def CNN(x,c_l1,c_l2,f_l1,f_l2,PP,ims):
print ims
#-------
#conv3D get rid of dependency of the number of input image channel
b=numpy.zeros(c_l1.get_value().shape[0])
conv1=tensor.nnet.relu(conv3D(x.dimshuffle(0,2,3,1,'x'),c_l1.dimshuffle(0,2,3,1,'x'),b,d=(1,1,1))) # shuffle dimensions
conv1=tensor.sum(conv1,axis=3) #add the dimension of channels
conv1=conv1.dimshuffle(0,3,1,2) #shuffle back to same dimension as conv2D
#---------
#conv1=tensor.nnet.relu(conv2d(x,c_l1)) #default stride=1 --subsample=(1,1)
conv1_shp=get_conv_output_shape(ims,c_l1.get_value().shape,border_mode='valid',subsample=(1,1))
print conv1_shp
#pp=tensor.reshape(conv1,conv1_shp[:2]+(conv1_shp[2]*conv1_shp[3],))
#print pp
pool1=pool_2d(conv1,(2,2),st=(2,2),ignore_border=True) #default maxpool
pool1_shp=get_pool_output_shape(conv1_shp,pool_size=(2,2),st=(2,2),ignore_border=True)
print pool1_shp
conv2=tensor.nnet.relu(conv2d(pool1,c_l2))
conv2_shp=get_conv_output_shape(pool1_shp,c_l2.get_value().shape,border_mode='valid',subsample=(1,1))
print conv2_shp
#pool2=pool_2d(conv2,(2,2),st=(2,2),ignore_border=True)
pool2=spp(conv2,conv2_shp,PP,'max')
fpool2=tensor.flatten(pool2,outdim=2)
full1=tensor.nnet.relu(tensor.dot(fpool2,f_l1))
pyx=tensor.nnet.softmax(tensor.dot(full1,f_l2))
return c_l1, c_l2, f_l1, f_l2, pyx
def CNN(x,c_l1,c_l2,f_l1,f_l2,insize):
print "in size ", insize
conv1=tensor.nnet.relu(conv2d(x,c_l1)) #default stride=1 --subsample=(1,1)
conv1_shp=get_conv_output_shape(insize,c_l1.get_value().shape,border_mode='valid',subsample=(1,1))
print "conv1 size ", conv1_shp
pool1=pool_2d(conv1,(3,3),st=(3,3),ignore_border=True) #default maxpool
pool1_shp=get_pool_output_shape(conv1_shp,pool_size=(3,3),st=(3,3),ignore_border=True)
print "pool1 size ", pool1_shp
lrn1=LRN(pool1,pool1_shp)
lrn1_shp=tuple(pool1_shp)
print "cross map norm1 size ", lrn1_shp
conv2=tensor.nnet.relu(conv2d(lrn1,c_l2))
conv2_shp=get_conv_output_shape(lrn1_shp,c_l2.get_value().shape,border_mode='valid',subsample=(1,1))
print "conv2 size ", conv2_shp
pool2=pool_2d(conv2,(2,2),st=(2,2),ignore_border=True)
pool2_shp=get_pool_output_shape(conv2_shp,pool_size=(2,2),st=(2,2),ignore_border=True)
print "pool2 size ", pool2_shp
lrn2=LRN(pool2,pool2_shp)
lrn2_shp=tuple(pool2_shp)
print "cross map norm2 size " , lrn2_shp
fpool2=tensor.flatten(lrn2,outdim=2)
full1=tensor.nnet.relu(tensor.dot(fpool2,f_l1))
pyx=tensor.nnet.sigmoid(tensor.dot(full1,f_l2))
return c_l1, c_l2, f_l1, f_l2, pyx
def burn():
sz = 128
img_shp = [sz, sz, sz, sz]
kern_shp = [sz // 2, sz, 3, 3]
out_shp = get_conv_output_shape(img_shp, kern_shp, "valid", (1, 1))
img = tt.tensor4("img")
kern = tt.tensor4("kern")
out = tt.tensor4("out")
def rand(shp):
return np.random.rand(*shp).astype(theano.config.floatX)
img = theano.shared(rand(img_shp))
kern = theano.shared(rand(kern_shp))
out = theano.shared(rand(out_shp))
# beta 1 is needed to force the reuse of out, otherwise, it is
# replaced by a GpuAllocEmpty
o1 = dnn._dnn_conv(img, kern, conv_mode="conv", out=out, beta=1.0)
mode = theano.compile.get_default_mode().including("local_remove_all_assert")
f = theano.function([], [o1], mode=mode)
theano.printing.debugprint(f)
print("Start computation")
for i in range(10000):
f.fn()
print("Computation stopped")
def infer_shape(self, node, input_shape):
imshp = input_shape[0]
kshp = input_shape[1]
res = get_conv_output_shape(
imshp, kshp, self.border_mode, self.subsample, self.filter_dilation
)
return [res]
def array_like_conv_output(self, inputs_shape, filters_shape, border_mode,
subsample, dilation, dtype):
# Return a random array with inferred convolution output shape.
out_shp = get_conv_output_shape(inputs_shape, filters_shape,
border_mode, subsample, dilation)
out_shp = assert_conv_shape(out_shp)
return np.random.random(out_shp).astype(dtype)
def burn():
sz = 128
img_shp = [sz, sz, sz, sz]
kern_shp = [sz // 2, sz, 3, 3]
out_shp = get_conv_output_shape(img_shp, kern_shp, 'valid', (1, 1))
img = T.tensor4('img')
kern = T.tensor4('kern')
out = T.tensor4('out')
def rand(shp):
return np.random.rand(*shp).astype(theano.config.floatX)
img = theano.shared(rand(img_shp))
kern = theano.shared(rand(kern_shp))
out = theano.shared(rand(out_shp))
# beta 1 is needed to force the reuse of out, otherwise, it is
# replaced by a GpuAllocEmpty
o1 = dnn._dnn_conv(img, kern, conv_mode='conv', out=out, beta=1.)
mode = theano.compile.get_default_mode().including(
"local_remove_all_assert")
f = theano.function([], [o1], mode=mode)
theano.printing.debugprint(f)
print("Start computation")
for i in range(10000):
f.fn()
print("Computation stopped")
def test_basic(self):
image_shape, kernel_shape = (3, 2, 8, 9), (4, 2, 5, 6)
sub_sample = (1, 2)
test1_params = get_conv_output_shape(
image_shape, kernel_shape, 'valid', sub_sample)
test2_params = get_conv_output_shape(
image_shape, kernel_shape, 'half', sub_sample)
test3_params = get_conv_output_shape(
image_shape, kernel_shape, 'full', sub_sample)
test4_params = get_conv_output_shape(
image_shape, kernel_shape, (1, 2), sub_sample)
self.assertTrue(test1_params == (3, 4, 4, 2))
self.assertTrue(test2_params == (3, 4, 8, 5))
self.assertTrue(test3_params == (3, 4, 12, 7))
self.assertTrue(test4_params == (3, 4, 6, 4))
def get_out_shape(ishape, kshape, border_mode, subsample):
"""
This function computes the output shape for a convolution with
the specified parameters. `ishape` and `kshape` can be symbolic
or scalar.
