def local_inv_1_plus_exp(node):
"""
1/(1+exp(x)) -> sigm(-x)
"""
# this optimization should be done for numerical stability
# so we don't care to check client counts
if node.op == tensor.inv:
inv_arg = node.inputs[0]
if inv_arg.owner and inv_arg.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = \
opt.scalarconsts_rest(inv_arg.owner.inputs)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
if nonconsts[0].owner and nonconsts[0].owner.op == tensor.exp:
if scalars and numpy.allclose(numpy.sum(scalars), 1):
out = opt._fill_chain(
sigmoid(tensor.neg(nonconsts[0].owner.inputs[0])),
scalar_inputs)
# keep combined stack traces of
# exp(x): nonconsts[0],
# 1 + exp(x): inv_arg,
# 1 / (1 + exp(x)): node.outputs[0]
copy_stack_trace(
[nonconsts[0], inv_arg, node.outputs[0]], out)
return out
def local_ultra_fast_sigmoid(node):
"""
When enabled, change all sigmoid to ultra_fast_sigmoid.
For example do mode.including('local_ultra_fast_sigmoid')
or use the Theano flag optimizer_including=local_ultra_fast_sigmoid.
This speeds up the sigmoid op by using an approximation.
This is done after the stabilization and specialize phases
to avoid interacting with them.
"""
if (isinstance(node.op, tensor.Elemwise)
and node.op.scalar_op == scalar_sigmoid):
out = ultra_fast_sigmoid(node.inputs[0])
copy_stack_trace(node.outputs[0], out)
def values_eq_approx_remove_low_prec(a, b):
# atol is found by trial/error.
# Other test could fail without good reason.
return tensor.TensorType.values_eq_approx(a, b, atol=0.02)
# Let DebugMode know that there this opt approx the values.
out.tag.values_eq_approx = values_eq_approx_remove_low_prec
return [out]
def local_inv_1_plus_exp(node):
"""
1/(1+exp(x)) -> sigm(-x)
"""
# this optimization should be done for numerical stability
# so we don't care to check client counts
if node.op == tensor.inv:
inv_arg = node.inputs[0]
if inv_arg.owner and inv_arg.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = \
opt.scalarconsts_rest(inv_arg.owner.inputs)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
if nonconsts[0].owner and nonconsts[0].owner.op == tensor.exp:
if scalars and numpy.allclose(numpy.sum(scalars), 1):
out = opt._fill_chain(
sigmoid(
tensor.neg(nonconsts[0].owner.inputs[0])),
scalar_inputs)
# keep combined stack traces of
# exp(x): nonconsts[0],
# 1 + exp(x): inv_arg,
# 1 / (1 + exp(x)): node.outputs[0]
copy_stack_trace(
[nonconsts[0], inv_arg, node.outputs[0]], out)
return out
def local_ultra_fast_sigmoid(node):
"""
When enabled, change all sigmoid to ultra_fast_sigmoid.
For example do mode.including('local_ultra_fast_sigmoid')
or use the Theano flag optimizer_including=local_ultra_fast_sigmoid.
This speeds up the sigmoid op by using an approximation.
This is done after the stabilization and specialize phases
to avoid interacting with them.
"""
if (isinstance(node.op, tensor.Elemwise) and
node.op.scalar_op == scalar_sigmoid):
out = ultra_fast_sigmoid(node.inputs[0])
copy_stack_trace(node.outputs[0], out)
def values_eq_approx_remove_low_prec(a, b):
# atol is found by trial/error.
# Other test could fail without good reason.
return tensor.TensorType.values_eq_approx(a, b, atol=0.02)
# Let DebugMode know that there this opt approx the values.
out.tag.values_eq_approx = values_eq_approx_remove_low_prec
return [out]
def local_inplace_DiagonalSubtensor(node):
"""Also work for IncDiagonalSubtensor."""
if isinstance(node.op, (DiagonalSubtensor, IncDiagonalSubtensor)) and not node.op.inplace:
new_op = node.op.__class__(inplace=True)
new_node = new_op(*node.inputs)
copy_stack_trace(node.outputs[0], new_node)
return [new_node]
return False
def local_inplace_sparse_block_outer(node):
"""
SparseBlockOuter(inplace=False) -> SparseBlockOuter(inplace=True)
"""
if isinstance(node.op, SparseBlockOuter) and not node.op.inplace:
new_node = sparse_block_outer_inplace(*node.inputs)
copy_stack_trace(node.outputs[0], new_node)
return [new_node]
return False
def local_inplace_DiagonalSubtensor(node):
"""Also work for IncDiagonalSubtensor."""
