def _forward_cpu_core(self, x, gy):
col = conv.im2col_cpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
gW = numpy.tensordot(gy, col, ((0, 2, 3), (0, 4, 5))
).astype(self.W_dtype, copy=False)
return gW
def _forward_cpu_core(self, x, gy):
if self._use_ideep:
return self._forward_ideep(x, gy)
# NumPy raises an error when the array is not contiguous.
# See: https://github.com/chainer/chainer/issues/2744
# TODO(niboshi): Remove this code when NumPy is fixed.
if (not (gy.flags.c_contiguous or gy.flags.f_contiguous)
and 1 in gy.shape):
gy = numpy.ascontiguousarray(gy)
col = conv.im2col_cpu(x,
self.kh,
self.kw,
self.sy,
self.sx,
self.ph,
self.pw,
cover_all=self.cover_all,
dy=self.dy,
dx=self.dx)
gW = numpy.tensordot(gy, col,
((0, 2, 3), (0, 4, 5))).astype(self.W_dtype,
copy=False)
return gW,
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, x):
self.col = conv.im2col_cpu(x[0], self.kh, self.kw, self.sy, self.sx,
self.ph, self.pw)
y = numpy.tensordot(self.col, self.W, ([1, 2, 3], [1, 2, 3]))
if self.b is not None:
y += self.b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
out_c, input_c, kh, kw = W.shape
n, c, h, w = x.shape
"""
For mkldnn backend, only support float32 for x and W
"""
if mkld.enable_convF(inputs):
out_h = conv.get_conv_outsize(h, kh, self.sy, self.ph, cover_all=self.cover_all)
assert out_h > 0, 'Height in the output should be positive.'
out_w = conv.get_conv_outsize(w, kw, self.sx, self.pw, cover_all=self.cover_all)
assert out_w > 0, 'Width in the output should be positive.'
self.pd = self.sy*(out_h-1) + kh - h - self.ph
self.pr = self.sx*(out_w-1) + kw - w - self.pw
y = numpy.empty(shape=(n, out_c, out_h, out_w), dtype=x.dtype)
if b is not None:
mkldnn.Convolution2D_F32.do_forward(x, W, b, y, kh, kw, self.sx, self.sy, self.ph, self.pw, self.pd, self.pr)
else:
mkldnn.Convolution2D_F32.do_forward(x, W, y, kh, kw, self.sx, self.sy, self.ph, self.pw, self.pd, self.pr)
return y,
else:
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
# print "%f, %f" %(cos_module.cos_func(0.5), sin_module.sin_func(0.5))
y = numpy.tensordot(
self.col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
y = numpy.rollaxis(y, 3, 1)
return y,
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 forward(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
xp = cuda.get_array_module(*x)
if xp is numpy:
self.col = conv.im2col_cpu(x, kh, kw, self.sy, self.sx, self.ph,
self.pw)
else:
self.col = conv.im2col_gpu(x, kh, kw, self.sy, self.sx, self.ph,
self.pw)
B, C, KY, KX, IY, IX = self.col.shape
D = W.shape[0] # (D, C, KY, KX)
c_ = self.col.transpose(1, 0, 4, 5, 2, 3) \
.reshape((C, B * IY * IX, KY * KX))
w_ = W.transpose(1, 2, 3, 0).reshape((C, KY * KX, D))
# (C, B*IY*IX, KY*KX), (C, KY*KX, D)-> (C, B*IY*IX, D)
y = _matmul(c_, w_, xp).astype(x.dtype, copy=False)
# (C, B*IY*IX, D) -> (B, C*D, IY, IX)
y = y.reshape((C, B, IY * IX, D)).transpose(1, 0, 3, 2) \
.reshape((B, C * D, IY, IX))
if b is not None:
y += b[None, :, None, None]
return y,
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 forward_cpu(self, x):
self.col = conv.im2col_cpu(
x[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw)
y = numpy.tensordot(self.col, self.W, ([1, 2, 3], [1, 2, 3]))
if self.b is not None:
y += self.b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, x):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(x)):
return self._forward_ideep(x)
self._in_shape = x[0].shape
self._in_dtype = x[0].dtype
col = conv.im2col_cpu(x[0],
self.kh,
self.kw,
self.sy,
self.sx,
self.ph,
self.pw,
pval=-float('inf'),
cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = col.reshape(n, c, kh * kw, out_h, out_w)