"""
return get_conv_output_shape(ishape, kshape, border_mode, subsample)
def test_basic_3d(self):
image_shape, kernel_shape = (3, 2, 12, 9, 7), (4, 2, 5, 6, 4)
sub_sample = (1, 2, 1)
filter_dilation = (2, 1, 1)
test1_params = get_conv_output_shape(
image_shape, kernel_shape, 'valid', sub_sample, filter_dilation)
test2_params = get_conv_output_shape(
image_shape, kernel_shape, 'half', sub_sample, filter_dilation)
test3_params = get_conv_output_shape(
image_shape, kernel_shape, 'full', sub_sample, filter_dilation)
test4_params = get_conv_output_shape(
image_shape, kernel_shape, (1, 2, 3), sub_sample, filter_dilation)
self.assertTrue(test1_params == (3, 4, 4, 2, 4))
self.assertTrue(test2_params == (3, 4, 12, 5, 8))
self.assertTrue(test3_params == (3, 4, 20, 7, 10))
self.assertTrue(test4_params == (3, 4, 6, 4, 10))
def test_basic_3d(self):
image_shape, kernel_shape = (3, 2, 12, 9, 7), (4, 2, 5, 6, 4)
sub_sample = (1, 2, 1)
filter_dilation = (2, 1, 1)
test1_params = get_conv_output_shape(
image_shape, kernel_shape, 'valid', sub_sample, filter_dilation)
test2_params = get_conv_output_shape(
image_shape, kernel_shape, 'half', sub_sample, filter_dilation)
test3_params = get_conv_output_shape(
image_shape, kernel_shape, 'full', sub_sample, filter_dilation)
test4_params = get_conv_output_shape(
image_shape, kernel_shape, (1, 2, 3), sub_sample, filter_dilation)
self.assertTrue(test1_params == (3, 4, 4, 2, 4))
self.assertTrue(test2_params == (3, 4, 12, 5, 8))
self.assertTrue(test3_params == (3, 4, 20, 7, 10))
self.assertTrue(test4_params == (3, 4, 6, 4, 10))
def get_out_shape(ishape, kshape, border_mode, subsample):
"""
This function computes the output shape for a convolution with
the specified parameters. `ishape` and `kshape` can be symbolic
or scalar.
"""
return get_conv_output_shape(ishape, kshape, border_mode, subsample)
def infer_shape(self, node, input_shape):
imshp = input_shape[0]
kshp = input_shape[1]
res = get_conv_output_shape(
imshp,
kshp,
self.border_mode,
self.subsample)
return [res]
def get_conv_shape(input_shape, filter_shape, padding, stride):
"""
Helper method to calculate the shapes post-convolution operation given input parameters. This isn't used
for our output_size calculations because Theano provides a function specific to its conv op.
"""
if isinstance(input_shape, Iterable):
shape = get_conv_output_shape(input_shape, filter_shape, padding, stride)
else:
shape = get_conv_shape_1axis(input_shape, filter_shape, padding, stride)
return shape
def get_if_valid_conv_output_shape(case_tuple):
# Filter function to keep only cases that produce valid convolution output shapes.
out_shp = get_conv_output_shape(case_tuple[0], # input shape
case_tuple[1], # filter shape
case_tuple[4], # border mode
case_tuple[2], # subsample
case_tuple[3]) # dilation
try:
return assert_conv_shape(out_shp)
except ValueError:
return False
def CNN(x,c_l1,c_l2,f_l1,f_l2,PP,ims):
print ims
conv1=tensor.nnet.relu(conv2d(x,c_l1)) #default stride=1 --subsample=(1,1)
conv1_shp=get_conv_output_shape(ims,c_l1.get_value().shape,border_mode='valid',subsample=(1,1))
print conv1_shp
pp=tensor.reshape(conv1,conv1_shp[:2]+(conv1_shp[2]*conv1_shp[3],))
print pp
pool1=pool_2d(conv1,(2,2),st=(2,2),ignore_border=True) #default maxpool
pool1_shp=get_pool_output_shape(conv1_shp,pool_size=(2,2),st=(2,2),ignore_border=True)
print pool1_shp
conv2=tensor.nnet.relu(conv2d(pool1,c_l2))
conv2_shp=get_conv_output_shape(pool1_shp,c_l2.get_value().shape,border_mode='valid',subsample=(1,1))
print conv2_shp
#pool2=pool_2d(conv2,(2,2),st=(2,2),ignore_border=True)
pool2=spp(conv2,conv2_shp,PP,'max')
fpool2=tensor.flatten(pool2,outdim=2)
full1=tensor.nnet.relu(tensor.dot(fpool2,f_l1))
pyx=tensor.nnet.sigmoid(tensor.dot(full1,f_l2))
return c_l1, c_l2, f_l1, f_l2, pyx
def get_conv_shape(input_shape, filter_shape, padding, stride):
"""
Helper method to calculate the shapes post-convolution operation given input parameters. This isn't used
for our output_size calculations because Theano provides a function specific to its conv op.
