def pool(self, y):
print ("pool")
'''apply pooling to unpooled output'''
if self.vis_shape is None:
self.vis_shape = test_shape(y)[-2:]
(p_y, p_inds) = pooling.maxpool2d(y, winsize=self.poolsize,stride=self.poolstride)
return (p_y, p_inds)
def define_coarse_stack(self, imnet_feats):
full1 = self.create_unit('full1', ninput=test_shape(imnet_feats)[1])
f_1 = relu(full1.infer(imnet_feats))
f_1_drop = random_zero(f_1, 0.5)
f_1_mean = 0.5 * f_1
full2 = self.create_unit('full2', ninput=test_shape(f_1_mean)[1])
f_2_drop = full2.infer(f_1_drop)
f_2_mean = full2.infer(f_1_mean)
# prediction
(h, w) = self.output_size
pred_drop = f_2_drop.reshape((self.bsize, h, w))
pred_mean = f_2_mean.reshape((self.bsize, h, w))
self.coarse = MachinePart(locals())
def pool(self, y):
print "pool"
'''apply pooling to unpooled output'''
if self.vis_shape is None:
self.vis_shape = test_shape(y)[-2:]
(p_y, p_inds) = pooling.maxpool2d(y, winsize=self.poolsize,stride=self.poolstride)
return (p_y, p_inds)
def define_coarse_stack(self, imnet_feats):
full1 = self.create_unit('full1', ninput=test_shape(imnet_feats)[1])
f_1 = relu(full1.infer(imnet_feats))
f_1_drop = random_zero(f_1, 0.5)
f_1_mean = 0.5 * f_1
full2 = self.create_unit('full2', ninput=test_shape(f_1_mean)[1])
f_2_drop = full2.infer(f_1_drop)
f_2_mean = full2.infer(f_1_mean)
# prediction
(h, w) = self.coarse_output_size
pred_drop = f_2_drop.reshape((self.bsize, h, w))
pred_mean = f_2_mean.reshape((self.bsize, h, w))
self.coarse = MachinePart(locals())
def conv_theano_mm(x, k, border_mode, transpose=False, stride=1):
#输入图片x： (bsize, xchan, h, w)
#k为滤波器：(nfilt, xchan, filt_h, filt_w)
(xh, xw) = test_shape(x)[-2:]
(kh, kw) = test_shape(k)[-2:]
if border_mode == 'valid':
pad = (0, 0)
elif border_mode == 'same': #卷积后的图片大小与原图片的大小相同，因此左右两边都要加上卷积核宽度的一半
pad = (kh // 2, kw // 2)
elif border_mode == 'full':
pad = (kh - 1, kw - 1)
else:
raise ValueError(border_mode)
if stride != 1 and not transpose and _mm_enable_compatibility_padding:
print 'True'
# semi-compatibility with cudaconv
# cudaconv strided convs go one filter tile past the end at the
# bottom/right. Get the same size with some extra padding if needed.
# The padding is centered, so this results in up to a half-stride image
# shift to the right, not exactly the same as before.
if border_mode != 'valid':
raise NotImplementedError()
old_h = np.ceil((xh - kh) / float(stride)) * stride + kh
old_w = np.ceil((xw - kw) / float(stride)) * stride + kw
pad = (int(np.ceil(
(old_h - xh) / 2.0)), int(np.ceil((old_w - xw) / 2.0)))
if transpose:
(ph, pw) = pad
bottom_shape = T.constant(
np.array((stride * (xh - 1) - 2 * ph + kh,
stride * (xw - 1) - 2 * pw + kw)))
res = theano.sandbox.cuda.blas.GpuCorrMM_gradInputs(
pad=pad,
subsample=(stride, stride)) \
(k, x, shape=bottom_shape)
else:
res = theano.sandbox.cuda.blas.GpuCorrMM(
pad=pad,
subsample=(stride, stride)) \
(x, k)
return res
def conv_theano_mm(x, k, border_mode, transpose=False, stride=1):
#输入图片x： (bsize, xchan, h, w)
#k为滤波器：(nfilt, xchan, filt_h, filt_w)
(xh, xw) = test_shape(x)[-2:]
(kh, kw) = test_shape(k)[-2:]
if border_mode == 'valid':
pad = (0,0)
elif border_mode == 'same':#卷积后的图片大小与原图片的大小相同，因此左右两边都要加上卷积核宽度的一半
pad = (kh // 2, kw // 2)
elif border_mode == 'full':
pad = (kh - 1, kw - 1)
else:
raise ValueError(border_mode)
if stride != 1 and not transpose and _mm_enable_compatibility_padding:
print 'True'
# semi-compatibility with cudaconv
# cudaconv strided convs go one filter tile past the end at the
# bottom/right. Get the same size with some extra padding if needed.
# The padding is centered, so this results in up to a half-stride image
# shift to the right, not exactly the same as before.
if border_mode != 'valid':
raise NotImplementedError()
old_h = np.ceil((xh - kh) / float(stride)) * stride + kh
old_w = np.ceil((xw - kw) / float(stride)) * stride + kw
pad = (int(np.ceil((old_h - xh) / 2.0)),
int(np.ceil((old_w - xw) / 2.0)))
if transpose:
(ph, pw) = pad
bottom_shape = T.constant(np.array((stride * (xh - 1) - 2*ph + kh,
stride * (xw - 1) - 2*pw + kw)))
res = theano.sandbox.cuda.blas.GpuCorrMM_gradInputs(
pad=pad,
subsample=(stride, stride)) \
(k, x, shape=bottom_shape)
else:
res = theano.sandbox.cuda.blas.GpuCorrMM(
pad=pad,
subsample=(stride, stride)) \
(x, k)
return res
def define_cost(self, pred, y0, m0):
bsize = self.bsize
npix = int(np.prod(test_shape(y0)[1:]))
y0_target = y0.reshape((self.bsize, npix))
y0_mask = m0.reshape((self.bsize, npix))
pred = pred.reshape((self.bsize, npix))
p = pred * y0_mask
t = y0_target * y0_mask
d = (p - t)
nvalid_pix = T.sum(y0_mask, axis=1)
depth_cost = (T.sum(nvalid_pix * T.sum(d**2, axis=1))
- 0.5*T.sum(T.sum(d, axis=1)**2)) \
/ T.maximum(T.sum(nvalid_pix**2), 1)
return depth_cost
def define_cost(self, pred, y0, m0):
bsize = self.bsize
npix = int(np.prod(test_shape(y0)[1:]))
y0_target = y0.reshape((self.bsize, npix))
y0_mask = m0.reshape((self.bsize, npix))
pred = pred.reshape((self.bsize, npix))
p = pred * y0_mask
t = y0_target * y0_mask
d = (p - t)
nvalid_pix = T.sum(y0_mask, axis=1)
depth_cost = (T.sum(nvalid_pix * T.sum(d**2, axis=1))
- 0.5*T.sum(T.sum(d, axis=1)**2)) \
/ T.maximum(T.sum(nvalid_pix**2), 1)
return depth_cost