# We select maximum twice, since the implementation using numpy.choose
# hits its bug when kh * kw >= 32.
self.indexes = col.argmax(axis=2)
y = col.max(axis=2)
return y,
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(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
xp = cuda.get_array_module(*x)
if xp is numpy:
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw)
else:
self.col = conv.im2col_gpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw)
B, C, KY, KX, IY, IX = self.col.shape
D = W.shape[0] # (D, C, KY, KX)
c_ = self.col.transpose(1, 0, 4, 5, 2, 3) \
.reshape((C, B * IY * IX, KY * KX))
w_ = W.transpose(1, 2, 3, 0).reshape((C, KY * KX, D))
# (C, B*IY*IX, KY*KX), (C, KY*KX, D)-> (C, B*IY*IX, D)
y = _matmul(c_, w_, xp).astype(x.dtype, copy=False)
# (C, B*IY*IX, D) -> (B, C*D, IY, IX)
y = y.reshape((C, B, IY * IX, D)).transpose(1, 0, 3, 2) \
.reshape((B, C * D, IY, IX))
if b is not None:
y += b[None, :, None, None]
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]
kh, kw = W.shape[2:]
col = conv.im2col_cpu(
gy, kh, kw, self.sy, self.sx, self.ph, self.pw)
gW = numpy.tensordot(
x, col, ([0, 2, 3], [0, 4, 5])).astype(W.dtype, copy=False)
gx = numpy.tensordot(
col, W, ([1, 2, 3], [1, 2, 3])).astype(x.dtype, copy=False)
gx = numpy.rollaxis(gx, 3, 1)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def forward(self, inputs):
x, = inputs
xp = cuda.get_array_module(x)
if xp == numpy:
y = im2col_cpu(x,
self.kh,
self.kw,
self.sy,
self.sx,
self.ph,
self.pw,
cover_all=self.cover_all,
dy=self.dy,
dx=self.dx)
else:
y = im2col_gpu(x,
self.kh,
self.kw,
self.sy,
self.sx,
self.ph,
self.pw,
cover_all=self.cover_all,
dy=self.dy,
dx=self.dx)
n, c, kh, kw, out_h, out_w = y.shape
y = y.reshape(n, c * kh * kw, out_h, out_w)
return y,
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
olen, ilen, hlen, wlen = W.shape
if self.coeffs is None:
self.coeffs = numpy.ones(ilen)
coeffs = numpy.copy(self.coeffs)
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)
M = numpy.asarray(coeffs,numpy.float32).reshape(W.shape)
self.M = M
W = self.M*W
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:]
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
y = numpy.tensordot(
self.col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, inputs):
self.retain_inputs((0, 1)) # retain only x and W
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:]
col = conv.im2col_cpu(x,
kh,
kw,
self.sy,
self.sx,
self.ph,
self.pw,
cover_all=self.cover_all,
dy=self.dy,
dx=self.dx)
y = numpy.tensordot(col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype,
copy=False)
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
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]
kh, kw = W.shape[2:]
col = conv.im2col_cpu(gy, kh, kw, self.sy, self.sx, self.ph, self.pw)
gW = numpy.tensordot(x, col, ([0, 2, 3], [0, 4, 5])).astype(W.dtype,
copy=False)
gx = numpy.tensordot(col, W, ([1, 2, 3], [1, 2, 3])).astype(x.dtype,
copy=False)
gx = numpy.rollaxis(gx, 3, 1)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
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:]
self.col = conv.im2col_cpu(x,
kh,
kw,
self.sy,
self.sx,
self.ph,
self.pw,
cover_all=self.cover_all,
dy=self.dy,
dx=self.dx)
y = numpy.tensordot(self.col, W,
((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, x):
self._in_shape = x[0].shape
self._in_dtype = x[0].dtype
col = conv.im2col_cpu(x[0], self.kh, self.kw, self.sy, self.sx,
self.ph, self.pw)
y = col.mean(axis=(2, 3))
return y,
def forward_cpu(self, x):
self._in_shape = x[0].shape
self._in_dtype = x[0].dtype
col = conv.im2col_cpu(x[0], self.kh, self.kw, self.sy, self.sx,
self.ph, self.pw)
y = col.mean(axis=(2, 3))
return y,
def backward_cpu(self, x, gy):
if self.gb is not None:
self.gb += gy[0].sum(axis=(0, 2, 3))
col = conv.im2col_cpu(
gy[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw)
self.gW += numpy.tensordot(x[0], col, ([0, 2, 3], [0, 4, 5]))
gx = numpy.tensordot(col, self.W, ([1, 2, 3], [1, 2, 3]))
gx = numpy.rollaxis(gx, 3, 1)
return gx,
def forward_cpu(self, gy):
gcol = conv.im2col_cpu(