"""
if isinstance(input_shape, Iterable):
shape = get_conv_output_shape(input_shape, filter_shape, padding,
stride)
else:
shape = get_conv_shape_1axis(input_shape, filter_shape, padding,
stride)
return shape
def _build(self, input_tensor):
"""Build 2D conolution operation of the input tensor
Parameters
----------
input_tensor : Tensor
4D Tensor with shape (batch, #input channel, row, col)
Returns
-------
Tensor
4D Tensor with shape (batch, #output channel, row, col)
"""
input_shape = input_tensor.shape
_LG.debug(' input_shape: %s', input_shape)
if not len(input_shape) == 4:
raise ValueError(
'Input tensor must be 4D. ({})'.format(input_tensor))
border_mode = _map_border_mode(self.args['padding'])
subsample = _get_subsample(self.args['strides'])
filter_shape = self._get_filter_shape(input_shape[1])
bias_shape = (filter_shape[0], )
output_shape = get_conv_output_shape(input_shape, filter_shape,
border_mode, subsample)
_check_output_shape(input_shape, filter_shape, border_mode, subsample)
_LG.debug(' border_mode: %s', border_mode)
_LG.debug(' subsample: %s', subsample)
_LG.debug(' filter_shape: %s', filter_shape)
_LG.debug(' output_shape: %s', output_shape)
self._build_parameters(filter_shape, bias_shape, input_tensor.dtype)
filters = self.get_parameter_variable('filter')
output_tensor = T.nnet.conv2d(input_tensor.unwrap(),
filters=filters.unwrap(),
input_shape=input_shape,
filter_shape=filter_shape,
border_mode=border_mode,
subsample=subsample)
if self.args['with_bias']:
bias = self.get_parameter_variable('bias').unwrap()
bias = bias.dimshuffle(('x', 0, 'x', 'x'))
output_tensor = bias + output_tensor
return wrapper.Tensor(output_tensor, shape=output_shape, name='output')
def _build(self, input_tensor):
"""Build 2D conolution operation of the input tensor
Parameters
----------
input_tensor : Tensor
4D Tensor with shape (batch, #input channel, row, col)
Returns
-------
Tensor
4D Tensor with shape (batch, #output channel, row, col)
"""
input_shape = input_tensor.shape
_LG.debug(' input_shape: %s', input_shape)
if not len(input_shape) == 4:
raise ValueError(
'Input tensor must be 4D. ({})'.format(input_tensor))
border_mode = _map_border_mode(self.args['padding'])
subsample = _get_subsample(self.args['strides'])
filter_shape = self._get_filter_shape(input_shape[1])
bias_shape = (filter_shape[0],)
output_shape = get_conv_output_shape(
input_shape, filter_shape, border_mode, subsample)
_check_output_shape(input_shape, filter_shape, border_mode, subsample)
_LG.debug(' border_mode: %s', border_mode)
_LG.debug(' subsample: %s', subsample)
_LG.debug(' filter_shape: %s', filter_shape)
_LG.debug(' output_shape: %s', output_shape)
self._build_parameters(filter_shape, bias_shape, input_tensor.dtype)
filters = self.get_parameter_variable('filter')
output_tensor = T.nnet.conv2d(
input_tensor.unwrap(), filters=filters.unwrap(),
input_shape=input_shape, filter_shape=filter_shape,
border_mode=border_mode, subsample=subsample)
if self.args['with_bias']:
bias = self.get_parameter_variable('bias').unwrap()
bias = bias.dimshuffle(('x', 0, 'x', 'x'))
output_tensor = bias + output_tensor
return wrapper.Tensor(output_tensor, shape=output_shape, name='output')
def array_like_conv_output(self, inputs_shape, filters_shape, border_mode, subsample, dilation, dtype):
# Return a random array with inferred convolution output shape.
out_shp = get_conv_output_shape(inputs_shape, filters_shape, border_mode, subsample, dilation)
out_shp = assert_conv_shape(out_shp)
return np.random.random(out_shp).astype(dtype)
def local_conv2d_gradinputs_cpu(node):
if not isinstance(node.op, AbstractConv2d_gradInputs):
return None
kern, topgrad, shape = node.inputs
if ((not isinstance(kern.type, TensorType) or
not isinstance(topgrad.type, TensorType))):
return None
if node.op.border_mode not in ['full', 'valid']:
return None
if not node.op.filter_flip:
# Not tested yet
return None
# Conv 3d implementation, needed when subsample > 2
if node.op.border_mode == 'valid' and node.op.subsample != (1, 1):
kern = kern[:, :, ::-1, ::-1]
shuffled_kern = kern.dimshuffle(0, 2, 3, 'x', 1)
shuffled_topgrad = topgrad.dimshuffle(0, 2, 3, 'x', 1)
b = theano.tensor.zeros_like(shuffled_kern[0, 0, 0, 0, :])
rval = convTransp3D(W=shuffled_kern, b=b,
d=(node.op.subsample[0], node.op.subsample[1], 1),
H=shuffled_topgrad,
RShape=(shape[0], shape[1], 1))
copy_stack_trace(node.outputs[0], rval)
rval = theano.tensor.addbroadcast(rval, 3)
rval = rval.dimshuffle(0, 4, 1, 2)
rval = theano.tensor.patternbroadcast(rval,
node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], rval)
return [rval]
# Conv2d Implementation
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
return None
mode = 'valid'
if not node.op.border_mode == 'full':
mode = 'full'
filters = kern.dimshuffle((1, 0, 2, 3))
filters = filters[:, :, ::-1, ::-1]
outshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode, node.op.subsample)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode, (1, 1))[2:]
nkern = op_imshp[1]
imshp = (op_kshp[0], outshp[0], outshp[1])
imshp_logical = (op_kshp[0], fulloutshp[0], fulloutshp[1])
din = ConvOp(imshp,
op_kshp[2:],
nkern,
op_imshp[0],
1, 1, output_mode=mode,
unroll_batch=None, unroll_kern=None,
unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=None,
version=-1,
direction_hint='bprop inputs')
din = din(topgrad, filters)
copy_stack_trace(node.outputs[0], din)
din = theano.tensor.patternbroadcast(din, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], din)
return [din]
def local_conv2d_gradweight_cpu(node):
if not isinstance(node.op, AbstractConv2d_gradWeights):
return None
img, topgrad, shape = node.inputs
if ((not isinstance(img.type, TensorType) or
not isinstance(topgrad.type, TensorType))):
return None
if node.op.border_mode not in ['full', 'valid']:
return None
if not node.op.filter_flip:
# Not tested yet
return
if node.op.border_mode == 'valid' and \
(node.op.subsample != (1, 1)):
