r.shape[0]):
gcol_r[j][i] = x[0][i]
y = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, self.h,
self.w)
return y,
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(gy, self.col, ((0, 2, 3), (0, 4, 5)))
gcol = numpy.tensordot(W, gy, (0, 1))
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
if len(inputs) == 3:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
else:
return gx, gW
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
Wb = numpy.where(W >= 0, 1, -1).astype(numpy.float32, copy=False)
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(gy, self.col, ((0, 2, 3), (0, 4, 5)))
gcol = numpy.tensordot(Wb, gy, (0, 1))
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(gy, self.col, ((0, 2, 3), (0, 4, 5))).astype(W.dtype, copy=False)
gcol = numpy.tensordot(W, gy, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w, dy=self.dy, dx=self.dx)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
Wb = numpy.where(W>=0, 1, -1).astype(numpy.float32, copy=False)
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
h, w = x.shape[2:]
gW = numpy.tensordot(gy, self.col, ((0, 2, 3), (0, 4, 5)))
gcol = numpy.tensordot(Wb, gy, (0, 1))
gcol = numpy.rollaxis(gcol, 3)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def backward_cpu(self, x, gy):
n, c, out_h, out_w = gy[0].shape
h, w = x[0].shape[2:]
kh, kw = self.kh, self.kw
gcol = numpy.zeros((n * c * out_h * out_w * kh * kw), dtype=x[0].dtype)
indexes = self.indexes.ravel()
indexes += numpy.arange(0, indexes.size * kh * kw, kh * kw)
gcol[indexes] = gy[0].ravel()
gcol = gcol.reshape(n, c, out_h, out_w, kh, kw)
gcol = numpy.swapaxes(gcol, 2, 4)
gcol = numpy.swapaxes(gcol, 3, 5)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
return gx,
def backward(self, inputs, grad_outputs):
x, W = inputs[:2]
xp = cuda.get_array_module(*x)
W = xp.where(W >= 0, 1, -1).astype(numpy.float32, copy=False)
W = self.M * W
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
h, w = x.shape[2:]
xp = cuda.get_array_module(*x)
B, C, KY, KX, IY, IX = self.col.shape
D = W.shape[0]
# (B, C*D, IY, IX) -> (C, D, B*IY*IX, D)
gy_ = gy.reshape((B, C, D, IY * IX)).transpose(1, 2, 0, 3) \
.reshape((C, D, B * IY * IX))
c_ = self.col.transpose(1, 0, 4, 5, 2, 3) \
.reshape((C, B * IY * IX, KY * KX))
# (C, D, B*IY*IX), (C, B*IY*IX, KY*KX) -> (C, D, KY*KX)
gW_ = _matmul(gy_, c_, xp)
gW = gW_.reshape((C, D, KY, KX)).transpose(1, 0, 2, 3)
gW = gW.astype(W.dtype, copy=False)
w_ = W.transpose(1, 2, 3, 0).reshape((C, KY * KX, D))
# (C, KY*KX, D), (C, D, B*IY*IX) -> (C, KY*KX, B*IY*IX)
gcol = _matmul(w_, gy_, xp).reshape((C, KY, KX, B, IY, IX))
gcol = gcol.astype(x.dtype, copy=False)
gcol = xp.rollaxis(gcol, 3)
if xp is numpy:
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h,
w)
else:
gx = conv.col2im_gpu(gcol, self.sy, self.sx, self.ph, self.pw, h,
w)
if b is None:
return gx, gW
else:
gy = xp.rollaxis(gy, 1, 4)
gb = gy.sum(axis=(0, 1, 2))
return gx, gW, gb
def backward_cpu(self, x, gy):
n, c, out_h, out_w = gy[0].shape
h, w = self._in_shape[2:]
kh, kw = self.kh, self.kw
gcol = numpy.zeros(
(n * c * out_h * out_w * kh * kw), dtype=self._in_dtype)
indexes = self.indexes.flatten()
indexes += numpy.arange(0, indexes.size * kh * kw, kh * kw)
gcol[indexes] = gy[0].ravel()
gcol = gcol.reshape(n, c, out_h, out_w, kh, kw)
gcol = numpy.swapaxes(gcol, 2, 4)
gcol = numpy.swapaxes(gcol, 3, 5)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
return gx,
def backward_cpu(self, x, gy):
if mkld.enable_avg_poolingF((x, gy)):
n, c, h, w = x[0].shape
gx = numpy.empty((n, c, h, w), dtype=x[0].dtype)