gy[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = gcol.shape
gcol = gcol.transpose(0, 1, 4, 5, 2, 3).reshape(-1, kh * kw)
indexes = self.indexes.ravel()
gx = gcol[numpy.arange(len(indexes)), indexes]
return gx.reshape(n, c, out_h, out_w),
def forward_cpu(self, gy):
gcol = conv.im2col_cpu(
gy[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = gcol.shape
gcol = gcol.transpose(0, 1, 4, 5, 2, 3).reshape(-1, kh * kw)
indexes = self.indexes.ravel()
gx = gcol[numpy.arange(len(indexes)), indexes]
return gx.reshape(n, c, out_h, out_w),
def forward_cpu(self, inputs):
self.retain_inputs((0, 1))
x, gy = inputs
col = conv.im2col_cpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
gW = numpy.tensordot(
gy, col, ((0, 2, 3), (0, 4, 5))).astype(self.W_dtype, copy=False)
return gW,
def forward_cpu(self, x):
col = conv.im2col_cpu(
x[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
pval=-float('inf'), cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = numpy.rollaxis(col.reshape(n, c, kh * kw, out_h, out_w), 2)
self.indexes = col.argmax(axis=0)
y = self.indexes.choose(col)
return y,
def forward_cpu(self, x):
col = conv.im2col_cpu(
x[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
pval=-float('inf'), cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = col.reshape(n, c, kh * kw, out_h, out_w)
col = col.transpose(0, 1, 3, 4, 2).reshape(-1, kh * kw)
indexes = self.indexes.ravel()
col = col[numpy.arange(len(indexes)), indexes]
return col.reshape(n, c, out_h, out_w),
def forward_cpu(self, x):
col = conv.im2col_cpu(
x[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
pval=-float('inf'), cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = col.reshape(n, c, kh * kw, out_h, out_w)
col = col.transpose(0, 1, 3, 4, 2).reshape(-1, kh * kw)
indexes = self.indexes.ravel()
col = col[numpy.arange(len(indexes)), indexes]
return col.reshape(n, c, out_h, out_w),
def forward_cpu(self, inputs):
x, W = inputs[:2]
kh, kw = W.shape[2:]
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw)
y = numpy.tensordot(self.col, W, ((1, 2, 3), (1, 2, 3)))
if len(inputs) == 3:
b = inputs[2]
y += b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
self.col = conv.im2col_cpu(x, kh, kw, self.sy, self.sx, self.ph,
self.pw)
y = numpy.tensordot(self.col, W, ((1, 2, 3), (1, 2, 3)))
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
y = numpy.tensordot(self.col, W, ((1, 2, 3), (1, 2, 3)))
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
self.col = conv.im2col_cpu(x, kh, kw, self.sy, self.sx, self.ph,
self.pw)
Wb = numpy.where(W >= 0, 1, -1).astype(numpy.float32, copy=False)
y = numpy.tensordot(self.col, Wb, ((1, 2, 3), (1, 2, 3)))
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def _forward_cpu_core(self, x, W, b):
kh, kw = W.shape[2:]
col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
y = numpy.tensordot(
col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
y = numpy.rollaxis(y, 3, 1)
return y
def backward(self, x, gy):
if isinstance(gy[0], cuda.ndarray):
gcol = conv.im2col_gpu(
gy[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw, cover_all=self.cover_all
)
else:
gcol = conv.im2col_cpu(
gy[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw, cover_all=self.cover_all
)
gx = gcol.sum(axis=(2, 3))
return (gx,)
def forward_cpu(self, inputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
kh, kw = W.shape[2:]
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw)
Wb = numpy.where(W>=0, 1, -1).astype(numpy.float32, copy=False)
y = numpy.tensordot(self.col, Wb, ((1, 2, 3), (1, 2, 3)))
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, x):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(x)):
return self._forward_ideep(x)
self._in_shape = x[0].shape
self._in_dtype = x[0].dtype
col = conv.im2col_cpu(x[0], self.kh, self.kw, self.sy, self.sx,
self.ph, self.pw)
y = col.mean(axis=(2, 3))
return y,
def forward_cpu(self, x):
col = conv.im2col_cpu(
x[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
pval=-float('inf'), cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = col.reshape(n, c, kh * kw, out_h, out_w)