# Use the gradient as defined in conv3D, because the implementation
# by Conv is slow (about 3x slower than conv3D, and probably 10x
# slower than it could be), and incorrect when subsample > 2.
# build a "node", that should be equivalent to the one given by
# self.make_node, but using convGrad3D instead.
shuffled_img = img.dimshuffle(0, 2, 3, 'x', 1)
shuffled_topgrad = topgrad.dimshuffle(0, 2, 3, 'x', 1)
rval = convGrad3D(V=shuffled_img,
d=(node.op.subsample[0], node.op.subsample[1], 1),
WShape=(shuffled_topgrad.shape[4],
shape[0], shape[1], 1,
shuffled_img.shape[4]),
dCdH=shuffled_topgrad)
copy_stack_trace(node.outputs[0], rval)
rval = theano.tensor.addbroadcast(rval, 3)
rval = rval.dimshuffle(0, 4, 1, 2)
rval = rval[:, :, ::-1, ::-1]
rval = theano.tensor.patternbroadcast(rval,
node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], rval)
return [rval]
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
# We cannot infer the shapes
return None
# Determine gradient on kernels
assert len(op_imshp) == 4 and len(op_kshp) == 4
outshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode, node.op.subsample)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode, (1, 1))[2:]
newimg = img.dimshuffle((1, 0, 2, 3))
newtopgrad = topgrad.dimshuffle((1, 0, 2, 3))
if node.op.border_mode == 'valid':
(img, filters) = (newimg, newtopgrad)
kshp_logical = fulloutshp
kshp_logical_top_aligned = False
imshp_logical = None
(bsize, nkern) = (op_imshp[1], op_kshp[0])
imshp = (op_imshp[0], op_imshp[2], op_imshp[3])
kshp = outshp
elif node.op.border_mode == 'full':
(img, filters) = (newtopgrad, newimg)
kshp_logical = None
kshp_logical_top_aligned = True
imshp_logical = (op_imshp[0],
fulloutshp[0],
fulloutshp[1])
(bsize, nkern) = (op_kshp[0], op_imshp[1])
imshp = (op_imshp[0], outshp[0], outshp[1])
kshp = op_imshp[2:]
else:
raise NotImplementedError(
'Only [full,valid] modes are currently supported.')
# Flip the kernels
filters = filters[:, :, ::-1, ::-1]
dw = ConvOp(imshp, kshp, nkern, bsize, 1, 1, output_mode='valid',
unroll_batch=None, unroll_kern=None, unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=kshp_logical,
kshp_logical_top_aligned=kshp_logical_top_aligned,
direction_hint='bprop weights')
res = dw(img, filters)
copy_stack_trace(node.outputs[0], res)
if node.op.border_mode == 'valid':
res = res.dimshuffle((1, 0, 2, 3))
res = res[:, :, ::-1, ::-1]
res = theano.tensor.patternbroadcast(res, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], res)
return [res]
def c_code(self, node, name, inp, out, sub):
x, = inp
dH, dW = self.subsample
if self.imshp is None:
self.imshp = x.shape
i_n, i_c, i_h, i_w = self.imshp
if len(self.kshp) == 5:
grp, k_n, k_c, k_h, k_w = self.kshp
assert i_c == k_c * grp
else:
k_n, k_c, k_h, k_w = self.kshp
grp = 1
o_n, o_c, o_h, o_w = get_conv_output_shape(image_shape=self.imshp,
kernel_shape=self.kshp,
border_mode=self.border_mode,
filter_dilation=self.filter_dilation,
subsample=self.subsample)
if self.border_mode == 'valid':
padH, padW = (0, 0)
elif self.border_mode == 'full':
padH, padW = ((k_h - 1), (k_w - 1))
elif self.border_mode == 'half':
padH, padW = ((k_h / 2), (k_w / 2))
elif isinstance(self.border_mode, tuple):
padH, padW = self.border_mode
else:
raise ValueError("border_mode must have two elements")
z, = out
if 'float32' == node.inputs[0].type.dtype:
precision = 'F32'
elif 'float64' == node.inputs[0].type.dtype:
precision = 'F64'
else:
raise Exception("Type %s is not supported!" %
node.inputs[0].type.dtype)
fail = sub['fail']
ccode = """
if (1 == first_run) {
int convPadding[2];
size_t convStride[2], weightSize[5], weightStride[5], imageSize[4], imageStride[4], zSize[4], zStride[4];
convStride[0] = %(dW)s;
convStride[1] = %(dH)s;
convPadding[0] = -%(padW)s;
convPadding[1] = -%(padH)s;
imageSize[0] = %(i_w)s; //w
imageSize[1] = %(i_h)s; //h
imageSize[2] = %(i_c)s; //c
imageSize[3] = %(i_n)s; //n
imageStride[0] = 1;
imageStride[1] = imageSize[0];
imageStride[2] = imageSize[0] * imageSize[1];
imageStride[3] = imageSize[0] * imageSize[1] * imageSize[2];
weightSize[0] = %(k_w)s;
weightSize[1] = %(k_h)s;
weightSize[2] = %(k_c)s;
weightSize[3] = %(k_n)s;
weightSize[4] = %(grp)s;
weightStride[0] = 1;
weightStride[1] = weightSize[0];
weightStride[2] = weightSize[0] * weightSize[1];
weightStride[3] = weightSize[0] * weightSize[1] * weightSize[2];
weightStride[4] = weightSize[0] * weightSize[1] * weightSize[2] * weightSize[3];
zSize[0] = %(o_w)s;
zSize[1] = %(o_h)s;
zSize[2] = %(o_c)s;
zSize[3] = %(o_n)s;
zStride[0] = 1;
zStride[1] = zSize[0];
zStride[2] = zSize[0] * zSize[1];
zStride[3] = zSize[0] * zSize[1] * zSize[2];
const int group = %(grp)s;
//create user layout
CHECK_ERR( dnnLayoutCreate_%(precision)s(&layout_user, DIMENSION, imageSize, imageStride), err );
CHECK_ERR( dnnGroupsConvolutionCreateForward_%(precision)s(&primitive, NULL,
dnnAlgorithmConvolutionDirect, group, DIMENSION, imageSize, zSize,
weightSize, convStride, convPadding, dnnBorderZeros), err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_%(precision)s(&layout_internal, primitive, dnnResourceSrc), err );
}
if (!dnnLayoutCompare_%(precision)s(layout_user, layout_internal))
{
if (NULL == to_internal)
{
CHECK_ERR( dnnConversionCreate_%(precision)s(&to_internal, layout_user, layout_internal), err );
}
}
if (NULL == %(z)s)
{
//Create PyArrayObject for output
%(z)s = (PyArrayObject*)PyArray_ZEROS(DIMENSION, PyArray_DIMS(%(x)s), PyArray_TYPE(%(x)s), 0);
if (NULL == %(z)s)
{
%(fail)s
}
}
if (NULL == internal_buf)
{
CHECK_ERR( dnnAllocateBuffer_%(precision)s((void**)&internal_buf, layout_internal), err );
}
if (to_internal)
{
convert_resources[dnnResourceFrom] = (PyArray_DATA(%(x)s));
convert_resources[dnnResourceTo] = (void*)(internal_buf);
CHECK_ERR( dnnExecute_%(precision)s(to_internal, convert_resources), err );
}
else
{
internal_buf = (PyArray_DATA(%(x)s));
}
if (layout_internal != ((dnnLayout_t*)PyArray_DATA(%(z)s))[0])
{
((dnnLayout_t*)PyArray_DATA(%(z)s))[0] = layout_internal;
}
if (internal_buf != ((void**)PyArray_DATA(%(z)s))[1])
{
((void**)PyArray_DATA(%(z)s))[1] = internal_buf;
}
first_run = 0;
#ifdef _MKL_DEBUG_
std::cout << "U2IConv2D: from buffer: " << convert_resources[dnnResourceFrom] << " to buffer: " << convert_resources[dnnResourceTo] << std::endl;
#endif
""" % locals()
return ccode
args.input_shape,
args.filter_shape,
args.subsample,
args.dilation,
args.border_mode,
args.conv_mode,
args.alpha,
args.beta,
)
if args.print_infos:
CheckDnn.print_infos(count_tests=False)
print("======================")
print("Running", test, algo, dtype, precision, *parameters)
if test == FWD:
tests.run_conv_fwd(algo, dtype, precision, parameters)
expected_output_shape = get_conv_output_shape(
args.input_shape,
args.filter_shape,
args.border_mode,
args.subsample,
args.dilation,
)
elif test == BWD_FILTER:
tests.run_conv_gradweight(algo, dtype, precision, parameters)
expected_output_shape = args.filter_shape
elif test == BWD_DATA:
tests.run_conv_gradinput(algo, dtype, precision, parameters)
expected_output_shape = args.input_shape
print("Computed shape:", expected_output_shape)
print("... OK")
def local_conv2d_gradinputs_cpu(node):
if (not isinstance(node.op, AbstractConv2d_gradInputs)
or node.inputs[0].dtype == "float16"):
return None
kern, topgrad, shape = node.inputs
if not isinstance(kern.type, TensorType) or not isinstance(
topgrad.type, TensorType):
return None
if node.op.border_mode not in ["full", "valid"]:
return None
if not node.op.filter_flip:
# Not tested yet
return None
if node.op.num_groups > 1 or node.op.unshared:
return None
# Conv 3d implementation, needed when subsample > 2
if node.op.border_mode == "valid" and node.op.subsample != (1, 1):
