def __init__(self, in_info, out_info, lua_fw_func, lua_bw_func=None, lua_file=None, name=None):
_init()
super(TorchWrapperOp, self).__init__()
self.in_info = make_hashable(in_info)
self.out_info = make_hashable(out_info)
self.lua_file = lua_file # if none, expect inplace definition
self.lua_fw_func = lua_fw_func
self.lua_bw_func = lua_bw_func
self.name = name or "<anonymous>"
for info in self.in_info + self.out_info:
assert "ndim" in info
assert "shape" in info
assert len(info["shape"]) == info["ndim"]
for info in self.out_info:
for s in info["shape"]:
assert s, "need output shape info or reference, %r" % info
def get_global_instance(cls, sprint_opts):
sprint_opts = make_hashable(sprint_opts)
if sprint_opts in cls.global_instances:
return cls.global_instances[sprint_opts]
instance = SprintInstancePool(sprint_opts=sprint_opts)
cls.global_instances[sprint_opts] = instance
return instance
def get_global_instance(cls, sprint_opts):
sprint_opts = make_hashable(sprint_opts)
if sprint_opts in cls.global_instances:
return cls.global_instances[sprint_opts]
instance = SprintInstancePool(sprint_opts=sprint_opts)
cls.global_instances[sprint_opts] = instance
return instance
def __init__(self, in_info, out_info,
c_fw_code, c_bw_code=None, c_extra_support_code=None, code_version=None,
grad_input_map=None, name=None):
"""
:param list[dict(str)] in_info: each dict describes one input var.
attribs in the dict:
int ndim: the ndim.
tuple shape: tuple and can contain None for specific dimensions.
optional attribs:
str dtype: "float32" by default.
bool need_contiguous: false by default.
int want_inplace: -1 by default. try to optimize to destroy input, on output-index.
"dummy_out" is a special value which will add another output.
bool is_inplace: false by default. whether the optimization was applied.
str gradient: can be "disconnected". see grad().
bool bw_input: True by default. add this param to the bw input.
other attribs are just ignored.
:param list[dict(str)] out_info: like in_info.
slightly different behavior for:
shape: we also allow refs to the in_info in the form (in-idx,dim). see infer_shape().
need_contiguous/want_inplace: used for bw, in case for bw_input == True.
:param str c_fw_code: C code for forward pass
:param str c_extra_support_code: C support code (for c_support_code)
:param str|None c_bw_code: C code for backward pass (for gradient)
:param tuple[int] code_version: will be returned by c_code_cache_version.
:param tuple|callable grad_input_map: selection of grad inputs.
by default, we get all inputs + all outputs + all grad outputs.
:param str name: name
"""
super(NativeOp, self).__init__()
assert isinstance(in_info, (list, tuple))
assert isinstance(out_info, (list, tuple))
in_info, out_info, num_dummy_outs = self._resolve_want_inplace_dummy(in_info, out_info)
self.in_info = make_hashable(in_info)
self.out_info = make_hashable(out_info)
self.num_dummy_outs = num_dummy_outs
self.c_fw_code = c_fw_code
self.c_bw_code = c_bw_code
self.c_extra_support_code = self._reduce_c_extra_support_code(c_extra_support_code)
self.code_version = code_version or ()
self.name = name or "<anonNativeOp>"
self.grad_input_map = self._convert_grad_input_map(grad_input_map, len(in_info) + len(out_info) * 2)
self.destroy_map = self._construct_destroy_map(in_info)
def _get_optimizer_item_for_opts(self, optimizer_opts, auto_create_new):
"""
:param dict[str]|str|None optimizer_opts:
:param bool auto_create_new:
:return: key, optimizer
:rtype: (object, tf.train.Optimizer)
"""
from Util import make_hashable
key = make_hashable(optimizer_opts)
if key in self.optimizers:
return key, self.optimizers[key]
assert auto_create_new, "no optimizer found for opts %r" % (optimizer_opts,)
optimizer = self._create_optimizer(optimizer_opts)
self.optimizers[key] = optimizer
return key, optimizer
def generate_hdf_from_other(opts):
"""
:param dict[str] opts:
:return: hdf filename
:rtype: str
"""
# See test_hdf_dump.py and tools/hdf_dump.py.
from Util import make_hashable
cache_key = make_hashable(opts)
if cache_key in _hdf_cache:
return _hdf_cache[cache_key]
fn = _get_tmp_file(suffix=".hdf")
from Dataset import init_dataset
dataset = init_dataset(opts)
hdf_dataset = HDFDatasetWriter(fn)
hdf_dataset.dump_from_dataset(dataset)
hdf_dataset.close()
_hdf_cache[cache_key] = fn
return fn
def generate_hdf_from_other(opts):
"""
:param dict[str] opts:
:return: hdf filename
:rtype: str
"""