if (isinstance(node.op, (DiagonalSubtensor, IncDiagonalSubtensor)) and
not node.op.inplace):
new_op = node.op.__class__(inplace=True)
new_node = new_op(*node.inputs)
copy_stack_trace(node.outputs[0], new_node)
return [new_node]
return False
def local_inplace_sparse_block_outer(node):
"""
SparseBlockOuter(inplace=False) -> SparseBlockOuter(inplace=True)
"""
if isinstance(node.op, SparseBlockOuter) and not node.op.inplace:
new_node = sparse_block_outer_inplace(*node.inputs)
copy_stack_trace(node.outputs[0], new_node)
return [new_node]
return False
def local_hard_sigmoid(node):
if (isinstance(node.op, tensor.Elemwise) and
node.op.scalar_op == scalar_sigmoid):
out = hard_sigmoid(node.inputs[0])
copy_stack_trace(node.outputs[0], out)
def values_eq_approx_remove_low_prec(a, b):
# atol is found by trial/error.
# Other test could fail without good reason.
return tensor.TensorType.values_eq_approx(a, b, atol=0.1)
# Let DebugMode know that there this opt approx the values.
out.tag.values_eq_approx = values_eq_approx_remove_low_prec
return [out]
def local_1msigmoid(node):
"""
1-sigm(x) -> sigm(-x)
"""
if node.op == tensor.sub:
sub_l, sub_r = node.inputs
if len(sub_r.clients) > 1:
return # graph is using both sigm and 1-sigm
if sub_r.owner and sub_r.owner.op == sigmoid:
try:
val_l = opt.get_scalar_constant_value(sub_l)
except Exception:
return
if numpy.allclose(numpy.sum(val_l), 1):
out = sigmoid(-sub_r.owner.inputs[0])
copy_stack_trace([sub_r, node.outputs[0]], out)
return [out]
def local_1msigmoid(node):
"""
1-sigm(x) -> sigm(-x)
"""
if node.op == tensor.sub:
sub_l, sub_r = node.inputs
if len(sub_r.clients) > 1:
return # graph is using both sigm and 1-sigm
if sub_r.owner and sub_r.owner.op == sigmoid:
try:
val_l = opt.get_scalar_constant_value(sub_l)
except Exception:
return
if numpy.allclose(numpy.sum(val_l), 1):
out = sigmoid(-sub_r.owner.inputs[0])
copy_stack_trace([sub_r, node.outputs[0]], out)
return [out]
def local_abstractconv_gemm(node):
if theano.config.cxx == "" or not theano.config.blas.ldflags:
return
if not isinstance(node.op, AbstractConv2d):
return None
img, kern = node.inputs
if not isinstance(img.type, TensorType) or \
not isinstance(kern.type, TensorType):
return None
# need to flip the kernel if necessary
if node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1]
rval = CorrMM(border_mode=node.op.border_mode,
subsample=node.op.subsample)(img, kern)
copy_stack_trace(node.outputs[0], rval)
return [rval]
def local_abstractconv_gemm(node):
if theano.config.cxx == "" or not theano.config.blas.ldflags:
return
if not isinstance(node.op, AbstractConv2d):
return None
img, kern = node.inputs
if not isinstance(img.type, TensorType) or \
not isinstance(kern.type, TensorType):
return None
# need to flip the kernel if necessary
if node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1]
rval = CorrMM(border_mode=node.op.border_mode,
subsample=node.op.subsample)(img, kern)
copy_stack_trace(node.outputs[0], rval)
return [rval]
def local_abstractconv_gradinputs_gemm(node):
if theano.config.cxx == "":
return
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
# need to flip the kernel if necessary
if node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1]
rval = CorrMM_gradInputs(border_mode=node.op.border_mode,
subsample=node.op.subsample)(kern, topgrad, shape)
copy_stack_trace(node.outputs[0], rval)
return [rval]
def local_abstractconv_gradinputs_gemm(node):
if theano.config.cxx == "":
return
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
# need to flip the kernel if necessary
if node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1]
rval = CorrMM_gradInputs(border_mode=node.op.border_mode,
subsample=node.op.subsample)(kern, topgrad,
shape)
copy_stack_trace(node.outputs[0], rval)
return [rval]
def local_abstractconv_gradweight_gemm(node):
if theano.config.cxx == "" or not theano.config.blas.ldflags:
return
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
rval = CorrMM_gradWeights(border_mode=node.op.border_mode,
subsample=node.op.subsample)(img, topgrad, shape)
copy_stack_trace(node.outputs[0], rval)
# need to flip the kernel if necessary
if node.op.filter_flip:
rval = rval[:, :, ::-1, ::-1]
rval = theano.tensor.patternbroadcast(rval, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], rval)
return [rval]
def local_abstractconv_gradweight_gemm(node):
if theano.config.cxx == "" or not theano.config.blas.ldflags:
return
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
rval = CorrMM_gradWeights(border_mode=node.op.border_mode,
subsample=node.op.subsample)(img, topgrad, shape)
copy_stack_trace(node.outputs[0], rval)
# need to flip the kernel if necessary
if node.op.filter_flip:
rval = rval[:, :, ::-1, ::-1]
rval = theano.tensor.patternbroadcast(rval, node.outputs[0].broadcastable)
copy_stack_trace(node.outputs[0], rval)
return [rval]
def local_to_gpu(node):
"""
op(host_from_gpu()) -> host_from_gpu(op)
gpu_from_host(op) -> op(gpu_from_host)
"""
if isinstance(node.op, op):
# op(host_from_gpu()) -> host_from_gpu(op)
# If any of the input that go on the GPU are on the GPU,
# move the op to the gpu.
if any(node.inputs[idx].owner and
isinstance(node.inputs[idx].owner.op, cuda.HostFromGpu)
for idx in to_gpu):
new_inp = list(node.inputs)
for idx in to_gpu:
new_inp[idx] = cuda.gpu_from_host(new_inp[idx])
result_node = op()(*new_inp)
copy_stack_trace(node.outputs[0], result_node)
transfer_node = cuda.host_from_gpu(result_node)
copy_stack_trace(node.outputs[0], transfer_node)
return [transfer_node]
if node.op == cuda.gpu_from_host:
# gpu_from_host(op) -> op(gpu_from_host)
host_input = node.inputs[0]
if host_input.owner and isinstance(host_input.owner.op,
op):
op_node = host_input.owner
new_inp = list(op_node.inputs)
for idx in to_gpu:
new_inp[idx] = cuda.gpu_from_host(new_inp[idx])
new_node = op()(*new_inp)
copy_stack_trace(host_input, new_node)
return [new_node]
return False
def local_conv2d_cpu(node):
if not isinstance(node.op, AbstractConv2d):
return None
img, kern = node.inputs
if ((not isinstance(img.type, TensorType) or
not isinstance(kern.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
rval = conv2d(img, kern,
node.op.imshp, node.op.kshp,
border_mode=node.op.border_mode,
subsample=node.op.subsample)
copy_stack_trace(node.outputs[0], rval)
return [rval]
def local_sigm_times_exp(node):
"""
exp(x) * sigm(-x) -> sigm(x)
exp(-x) * sigm(x) -> sigm(-x)
"""
# Bail early if it is not a multiplication.
if node.op != tensor.mul:
return None
# Obtain tree of multiplications starting at this node.
mul_tree = parse_mul_tree(node.outputs[0])
# Perform core optimization.
did_something = perform_sigm_times_exp(mul_tree)
if not did_something:
# No change.
return None
# The optimization may have introduced multiplications by 1 in the tree:
# get rid of them.
mul_tree = simplify_mul(mul_tree)
# Recompute final output based on the updated tree.
out = compute_mul(mul_tree)
# keep the stack trace
copy_stack_trace(node.outputs[0], out)
return [out]
def local_sigm_times_exp(node):
"""
exp(x) * sigm(-x) -> sigm(x)
exp(-x) * sigm(x) -> sigm(-x)
"""
# Bail early if it is not a multiplication.
if node.op != tensor.mul:
return None
# Obtain tree of multiplications starting at this node.
mul_tree = parse_mul_tree(node.outputs[0])
# Perform core optimization.
did_something = perform_sigm_times_exp(mul_tree)
if not did_something:
# No change.
return None
# The optimization may have introduced multiplications by 1 in the tree:
# get rid of them.
mul_tree = simplify_mul(mul_tree)
# Recompute final output based on the updated tree.
out = compute_mul(mul_tree)
# keep the stack trace
copy_stack_trace(node.outputs[0], out)
return [out]
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 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]