# The op don't support that anymore.
return False
# Conv2d Implementation
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
return None
mode = "valid"
if not node.op.border_mode == "full":
mode = "full"
filters = kern.dimshuffle((1, 0, 2, 3))
filters = filters[:, :, ::-1, ::-1]
outshp = get_conv_output_shape(
op_imshp,
op_kshp,
node.op.border_mode,
node.op.subsample,
node.op.filter_dilation,
)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp, node.op.border_mode,
(1, 1))[2:]
nkern = op_imshp[1]
imshp = (op_kshp[0], outshp[0], outshp[1])
imshp_logical = (op_kshp[0], fulloutshp[0], fulloutshp[1])
din = ConvOp(
imshp,
op_kshp[2:],
nkern,
op_imshp[0],
1,
1,
output_mode=mode,
unroll_batch=None,
unroll_kern=None,
unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=None,
version=-1,
direction_hint="bprop inputs",
)
din = din(topgrad, filters)
copy_stack_trace(node.outputs[0], din)
din = theano.tensor.patternbroadcast(din, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], din)
return [din]
def local_conv2d_gradweight_cpu(node):
if (not isinstance(node.op, AbstractConv2d_gradWeights)
or node.inputs[0].dtype == "float16"):
return None
img, topgrad, shape = node.inputs
if not isinstance(img.type, TensorType) or not isinstance(
topgrad.type, TensorType):
return None
if node.op.border_mode not in ["full", "valid"]:
return None
if not node.op.filter_flip:
# Not tested yet
return
if node.op.num_groups > 1 or node.op.unshared:
return None
if node.op.border_mode == "valid" and (node.op.subsample != (1, 1)):
return None
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
# We cannot infer the shapes
return None
# Determine gradient on kernels
assert len(op_imshp) == 4 and len(op_kshp) == 4
outshp = get_conv_output_shape(
op_imshp,
op_kshp,
node.op.border_mode,
node.op.subsample,
node.op.filter_dilation,
)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp, node.op.border_mode,
(1, 1))[2:]
newimg = img.dimshuffle((1, 0, 2, 3))
newtopgrad = topgrad.dimshuffle((1, 0, 2, 3))
if node.op.border_mode == "valid":
(img, filters) = (newimg, newtopgrad)
kshp_logical = fulloutshp
kshp_logical_top_aligned = False
imshp_logical = None
(bsize, nkern) = (op_imshp[1], op_kshp[0])
imshp = (op_imshp[0], op_imshp[2], op_imshp[3])
kshp = outshp
elif node.op.border_mode == "full":
(img, filters) = (newtopgrad, newimg)
kshp_logical = None
kshp_logical_top_aligned = True
imshp_logical = (op_imshp[0], fulloutshp[0], fulloutshp[1])
(bsize, nkern) = (op_kshp[0], op_imshp[1])
imshp = (op_imshp[0], outshp[0], outshp[1])
kshp = op_imshp[2:]
else:
raise NotImplementedError(
"Only [full,valid] modes are currently supported.")
# Flip the kernels
filters = filters[:, :, ::-1, ::-1]
dw = ConvOp(
imshp,
kshp,
nkern,
bsize,
1,
1,
output_mode="valid",
unroll_batch=None,
unroll_kern=None,
unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=kshp_logical,
kshp_logical_top_aligned=kshp_logical_top_aligned,
direction_hint="bprop weights",
)
res = dw(img, filters)
copy_stack_trace(node.outputs[0], res)
if node.op.border_mode == "valid":
res = res.dimshuffle((1, 0, 2, 3))
res = res[:, :, ::-1, ::-1]
res = theano.tensor.patternbroadcast(res, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], res)
return [res]
def local_conv2d_gradweight_cpu(node):
if (not isinstance(node.op, AbstractConv2d_gradWeights) or
node.inputs[0].dtype == 'float16'):
return None
img, topgrad, shape = node.inputs
if ((not isinstance(img.type, TensorType) or
not isinstance(topgrad.type, TensorType))):
return None
if node.op.border_mode not in ['full', 'valid']:
return None
if not node.op.filter_flip:
# Not tested yet
return
if node.op.num_groups > 1 or node.op.unshared:
return None
if node.op.border_mode == 'valid' and \
(node.op.subsample != (1, 1)):
return None
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
# We cannot infer the shapes
return None
# Determine gradient on kernels
assert len(op_imshp) == 4 and len(op_kshp) == 4
outshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode,
node.op.subsample,
node.op.filter_dilation)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode, (1, 1))[2:]
newimg = img.dimshuffle((1, 0, 2, 3))
newtopgrad = topgrad.dimshuffle((1, 0, 2, 3))
if node.op.border_mode == 'valid':
(img, filters) = (newimg, newtopgrad)
kshp_logical = fulloutshp
kshp_logical_top_aligned = False
imshp_logical = None
(bsize, nkern) = (op_imshp[1], op_kshp[0])
imshp = (op_imshp[0], op_imshp[2], op_imshp[3])
kshp = outshp
elif node.op.border_mode == 'full':
(img, filters) = (newtopgrad, newimg)
kshp_logical = None
kshp_logical_top_aligned = True
imshp_logical = (op_imshp[0],
fulloutshp[0],
fulloutshp[1])
(bsize, nkern) = (op_kshp[0], op_imshp[1])
imshp = (op_imshp[0], outshp[0], outshp[1])
kshp = op_imshp[2:]
else:
raise NotImplementedError(
'Only [full,valid] modes are currently supported.')