# See test_hdf_dump.py and tools/hdf_dump.py.
from Util import make_hashable
cache_key = make_hashable(opts)
if cache_key in _hdf_cache:
return _hdf_cache[cache_key]
fn = _get_tmp_file(suffix=".hdf")
from Dataset import init_dataset
dataset = init_dataset(opts)
hdf_dataset = HDFDatasetWriter(fn)
hdf_dataset.dump_from_dataset(dataset)
hdf_dataset.close()
_hdf_cache[cache_key] = fn
return fn
def get_apply_grads_op(self, loss, var_list):
"""
:param tf.Tensor loss:
:param list[tf.Variable] var_list:
:return: op with all variable updates combined, using the optimizer
:rtype: tf.Operation
"""
# The following code is basically extended self.optimizer.minimize(), to optionally modify gradients.
from Util import make_hashable
if not var_list:
return tf.no_op(name="no_grad_vars_no_op")
grads_and_vars = self._compute_gradients(loss, var_list=var_list)
if self.config.is_true("use_horovod") and self.config.value("horovod_reduce_type", "") == "grad":
# noinspection PyPackageRequirements,PyUnresolvedReferences
import horovod.tensorflow as hvd
grads_and_vars = [
(hvd.allreduce(grad, average=self.config.is_true("horovod_avg_grad")) if grad is not None else None, var)
for (grad, var) in grads_and_vars]
var_grads = {var: grad for (grad, var) in grads_and_vars if grad is not None}
if not var_grads:
raise Exception("no single variable to train")
global_info = self._GetGlobalInfo(optimizer=self, all_vars=var_list, var_grads=var_grads)
if self.config.bool_or_other("debug_grad_summaries", False):
tf.summary.scalar("global_grad_norm", global_info.get_global_grad_norm())
grads_per_apply_grad_opts = {} # dict apply_grad_opts -> list of (grad, var)
for grad, var in grads_and_vars:
assert var in var_list
if grad is None:
continue
new_grad, apply_grad_opts = self._post_process_grad(grad=grad, var=var, global_info=global_info)
grads_per_apply_grad_opts.setdefault(make_hashable(apply_grad_opts), []).append((new_grad, var))
all_apply_grads = []
assert grads_per_apply_grad_opts
for apply_grad_opts, grads_and_vars_per_opts in grads_per_apply_grad_opts.items():
all_apply_grads.append(self._apply_gradients(grads_and_vars_per_opts, **apply_grad_opts))
if len(all_apply_grads) == 1:
return all_apply_grads[0]
return tf.group(*all_apply_grads)
def __init__(self, sprint_opts): super(SprintAlignmentAutomataOp, self).__init__() self.sprint_opts = make_hashable(sprint_opts) self.sprint_instance_pool = None # type: typing.Optional[SprintInstancePool]
def __init__(self, sprint_opts): super(SprintErrorSigOp, self).__init__() self.sprint_opts = make_hashable(sprint_opts) self.sprint_instance_pool = None # type: typing.Optional[SprintInstancePool] self.debug_perform_time = None
def __init__(self, sprint_opts):
super(SprintAlignmentAutomataOp, self).__init__()
self.sprint_opts = make_hashable(sprint_opts)
self.sprint_instance_pool = None # type: typing.Optional[SprintInstancePool]
def __init__(self, sprint_opts):
super(SprintErrorSigOp, self).__init__()
self.sprint_opts = make_hashable(sprint_opts)
self.sprint_instance_pool = None
self.debug_perform_time = None
def __init__(self, target, sprint_opts): super(SprintErrorSigOp, self).__init__() self.target = target # default is "classes" self.sprint_opts = make_hashable(sprint_opts) self.sprint_instance_pool = None
def __init__(self, sprint_opts): super(SprintErrorSigOp, self).__init__() self.sprint_opts = make_hashable(sprint_opts) self.sprint_instance_pool = None self.debug_perform_time = None
def __init__(self, sprint_opts):
super(SprintAlignmentAutomataOp, self).__init__()
self.sprint_opts = make_hashable(sprint_opts)
self.sprint_instance_pool = None
def __init__(self, sprint_opts): super(SprintAlignmentAutomataOp, self).__init__() self.sprint_opts = make_hashable(sprint_opts) self.sprint_instance_pool = None
def __init__(self, sprint_opts):
super(SprintErrorSigOp, self).__init__()
self.sprint_opts = make_hashable(sprint_opts)
self.sprint_instance_pool = None # type: typing.Optional[SprintInstancePool]
self.debug_perform_time = None