# We select maximum twice, since the implementation using numpy.choose
# hits its bug when kh * kw >= 32.
self.indexes = col.argmax(axis=2)
y = col.max(axis=2)
return y,
def forward_cpu(self, x):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(x)):
return self._forward_ideep(x)
self._in_shape = x[0].shape
self._in_dtype = x[0].dtype
col = conv.im2col_cpu(x[0], self.kh, self.kw, self.sy, self.sx,
self.ph, self.pw)
y = col.mean(axis=(2, 3))
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:]
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
y = numpy.tensordot(
self.col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def test_im2col_consistency(self):
col_cpu = conv.im2col_cpu(self.x, 3, 3, 2, 2, 2, 2, dy=2, dx=2)
col_gpu = conv.im2col_gpu(cuda.to_gpu(self.x),
3,
3,
2,
2,
2,
2,
dy=2,
dx=2)
testing.assert_allclose(col_cpu, col_gpu.get(), atol=0, rtol=0)
def forward_cpu(self, x):
col = conv.im2col_cpu(
x[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
pval=-float('inf'), cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = col.reshape(n, c, kh * kw, out_h, out_w)
# We select maximum twice, since the implementation using numpy.choose
# hits its bug when kh * kw >= 32.
self.indexes = col.argmax(axis=2)
y = col.max(axis=2)
return y,
def backward_cpu(self, x, gy):
gcol = conv.im2col_cpu(gy[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw, cover_all=self.cover_all)
gcol = gcol.transpose(0, 1, 4, 5, 2, 3)
n, c, oy, ox, ky, kx = gcol.shape
gcol = gcol.reshape((n, c, oy, ox, ky * kx))
gx = numpy.empty((n, c, oy, ox), dtype=x[0].dtype)
for n in six.moves.range(gcol.shape[0]):
for c in six.moves.range(gcol.shape[1]):
for oy in six.moves.range(gcol.shape[2]):
for ox in six.moves.range(gcol.shape[3]):
gx[n, c, oy, ox] = gcol[n, c, oy, ox][self.indexes[n, c, oy, ox]]
return (gx,)
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
gy = grad_outputs[0]
kh, kw = W.shape[2:]
col = conv.im2col_cpu(gy, kh, kw, self.sy, self.sx, self.ph, self.pw)
gW = numpy.tensordot(x, col, ([0, 2, 3], [0, 4, 5]))
gx = numpy.tensordot(col, W, ([1, 2, 3], [1, 2, 3]))
gx = numpy.rollaxis(gx, 3, 1)
if len(inputs) == 3:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
else:
return gx, gW
def forward(self, inputs):
x, = inputs
xp = cuda.get_array_module(x)
if xp == numpy:
y = im2col_cpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
else:
y = im2col_gpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
n, c, kh, kw, out_h, out_w = y.shape
y = y.reshape(n, c * kh * kw, out_h, out_w)
return y,
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
gy = grad_outputs[0]
kh, kw = W.shape[2:]
col = conv.im2col_cpu(
gy, kh, kw, self.sy, self.sx, self.ph, self.pw)
gW = numpy.tensordot(x, col, ([0, 2, 3], [0, 4, 5]))
gx = numpy.tensordot(col, W, ([1, 2, 3], [1, 2, 3]))
gx = numpy.rollaxis(gx, 3, 1)
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]
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
kh, kw = W.shape[2:]
col = conv.im2col_cpu(gy, kh, kw, self.sy, self.sx, self.ph, self.pw)
gW = numpy.tensordot(x, col, ([0, 2, 3], [0, 4, 5])).astype(W.dtype, copy=False)
gx = numpy.tensordot(col, W, ([1, 2, 3], [1, 2, 3])).astype(x.dtype, copy=False)
gx = numpy.rollaxis(gx, 3, 1)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def forward_cpu(self, inputs):
self.retain_inputs((0, 1))
x, gy = inputs
col = conv.im2col_cpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
# NumPy raises an error when the array is not contiguous.
# See: https://github.com/chainer/chainer/issues/2744
# TODO(niboshi): Remove this code when NumPy is fixed.
if (not (gy.flags.c_contiguous or gy.flags.f_contiguous) and
1 in gy.shape):
gy = numpy.ascontiguousarray(gy)
gW = numpy.tensordot(
gy, col, ((0, 2, 3), (0, 4, 5))).astype(self.W_dtype, copy=False)
return gW,
def backward_cpu(self, inputs, grad_outputs):
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
gy = grad_outputs[0]
kh, kw = W.shape[2:]
col = conv.im2col_cpu(gy, kh, kw, self.sy, self.sx, self.ph, self.pw)
gW = numpy.tensordot(x, col, ([0, 2, 3], [0, 4, 5])).astype(W.dtype,
copy=False)
gx = numpy.tensordot(col, W, ([1, 2, 3], [1, 2, 3])).astype(x.dtype,
copy=False)
gx = numpy.rollaxis(gx, 3, 1)
if b is None:
return gx, gW
else:
gb = gy.sum(axis=(0, 2, 3))
return gx, gW, gb
def _forward_cpu_core(self, x, gy):
if self._use_ideep:
return self._forward_ideep(x, gy)