# Flip the kernels
filters = filters[:, :, ::-1, ::-1]
dw = ConvOp(imshp, kshp, nkern, bsize, 1, 1, output_mode='valid',
unroll_batch=None, unroll_kern=None, unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=kshp_logical,
kshp_logical_top_aligned=kshp_logical_top_aligned,
direction_hint='bprop weights')
res = dw(img, filters)
copy_stack_trace(node.outputs[0], res)
if node.op.border_mode == 'valid':
res = res.dimshuffle((1, 0, 2, 3))
res = res[:, :, ::-1, ::-1]
res = theano.tensor.patternbroadcast(res, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], res)
return [res]
elif test == BWD_FILTER:
check_config = cudnn.bwd_filter_algo_supports_dtype_config(args.algo, args.dtype, args.precision, ndim)
elif test == BWD_DATA:
check_config = cudnn.bwd_data_algo_supports_dtype_config(args.algo, args.dtype, args.precision, ndim)
if not check_config:
print('Warning: %s computation does not normally support configuration (%s, %s) for algo %s.' % (
test, args.dtype, args.precision, args.algo), file=sys.stderr)
algo = args.algo
dtype = args.dtype
precision = args.precision
parameters = (
args.input_shape, args.filter_shape, args.subsample, args.dilation, args.border_mode, args.conv_mode,
args.alpha, args.beta)
if args.print_infos:
CheckDnn.print_infos(count_tests=False)
print('======================')
print('Running', test, algo, dtype, precision, *parameters)
if test == FWD:
tests.run_conv_fwd(algo, dtype, precision, parameters)
expected_output_shape = get_conv_output_shape(args.input_shape, args.filter_shape, args.border_mode,
args.subsample, args.dilation)
elif test == BWD_FILTER:
tests.run_conv_gradweight(algo, dtype, precision, parameters)
expected_output_shape = args.filter_shape
elif test == BWD_DATA:
tests.run_conv_gradinput(algo, dtype, precision, parameters)
expected_output_shape = args.input_shape
print('Computed shape:', expected_output_shape)
print('... OK')
def local_abstractconv_cudnn_alt(node):
if not isinstance(node.op, (AbstractConv2d, AbstractConv2d_gradWeights,
AbstractConv2d_gradInputs)):
return
if version(raises=False) < 6000 and node.op.filter_dilation != (1, 1):
return None
if node.op.unshared:
return None
if isinstance(node.op.border_mode, tuple) and any(
isinstance(p, tuple) for p in node.op.border_mode):
# Asymmetric padding not yet supported
return None
inp1 = node.inputs[0]
inp2 = node.inputs[1]
if not dnn_available(inp1.type.context_name):
return
op = node.op
border_mode = node.op.border_mode
subsample = node.op.subsample
filter_dilation = node.op.filter_dilation
num_groups = node.op.num_groups
precision, _ = get_precision(None, [inp1, inp2])
if node.op.filter_flip:
conv_mode = "conv"
else:
conv_mode = "cross"
if isinstance(op, AbstractConv2d):
if border_mode == "half" or subsample != (1, 1) or num_groups != 1:
return None
if border_mode == "full":
direction_hint = "bprop inputs"
elif border_mode == "valid" and filter_dilation == (1, 1):
direction_hint = "bprop weights"
else:
return None
rval = dnn_conv(
inp1,
inp2,
border_mode=border_mode,
subsample=subsample,
dilation=filter_dilation,
direction_hint=direction_hint,
conv_mode=conv_mode,
num_groups=num_groups,
)
elif isinstance(op, AbstractConv2d_gradWeights):
if (border_mode == "valid" and subsample == (1, 1)
and filter_dilation == (1, 1) and num_groups == 1):
img = gpu_contiguous(inp1)
topgrad = gpu_contiguous(inp2)
ctx_name = infer_context_name(img, topgrad)
img = gpu_contiguous(img.dimshuffle(1, 0, 2, 3))
topgrad = gpu_contiguous(topgrad.dimshuffle(1, 0, 2, 3))
ishape = [shape_i_op(i)(img) for i in range(img.ndim)]
tshape = [shape_i_op(i)(topgrad) for i in range(topgrad.ndim)]
out_shp = get_conv_output_shape(
ishape,
tshape,
border_mode=border_mode,
subsample=subsample,
filter_dilation=filter_dilation,
)
out_shp = assert_conv_shape(out_shp)
out = GpuAllocEmpty(dtype=img.dtype,
context_name=ctx_name)(*out_shp)
desc = GpuDnnConvDesc(
border_mode=border_mode,
subsample=subsample,
dilation=filter_dilation,
conv_mode="cross",
precision=precision,
)(out.shape)
conv = GpuDnnConv(algo=None, num_groups=num_groups)(img, topgrad,
out, desc)
if conv_mode == "conv":
conv = conv[:, :, ::-1, ::-1]
rval = as_gpuarray_variable(conv.dimshuffle(1, 0, 2, 3), ctx_name)
else:
return None
elif isinstance(op, AbstractConv2d_gradInputs):
if border_mode == "valid" and subsample == (1, 1) and num_groups == 1:
kerns = gpu_contiguous(inp1.dimshuffle(1, 0, 2, 3))
topgrad = gpu_contiguous(inp2)
ctx_name = infer_context_name(kerns, topgrad)
conv_mode = "cross" if conv_mode == "conv" else "conv"
desc = GpuDnnConvDesc(
border_mode="full",
subsample=subsample,
dilation=filter_dilation,
conv_mode=conv_mode,
precision=precision,
)(kerns.shape)
tshape = [shape_i_op(i)(topgrad) for i in range(topgrad.ndim)]
kshape = [shape_i_op(i)(kerns) for i in range(kerns.ndim)]
shape = get_conv_output_shape(
tshape,
kshape,
border_mode="full",
subsample=subsample,
filter_dilation=filter_dilation,
)
shape = assert_conv_shape(shape)
out = GpuAllocEmpty(dtype=topgrad.dtype,
context_name=ctx_name)(*shape)
rval = GpuDnnConv(algo=None, num_groups=num_groups)(topgrad, kerns,
out, desc)
else:
return None
return [rval]
def local_conv2d_gradinputs_cpu(node):
if (not isinstance(node.op, AbstractConv2d_gradInputs) or
node.inputs[0].dtype == 'float16'):
return None
kern, topgrad, shape = node.inputs
if ((not isinstance(kern.type, TensorType) or
not isinstance(topgrad.type, TensorType))):
return None
if node.op.border_mode not in ['full', 'valid']:
return None
if not node.op.filter_flip:
# Not tested yet
return None
if node.op.num_groups > 1 or node.op.unshared:
return None
# Conv 3d implementation, needed when subsample > 2
if node.op.border_mode == 'valid' and node.op.subsample != (1, 1):
