def reshape(self, data_shapes, label_shapes=None):
"""Reshapes the module for new input shapes.
Parameters
----------
data_shapes : list of (str, tuple)
Typically is ``data_iter.provide_data``.
label_shapes : list of (str, tuple)
Typically is ``data_iter.provide_label``.
"""
assert self.binded
self._data_shapes, self._label_shapes = _parse_data_desc(
self.data_names, self.label_names, data_shapes, label_shapes)
self._exec_group.reshape(self._data_shapes, self._label_shapes)
def reshape(self, data_shapes, label_shapes=None):
"""Reshape the module for new input shapes.
Parameters
----------
data_shapes : list of (str, tuple)
Typically is `data_iter.provide_data`.
label_shapes : list of (str, tuple)
Typically is `data_iter.provide_label`.
"""
assert self.binded
# self._data_shapes, self._label_shapes = _parse_data_desc(
# self.data_names, self.label_names, data_shapes, label_shapes)
self._data_shapes, self._label_shapes = zip(*[_parse_data_desc(self.data_names, self.label_names, data_shape, label_shape)
for data_shape, label_shape in zip(data_shapes, label_shapes)])
self._exec_group.reshape(self._data_shapes, self._label_shapes)
def bind(self, data_shapes, label_shapes=None, for_training=True,
inputs_need_grad=False, force_rebind=False, shared_module=None,
grad_req='write'):
"""Bind the symbols to construct executors. This is necessary before one
can perform computation with the module.
Parameters
----------
data_shapes : list of (str, tuple)
Typically is `data_iter.provide_data`.
label_shapes : list of (str, tuple)
Typically is `data_iter.provide_label`.
for_training : bool
Default is `True`. Whether the executors should be bind for training.
inputs_need_grad : bool
Default is `False`. Whether the gradients to the input data need to be computed.
Typically this is not needed. But this might be needed when implementing composition
of modules.
force_rebind : bool
Default is `False`. This function does nothing if the executors are already
binded. But with this `True`, the executors will be forced to rebind.
shared_module : Module
Default is `None`. This is used in bucketing. When not `None`, the shared module
essentially corresponds to a different bucket -- a module with different symbol
but with the same sets of parameters (e.g. unrolled RNNs with different lengths).
"""
# force rebinding is typically used when one want to switch from
# training to prediction phase.
if force_rebind:
self._reset_bind()
if self.binded:
self.logger.warning('Already binded, ignoring bind()')
return
self.for_training = for_training
self.inputs_need_grad = inputs_need_grad
self.binded = True
self._grad_req = grad_req
if not for_training:
assert not inputs_need_grad
else:
pass
# this is not True, as some module might not contains a loss function
# that consumes the labels
# assert label_shapes is not None
# self._data_shapes, self._label_shapes = _parse_data_desc(
# self.data_names, self.label_names, data_shapes, label_shapes)
self._data_shapes, self._label_shapes = zip(*[_parse_data_desc(self.data_names, self.label_names, data_shape, label_shape)
for data_shape, label_shape in zip(data_shapes, label_shapes)])
if self._label_shapes.count(None) == len(self._label_shapes):
self._label_shapes = None
if shared_module is not None:
assert isinstance(shared_module, Module) and \
shared_module.binded and shared_module.params_initialized
shared_group = shared_module._exec_group
else:
shared_group = None
self._exec_group = DataParallelExecutorGroup(self._symbol, self._context,
self._work_load_list, self._data_shapes,
self._label_shapes, self._param_names,
for_training, inputs_need_grad,
shared_group, logger=self.logger,
fixed_param_names=self._fixed_param_names,
grad_req=grad_req,
state_names=self._state_names)
# self._total_exec_bytes = self._exec_group._total_exec_bytes
if shared_module is not None:
self.params_initialized = True
self._arg_params = shared_module._arg_params
self._aux_params = shared_module._aux_params
elif self.params_initialized:
# if the parameters are already initialized, we are re-binding
# so automatically copy the already initialized params
self._exec_group.set_params(self._arg_params, self._aux_params)
else:
assert self._arg_params is None and self._aux_params is None
param_arrays = [
nd.zeros(x[0].shape, dtype=x[0].dtype)
for x in self._exec_group.param_arrays
]
self._arg_params = {name:arr for name, arr in zip(self._param_names, param_arrays)}
aux_arrays = [
nd.zeros(x[0].shape, dtype=x[0].dtype)
for x in self._exec_group.aux_arrays
]
self._aux_params = {name:arr for name, arr in zip(self._aux_names, aux_arrays)}
if shared_module is not None and shared_module.optimizer_initialized:
self.borrow_optimizer(shared_module)
def bind(self, data_shapes, label_shapes=None, for_training=True,
inputs_need_grad=False, force_rebind=False, shared_module=None,
grad_req='write'):
"""Bind the symbols to construct executors. This is necessary before one
can perform computation with the module.
