def logpy_group_incsubtensor(node):
# TODO: how to design re-usable patterns? (dtype, ndim, etc.)
shape_of = node.fgraph.shape_feature.shape_of
shape_dimo = goalifyN(
shape_dim(shape_of))
ndimo = goalify(lambda x: getattr(x, 'ndim'))
x = node.outputs[0].type()
if x.ndim == 0:
return
y = x[0].type()
z = tensor.set_subtensor(x[1001], y)
incs = []
orig_out = node.outputs[0]
while node:
with variables(x, y, 1001):
match = run(1, (x, y, 1001), (eq, node.outputs[0], z))
if match:
xx, yy, ii = match[0]
incs.append((ii, xx, yy))
node = xx.owner
continue
break
if not incs:
return
incs.sort()
if zip(*incs)[0] == tuple(range(shape_dim(shape_of)(xx, 0))):
iin = tensor.concatenate([
tensor.shape_padleft(yy)
for ii, _, yy in incs])
print 'INCS', incs
return [iin]
def goalifyN(func):
funco = goalify(func)
def goalo(args, result):
tmp = var()
return (lall,
(eq, tmp, args),
(funco, tmp, result))
return goalo
def logpy_remove_dot_scalar_matrix(node):
# TODO: how to design re-usable patterns? (dtype, ndim, etc.)
shape_of = node.fgraph.shape_feature.shape_of
shape_dimo = goalifyN(
shape_dim(shape_of))
ndimo = goalify(lambda x: getattr(x, 'ndim'))
x = theano.tensor.matrix() # -- XXX type should not matter
y = theano.tensor.matrix() # -- XXX type should not matter
if isinstance(node.op, theano.tensor.Dot):
with variables(x, y):
#theano.printing.debugprint(tensor.dot(x, y))
result = run(1, (x, y),
(eq, node, tensor.dot(x, y).owner),
(ndimo, x, 2),
(shape_dimo, (x, 0), 1),
(shape_dimo, (x, 1), 1),
)
if result:
xx, yy = result[0]
#print 'MATCHED xx!', xx, shape_of[xx], xx.type
#print 'MATCHED yy!', yy, shape_of[yy], yy.type
#theano.printing.debugprint(xx)
return [tensor.addbroadcast(xx, 0, 1).dimshuffle() * yy]