# The op don't support that anymore.
return False
# Conv2d Implementation
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
return None
mode = 'valid'
if not node.op.border_mode == 'full':
mode = 'full'
filters = kern.dimshuffle((1, 0, 2, 3))
filters = filters[:, :, ::-1, ::-1]
outshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode,
node.op.subsample,
node.op.filter_dilation)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp,
node.op.border_mode, (1, 1))[2:]
nkern = op_imshp[1]
imshp = (op_kshp[0], outshp[0], outshp[1])
imshp_logical = (op_kshp[0], fulloutshp[0], fulloutshp[1])
din = ConvOp(imshp,
op_kshp[2:],
nkern,
op_imshp[0],
1, 1, output_mode=mode,
unroll_batch=None, unroll_kern=None,
unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=None,
version=-1,
direction_hint='bprop inputs')
din = din(topgrad, filters)
copy_stack_trace(node.outputs[0], din)
din = theano.tensor.patternbroadcast(din, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], din)
return [din]
def local_conv2d_gradweight_cpu(node):
if not isinstance(node.op, AbstractConv2d_gradWeights):
return None
img, topgrad, shape = node.inputs
if ((not isinstance(img.type, TensorType)
or not isinstance(topgrad.type, TensorType))):
return None
if node.op.border_mode not in ['full', 'valid']:
return None
if not node.op.filter_flip:
# Not tested yet
return
if node.op.border_mode == 'valid' and \
(node.op.subsample != (1, 1)):
# Use the gradient as defined in conv3D, because the implementation
# by Conv is slow (about 3x slower than conv3D, and probably 10x
# slower than it could be), and incorrect when subsample > 2.
# build a "node", that should be equivalent to the one given by
# self.make_node, but using convGrad3D instead.
shuffled_img = img.dimshuffle(0, 2, 3, 'x', 1)
shuffled_topgrad = topgrad.dimshuffle(0, 2, 3, 'x', 1)
rval = convGrad3D(V=shuffled_img,
d=(node.op.subsample[0], node.op.subsample[1], 1),
WShape=(shuffled_topgrad.shape[4], shape[0],
shape[1], 1, shuffled_img.shape[4]),
dCdH=shuffled_topgrad)
copy_stack_trace(node.outputs[0], rval)
rval = theano.tensor.addbroadcast(rval, 3)
rval = rval.dimshuffle(0, 4, 1, 2)
rval = rval[:, :, ::-1, ::-1]
rval = theano.tensor.patternbroadcast(rval,
node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], rval)
return [rval]
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
# We cannot infer the shapes
return None
# Determine gradient on kernels
assert len(op_imshp) == 4 and len(op_kshp) == 4
outshp = get_conv_output_shape(op_imshp, op_kshp, node.op.border_mode,
node.op.subsample)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp, node.op.border_mode,
(1, 1))[2:]
newimg = img.dimshuffle((1, 0, 2, 3))
newtopgrad = topgrad.dimshuffle((1, 0, 2, 3))
if node.op.border_mode == 'valid':
(img, filters) = (newimg, newtopgrad)
kshp_logical = fulloutshp
kshp_logical_top_aligned = False
imshp_logical = None
(bsize, nkern) = (op_imshp[1], op_kshp[0])
imshp = (op_imshp[0], op_imshp[2], op_imshp[3])
kshp = outshp
elif node.op.border_mode == 'full':
(img, filters) = (newtopgrad, newimg)
kshp_logical = None
kshp_logical_top_aligned = True
imshp_logical = (op_imshp[0], fulloutshp[0], fulloutshp[1])
(bsize, nkern) = (op_kshp[0], op_imshp[1])
imshp = (op_imshp[0], outshp[0], outshp[1])
kshp = op_imshp[2:]
else:
raise NotImplementedError(
'Only [full,valid] modes are currently supported.')
# Flip the kernels
filters = filters[:, :, ::-1, ::-1]
dw = ConvOp(imshp,
kshp,
nkern,
bsize,
1,
1,
output_mode='valid',
unroll_batch=None,
unroll_kern=None,
unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=kshp_logical,
kshp_logical_top_aligned=kshp_logical_top_aligned,
direction_hint='bprop weights')
res = dw(img, filters)
copy_stack_trace(node.outputs[0], res)
if node.op.border_mode == 'valid':
res = res.dimshuffle((1, 0, 2, 3))
res = res[:, :, ::-1, ::-1]
res = theano.tensor.patternbroadcast(res, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], res)
return [res]
def local_conv2d_gradinputs_cpu(node):
if not isinstance(node.op, AbstractConv2d_gradInputs):
return None
kern, topgrad, shape = node.inputs
if ((not isinstance(kern.type, TensorType)
or not isinstance(topgrad.type, TensorType))):
return None
if node.op.border_mode not in ['full', 'valid']:
return None
if not node.op.filter_flip:
# Not tested yet
return None
# Conv 3d implementation, needed when subsample > 2
if node.op.border_mode == 'valid' and node.op.subsample != (1, 1):
kern = kern[:, :, ::-1, ::-1]
shuffled_kern = kern.dimshuffle(0, 2, 3, 'x', 1)
shuffled_topgrad = topgrad.dimshuffle(0, 2, 3, 'x', 1)
b = theano.tensor.zeros_like(shuffled_kern[0, 0, 0, 0, :])
rval = convTransp3D(W=shuffled_kern,
b=b,
d=(node.op.subsample[0], node.op.subsample[1], 1),
H=shuffled_topgrad,
RShape=(shape[0], shape[1], 1))
copy_stack_trace(node.outputs[0], rval)
rval = theano.tensor.addbroadcast(rval, 3)
rval = rval.dimshuffle(0, 4, 1, 2)
rval = theano.tensor.patternbroadcast(rval,
node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], rval)
return [rval]