mkldnn.AvgPooling_F32.do_backward(gy[0], x[0], gx, self.sy,
self.sx, self.ph, self.pd,
self.pw, self.pr, self.kh,
self.kw)
return gx,
else:
h, w = x[0].shape[2:]
gcol = numpy.tile(gy[0][:, :, None, None],
(1, 1, self.kh, self.kw, 1, 1))
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h,
w)
gx /= self.kh * self.kw
return gx,
def forward(self, x):
h, w = x[0].shape[2:]
if self.outh is None:
self.outh = conv.get_deconv_outsize(
h, self.kh, self.sy, self.ph, cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(
w, self.kw, self.sx, self.pw, cover_all=self.cover_all)
xp = cuda.get_array_module(*x)
col = xp.tile(x[0][:, :, None, None],
(1, 1, self.kh, self.kw, 1, 1))
if xp is numpy:
y = conv.col2im_cpu(col, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
else:
y = conv.col2im_gpu(col, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
return y,
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
_, _, h, w = x.shape
gcol = numpy.tensordot(W, x, (0, 1))
# - k, m, n: shape of out_channel
# - b: number of inputs
# - h, w: height and width of kernels
# k, m, n, b, h, w -> b, k, m, n, h, w
gcol = numpy.rollaxis(gcol, 3)
if self.outh is None:
self.outh = conv.get_deconv_outsize(h, kh, self.sy, self.ph)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w, kw, self.sx, self.pw)
y = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return y,
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
_, _, h, w = x.shape
gcol = numpy.tensordot(W, x, (0, 1))
# - k, m, n: shape of out_channel
# - b: number of inputs
# - h, w: height and width of kernels
# k, m, n, b, h, w -> b, k, m, n, h, w
gcol = numpy.rollaxis(gcol, 3)
if self.outh is None:
self.outh = conv.get_deconv_outsize(h, kh, self.sy, self.ph)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w, kw, self.sx, self.pw)
y = conv.col2im_cpu(
gcol, self.sy, self.sx, self.ph, self.pw, self.outh, self.outw)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return y,
def backward(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
h, w = x.shape[2:]
xp = cuda.get_array_module(*x)
B, C, KY, KX, IY, IX = self.col.shape
D = W.shape[0]
# (B, C*D, IY, IX) -> (C, D, B*IY*IX, D)
gy_ = gy.reshape((B, C, D, IY * IX)).transpose(1, 2, 0, 3) \
.reshape((C, D, B * IY * IX))
c_ = self.col.transpose(1, 0, 4, 5, 2, 3) \
.reshape((C, B * IY * IX, KY * KX))
# (C, D, B*IY*IX), (C, B*IY*IX, KY*KX) -> (C, D, KY*KX)
gW_ = _matmul(gy_, c_, xp)
gW = gW_.reshape((C, D, KY, KX)).transpose(1, 0, 2, 3)
gW = gW.astype(W.dtype, copy=False)
w_ = W.transpose(1, 2, 3, 0).reshape((C, KY * KX, D))
# (C, KY*KX, D), (C, D, B*IY*IX) -> (C, KY*KX, B*IY*IX)
gcol = _matmul(w_, gy_, xp).reshape((C, KY, KX, B, IY, IX))
gcol = gcol.astype(x.dtype, copy=False)
gcol = xp.rollaxis(gcol, 3)
if xp is numpy:
gx = conv.col2im_cpu(gcol, self.sy, self.sx,
self.ph, self.pw, h, w)
else:
gx = conv.col2im_gpu(gcol, self.sy, self.sx,
self.ph, self.pw, h, w)
if b is None:
return gx, gW
else:
gy = xp.rollaxis(gy, 1, 4)
gb = gy.sum(axis=(0, 1, 2))
return gx, gW, gb
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
_, _, h, w = x.shape
gcol = numpy.tensordot(W, x, (0, 1)).astype(x.dtype, copy=False)
# - k, m, n: shape of out_channel
# - b: number of inputs
# - h, w: height and width of kernels
# k, m, n, b, h, w -> b, k, m, n, h, w
gcol = numpy.rollaxis(gcol, 3)
if self.outh is None:
self.outh = conv.get_deconv_outsize(h, kh, self.sy, self.ph)
assert self.outh > 0, "Height in the output should be positive."
if self.outw is None:
self.outw = conv.get_deconv_outsize(w, kw, self.sx, self.pw)
assert self.outw > 0, "Width in the output should be positive."
y = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, self.outh, self.outw)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return (y,)
def forward_cpu(self, gy):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(gy)):
return self._forward_ideep(gy)
n, c, out_h, out_w = gy[0].shape
h, w = self._in_shape[2:]
kh, kw = self.kh, self.kw
gcol = numpy.zeros(
(n * c * out_h * out_w * kh * kw), dtype=self._in_dtype)
indexes = self.indexes.flatten()
indexes += numpy.arange(0, indexes.size * kh * kw, kh * kw)
gcol[indexes] = gy[0].ravel()
gcol = gcol.reshape(n, c, out_h, out_w, kh, kw)
gcol = numpy.swapaxes(gcol, 2, 4)
gcol = numpy.swapaxes(gcol, 3, 5)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
return gx,
def forward_cpu(self, gy):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(gy)):
return self._forward_ideep(gy)
n, c, out_h, out_w = gy[0].shape
h, w = self._in_shape[2:]
kh, kw = self.kh, self.kw
gcol = numpy.zeros((n * c * out_h * out_w * kh * kw),
dtype=self._in_dtype)
indexes = self.indexes.ravel() + numpy.arange(
0, self.indexes.size * kh * kw, kh * kw)
gcol[indexes] = gy[0].ravel()
gcol = gcol.reshape(n, c, out_h, out_w, kh, kw)
gcol = numpy.swapaxes(gcol, 2, 4)
gcol = numpy.swapaxes(gcol, 3, 5)
gx = conv.col2im_cpu(gcol, self.sy, self.sx, self.ph, self.pw, h, w)
return gx,
def forward(self, x):
h, w = x[0].shape[2:]
if self.outh is None:
self.outh = conv.get_deconv_outsize(h,
self.kh,
self.sy,
self.ph,
cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(w,
self.kw,
self.sx,
self.pw,
cover_all=self.cover_all)
xp = cuda.get_array_module(*x)
col = xp.tile(x[0][:, :, xp.newaxis, xp.newaxis],
(1, 1, self.kh, self.kw, 1, 1))
if isinstance(x[0], cuda.ndarray):
y = conv.col2im_gpu(col, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
else:
y = conv.col2im_cpu(col, self.sy, self.sx, self.ph, self.pw,
self.outh, self.outw)
return y,
def forward_cpu(self, inputs):
self.retain_inputs((0, 1)) # only retain x and W
if len(inputs) == 2:
(x, W), b = inputs, None
else:
x, W, b = inputs
self._calc_out_size(x, W)
gcol = numpy.tensordot(W, x, (0, 1)).astype(x.dtype, copy=False)
gcol = numpy.rollaxis(gcol, 3)
y = conv.col2im_cpu(gcol,
self.sy,
self.sx,
self.ph,
self.pw,
self.outh,
self.outw,
dy=self.dy,
dx=self.dx)
# b, k, h, w
if b is not None:
y += b.reshape(1, b.size, 1, 1)
return y,
def forward_cpu(self, x):
self._in_dtype = x[0].dtype
n, c, h, w = x[0].shape
if self.outh is None:
self.outh = conv.get_deconv_outsize(
h, self.kh, self.sy, self.ph, cover_all=self.cover_all)
if self.outw is None:
self.outw = conv.get_deconv_outsize(
w, self.kw, self.sx, self.pw, cover_all=self.cover_all)
up_y = numpy.zeros((n, c, self.outh, self.outw), dtype=self._in_dtype)
up_y = conv.im2col_cpu(
up_y, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all).transpose(0, 1, 4, 5, 2, 3)
colh, colw = up_y.shape[2:4]
up_y = up_y.reshape(-1, self.kh * self.kw)
indexes = self.indexes.ravel()
up_y[numpy.arange(len(indexes)), indexes] = x[0].ravel()
up_y = up_y.reshape(n, c, colh, colw, self.kh, self.kw)
up_y = conv.col2im_cpu(
up_y.transpose(0, 1, 4, 5, 2, 3), self.sy, self.sx, self.ph,
self.pw, self.outh, self.outw)
return up_y,
def test_col2im_consistency(self):
col = conv.im2col_cpu(self.x, 3, 3, 2, 2, 1, 1)
h, w = self.x.shape[2:]
im_cpu = conv.col2im_cpu(col, 2, 2, 1, 1, h, w)
im_gpu = conv.col2im_gpu(cuda.to_gpu(col), 2, 2, 1, 1, h, w)
gradient_check.assert_allclose(im_cpu, im_gpu.get())
def _col2im(x, *args, **kwargs):
if isinstance(x, numpy.ndarray):
return col2im_cpu(x, *args, **kwargs)
return col2im_gpu(x, *args, **kwargs)