def f(local_opt):
name = (kwargs and kwargs.pop('name')) or local_opt.__name__
optdb.register(
name, TopoOptimizer(
local_opt, failure_callback=TopoOptimizer.warn_inplace),
60, 'fast_run', 'inplace', 'gpuarray', *tags)
return local_opt
for idx in to_gpu:
new_inp[idx] = cuda.gpu_from_host(new_inp[idx])
return [op()(*new_inp)]
return False
local_to_gpu.__name__ = "local_to_gpu_" + op.__name__
cuda.opt.register_opt()(local_to_gpu)
if cuda.cuda_available:
make_gpu_optimizer(DiagonalSubtensor, [0])
make_gpu_optimizer(IncDiagonalSubtensor, [0, 3])
@theano.gof.local_optimizer([DiagonalSubtensor, IncDiagonalSubtensor])
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)
return [new_node]
return False
theano.compile.optdb.register(
'local_inplace_DiagonalSubtensor',
TopoOptimizer(local_inplace_DiagonalSubtensor,
failure_callback=TopoOptimizer.warn_inplace), 60, 'fast_run',
'inplace')
self._inplace))
def c_code_cache_version(self):
"""
.. todo::
WRITEME
"""
return (8, )
@local_optimizer([CrossMapNormUndo])
def local_crossmapnormundo_inplace(node):
"""
.. todo::
WRITEME
"""
if isinstance(node.op, CrossMapNormUndo) and not node.op.inplace:
new_op = node.op.as_inplace()
new_node = new_op(*node.inputs)
return new_node
return False
theano.compile.optdb.register(
'local_crossmapnormundo_inplace',
TopoOptimizer(local_crossmapnormundo_inplace,
failure_callback=TopoOptimizer.warn_inplace), 80, 'fast_run',
'inplace')
from theano import compile, gof
from theano.gof import TopoOptimizer
from theano.typed_list.basic import Append, Extend, Insert, Remove, Reverse
@gof.local_optimizer([Append, Extend, Insert, Reverse, Remove], inplace=True)
def typed_list_inplace_opt(node):
if (isinstance(node.op, (Append, Extend, Insert, Reverse, Remove))
and not node.op.inplace):
new_op = node.op.__class__(inplace=True)
new_node = new_op(*node.inputs)
return [new_node]
return False
compile.optdb.register(
"typed_list_inplace_opt",
TopoOptimizer(typed_list_inplace_opt,
failure_callback=TopoOptimizer.warn_inplace),
60,
"fast_run",
"inplace",
)
if node.op.what_to_output == 0:
return [GpuLargeSparseTargets(node.op.what_to_output)(*node.inputs)]
elif node.op.what_to_output == 1:
return [
host_from_gpu(
GpuLargeSparseTargets(node.op.what_to_output)(*node.inputs))
]
else:
out = GpuLargeSparseTargets(node.op.what_to_output)(*node.inputs)
return [out[0], host_from_gpu(out[1])]
optdb.register(
"local_large_sparse_targets_gpu",
TopoOptimizer(
local_optimizer([LargeSparseTargets])(local_large_sparse_targets_gpu)),
49, "fast_run")
def optimize_large_sparse_target(inputs, H, outputs, updates):
"""
TODO: WRITEME
"""
# need to rewrite MergeLargeSparseTargetOps because there will be multiple
# updates containing gradH!
if not isinstance(updates, OrderedDict):
raise ValueError("Updates needs to be OrderedDict otherwise keys, and"
" values may not match after optimization")