# Conv2d Implementation
dx, dy = node.op.subsample
if dx not in (1, 2) or dy not in (1, 2):
# Not implemented in the gradient of ConvOp
return None
if node.op.imshp is None:
op_imshp = (None, None, None, None)
else:
op_imshp = node.op.imshp
if node.op.kshp is None:
op_kshp = (None, None, None, None)
else:
op_kshp = node.op.kshp
if None in op_imshp or None in op_kshp:
if (dx, dy) != (1, 1):
return None
mode = 'valid'
if not node.op.border_mode == 'full':
mode = 'full'
filters = kern.dimshuffle((1, 0, 2, 3))
filters = filters[:, :, ::-1, ::-1]
outshp = get_conv_output_shape(op_imshp, op_kshp, node.op.border_mode,
node.op.subsample)[2:]
fulloutshp = get_conv_output_shape(op_imshp, op_kshp, node.op.border_mode,
(1, 1))[2:]
nkern = op_imshp[1]
imshp = (op_kshp[0], outshp[0], outshp[1])
imshp_logical = (op_kshp[0], fulloutshp[0], fulloutshp[1])
din = ConvOp(imshp,
op_kshp[2:],
nkern,
op_imshp[0],
1,
1,
output_mode=mode,
unroll_batch=None,
unroll_kern=None,
unroll_patch=None,
imshp_logical=imshp_logical,
kshp_logical=None,
version=-1,
direction_hint='bprop inputs')
din = din(topgrad, filters)
copy_stack_trace(node.outputs[0], din)
din = theano.tensor.patternbroadcast(din, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], din)
return [din]
def c_code(self, node, name, inp, out, sub):
x, = inp
dH, dW = self.subsample
if self.imshp is None:
self.imshp = x.shape
i_n, i_c, i_h, i_w = self.imshp
if len(self.kshp) == 5:
grp, k_n, k_c, k_h, k_w = self.kshp
assert i_c == k_c * grp
else:
k_n, k_c, k_h, k_w = self.kshp
grp = 1
o_n, o_c, o_h, o_w = get_conv_output_shape(
image_shape=self.imshp,
kernel_shape=self.kshp,
border_mode=self.border_mode,
filter_dilation=self.filter_dilation,
subsample=self.subsample)
if self.border_mode == 'valid':
padH, padW = (0, 0)
elif self.border_mode == 'full':
padH, padW = ((k_h - 1), (k_w - 1))
elif self.border_mode == 'half':
padH, padW = ((k_h / 2), (k_w / 2))
elif isinstance(self.border_mode, tuple):
padH, padW = self.border_mode
else:
raise ValueError("border_mode must have two elements")
z, = out
if 'float32' == node.inputs[0].type.dtype:
precision = 'F32'
elif 'float64' == node.inputs[0].type.dtype:
precision = 'F64'
else:
raise Exception("Type %s is not supported!" %
node.inputs[0].type.dtype)
fail = sub['fail']
ccode = """
if (1 == first_run) {
int convPadding[2];
size_t convStride[2], weightSize[5], weightStride[5], imageSize[4], imageStride[4], zSize[4], zStride[4];
convStride[0] = %(dW)s;
convStride[1] = %(dH)s;
convPadding[0] = -%(padW)s;
convPadding[1] = -%(padH)s;
imageSize[0] = %(i_w)s; //w
imageSize[1] = %(i_h)s; //h
imageSize[2] = %(i_c)s; //c
imageSize[3] = %(i_n)s; //n
imageStride[0] = 1;
imageStride[1] = imageSize[0];
imageStride[2] = imageSize[0] * imageSize[1];
imageStride[3] = imageSize[0] * imageSize[1] * imageSize[2];
weightSize[0] = %(k_w)s;
weightSize[1] = %(k_h)s;
weightSize[2] = %(k_c)s;
weightSize[3] = %(k_n)s;
weightSize[4] = %(grp)s;
weightStride[0] = 1;
weightStride[1] = weightSize[0];
weightStride[2] = weightSize[0] * weightSize[1];
weightStride[3] = weightSize[0] * weightSize[1] * weightSize[2];
weightStride[4] = weightSize[0] * weightSize[1] * weightSize[2] * weightSize[3];
zSize[0] = %(o_w)s;
zSize[1] = %(o_h)s;
zSize[2] = %(o_c)s;
zSize[3] = %(o_n)s;
zStride[0] = 1;
zStride[1] = zSize[0];
zStride[2] = zSize[0] * zSize[1];
zStride[3] = zSize[0] * zSize[1] * zSize[2];
const int group = %(grp)s;
//create user layout
CHECK_ERR( dnnLayoutCreate_%(precision)s(&layout_user, DIMENSION, imageSize, imageStride), err );
CHECK_ERR( dnnGroupsConvolutionCreateForward_%(precision)s(&primitive, NULL,
dnnAlgorithmConvolutionDirect, group, DIMENSION, imageSize, zSize,
weightSize, convStride, convPadding, dnnBorderZeros), err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_%(precision)s(&layout_internal, primitive, dnnResourceSrc), err );
}
if (!dnnLayoutCompare_%(precision)s(layout_user, layout_internal))
{
if (NULL == to_internal)
{
CHECK_ERR( dnnConversionCreate_%(precision)s(&to_internal, layout_user, layout_internal), err );
}
}
if (NULL == %(z)s)
{
//Create PyArrayObject for output
%(z)s = (PyArrayObject*)PyArray_ZEROS(DIMENSION, PyArray_DIMS(%(x)s), PyArray_TYPE(%(x)s), 0);
if (NULL == %(z)s)
{
%(fail)s
}
}
if (NULL == internal_buf)
{
CHECK_ERR( dnnAllocateBuffer_%(precision)s((void**)&internal_buf, layout_internal), err );
}
if (to_internal)
{
convert_resources[dnnResourceFrom] = (PyArray_DATA(%(x)s));
convert_resources[dnnResourceTo] = (void*)(internal_buf);
CHECK_ERR( dnnExecute_%(precision)s(to_internal, convert_resources), err );
}
else
{
internal_buf = (PyArray_DATA(%(x)s));
}
if (layout_internal != ((dnnLayout_t*)PyArray_DATA(%(z)s))[0])
{
((dnnLayout_t*)PyArray_DATA(%(z)s))[0] = layout_internal;
}
if (internal_buf != ((void**)PyArray_DATA(%(z)s))[1])
{
((void**)PyArray_DATA(%(z)s))[1] = internal_buf;
}
first_run = 0;
#ifdef _MKL_DEBUG_
std::cout << "U2IConv2D: from buffer: " << convert_resources[dnnResourceFrom] << " to buffer: " << convert_resources[dnnResourceTo] << std::endl;
#endif
""" % locals()
return ccode