# NumPy raises an error when the array is not contiguous.
# See: https://github.com/chainer/chainer/issues/2744
# TODO(niboshi): Remove this code when NumPy is fixed.
if (not (gy.flags.c_contiguous or gy.flags.f_contiguous) and
1 in gy.shape):
gy = numpy.ascontiguousarray(gy)
col = conv.im2col_cpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
gW = numpy.tensordot(gy, col, ((0, 2, 3), (0, 4, 5))
).astype(self.W_dtype, copy=False)
return gW,
def backward_cpu(self, x, gy):
# x is a dummy variable, which is required only for compatibility with pooling_2d.Pooling2D
col = conv.im2col_cpu(gy[0],
self.kh,
self.kw,
self.sy,
self.sx,
self.ph,
self.pw,
pval=-float('inf'),
cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = col.reshape(n, c, kh * kw, out_h, out_w)
# We select maximum twice, since the implementation using numpy.choose
# hits its bug when kh * kw >= 32.
gx = col.max(axis=2)
return gx,
def forward_cpu(self, x):
if (intel64.should_use_ideep('>=auto')
and intel64.inputs_all_ready(x)):
return self._forward_ideep(x)
self._in_shape = x[0].shape
self._in_dtype = x[0].dtype
col = conv.im2col_cpu(
x[0], self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
pval=-float('inf'), cover_all=self.cover_all)
n, c, kh, kw, out_h, out_w = col.shape
col = col.reshape(n, c, kh * kw, out_h, out_w)
# We select maximum twice, since the implementation using numpy.choose
# hits its bug when kh * kw >= 32.
self.indexes = col.argmax(axis=2)
y = col.max(axis=2)
return y,
def check_forward(self, y):
y = F.upsampling_2d(
self.pooled_y, self.p.indexes, ksize=(self.p.kh, self.p.kw),
stride=(self.p.sy, self.p.sx), pad=(self.p.ph, self.p.pw),
outsize=self.in_shape[2:], cover_all=self.p.cover_all)
if isinstance(y.data, numpy.ndarray):
y = conv.im2col_cpu(y.data, self.p.kh, self.p.kw,
self.p.sy, self.p.sx, self.p.ph, self.p.pw)
else:
y = conv.im2col_gpu(y.data, self.p.kh, self.p.kw,
self.p.sy, self.p.sx, self.p.ph, self.p.pw)
for i in numpy.ndindex(y.shape):
n, c, ky, kx, oy, ox = i
up_y = y[n, c, ky, kx, oy, ox]
if ky * y.shape[3] + kx == self.p.indexes[n, c, oy, ox]:
in_y = self.pooled_y.data[n, c, oy, ox]
testing.assert_allclose(in_y, up_y)
else:
testing.assert_allclose(up_y, 0)
def check_forward(self, y):
y = F.upsampling_2d(
self.pooled_y, self.indices, ksize=self.ksize,
stride=self.stride, outsize=self.in_shape[2:])
if isinstance(y.array, numpy.ndarray):
y = conv.im2col_cpu(
y.array, self.ksize, self.ksize, self.stride, self.stride,
0, 0)
else:
y = conv.im2col_gpu(
y.array, self.ksize, self.ksize, self.stride, self.stride,
0, 0)
for i in numpy.ndindex(y.shape):
n, c, ky, kx, oy, ox = i
up_y = y[n, c, ky, kx, oy, ox]
if ky * y.shape[3] + kx == self.indices[n, c, oy, ox]:
in_y = self.pooled_y.array[n, c, oy, ox]
testing.assert_allclose(in_y, up_y)
else:
testing.assert_allclose(up_y, 0)
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:]
self.col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
y = numpy.tensordot(
self.col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
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 forward_cpu(self, inputs):
self.retain_inputs((0, 1)) # retain only x and W
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:]
col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all, dy=self.dy, dx=self.dx)
y = numpy.tensordot(
col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def forward_cpu(self, x):
col = conv.im2col_cpu(x[0], self.kh, self.kw, self.sy, self.sx,
self.ph, self.pw)
y = col.mean(axis=(2, 3))
return y,