Parameters
----------
data_shapes : list of (str, tuple)
Typically is `data_iter.provide_data`.
label_shapes : list of (str, tuple)
Typically is `data_iter.provide_label`.
for_training : bool
Default is `True`. Whether the executors should be bind for training.
inputs_need_grad : bool
Default is `False`. Whether the gradients to the input data need to be computed.
Typically this is not needed. But this might be needed when implementing composition
of modules.
force_rebind : bool
Default is `False`. This function does nothing if the executors are already
binded. But with this `True`, the executors will be forced to rebind.
shared_module : Module
Default is `None`. This is used in bucketing. When not `None`, the shared module
essentially corresponds to a different bucket -- a module with different symbol
but with the same sets of parameters (e.g. unrolled RNNs with different lengths).
"""
# force rebinding is typically used when one want to switch from
# training to prediction phase.
if force_rebind:
self._reset_bind()
if self.binded:
self.logger.warning('Already binded, ignoring bind()')
return
self.for_training = for_training
self.inputs_need_grad = inputs_need_grad
self.binded = True
self._grad_req = grad_req
if not for_training:
assert not inputs_need_grad
else:
pass
# this is not True, as some module might not contains a loss function
# that consumes the labels
# assert label_shapes is not None
# self._data_shapes, self._label_shapes = _parse_data_desc(
# self.data_names, self.label_names, data_shapes, label_shapes)
self._data_shapes, self._label_shapes = zip(*[_parse_data_desc(self.data_names, self.label_names, data_shape, label_shape)
for data_shape, label_shape in zip(data_shapes, label_shapes)])
if self._label_shapes.count(None) == len(self._label_shapes):
self._label_shapes = None
if shared_module is not None:
assert isinstance(shared_module, Module) and \
shared_module.binded and shared_module.params_initialized
shared_group = shared_module._exec_group
else:
shared_group = None
self._exec_group = DataParallelExecutorGroup(self._symbol, self._context,
self._work_load_list, self._data_shapes,
self._label_shapes, self._param_names,
for_training, inputs_need_grad,
shared_group, logger=self.logger,
fixed_param_names=self._fixed_param_names,
grad_req=grad_req,
state_names=self._state_names)
# self._total_exec_bytes = self._exec_group._total_exec_bytes
if shared_module is not None:
self.params_initialized = True
self._arg_params = shared_module._arg_params
self._aux_params = shared_module._aux_params
elif self.params_initialized:
# if the parameters are already initialized, we are re-binding
# so automatically copy the already initialized params
self._exec_group.set_params(self._arg_params, self._aux_params)
else:
assert self._arg_params is None and self._aux_params is None
param_arrays = [
nd.zeros(x[0].shape, dtype=x[0].dtype)
for x in self._exec_group.param_arrays
]
self._arg_params = {name:arr for name, arr in zip(self._param_names, param_arrays)}
aux_arrays = [
nd.zeros(x[0].shape, dtype=x[0].dtype)
for x in self._exec_group.aux_arrays
]
self._aux_params = {name:arr for name, arr in zip(self._aux_names, aux_arrays)}
if shared_module is not None and shared_module.optimizer_initialized:
self.borrow_optimizer(shared_module)
def bind(modQ,
data_shapes,
label_shapes=None,
for_training=True,
inputs_need_grad=False,
force_rebind=False,
shared_module=None,
grad_req='write'):
if force_rebind:
modQ._reset_bind()
if modQ.binded:
modQ.logger.warning('Already binded, ignoring bind()')
return
modQ.for_training = for_training
modQ.inputs_need_grad = inputs_need_grad
modQ.binded = True
modQ._grad_req = grad_req
if not for_training:
assert not inputs_need_grad
else:
pass
# this is not True, as some module might not contains a loss function
# that consumes the labels
# assert label_shapes is not None
modQ._data_shapes, modQ._label_shapes = _parse_data_desc(
modQ.data_names, modQ.label_names, data_shapes, label_shapes)
if shared_module is not None:
assert isinstance(shared_module, Module) and \
shared_module.binded and shared_module.params_initialized
shared_group = shared_module._exec_group
else:
shared_group = None
modQ._exec_group = DataParallelExecutorGroup(
modQ._symbol,
modQ._context,
modQ._work_load_list,
modQ._data_shapes,
modQ._label_shapes,
modQ._param_names,
for_training,
inputs_need_grad,
shared_group,
logger=modQ.logger,
fixed_param_names=modQ._fixed_param_names,
grad_req=grad_req,
state_names=modQ._state_names)
modQ._total_exec_bytes = modQ._exec_group._total_exec_bytes
if shared_module is not None:
modQ.params_initialized = True
modQ._arg_params = shared_module._arg_params
modQ._aux_params = shared_module._aux_params
elif modQ.params_initialized:
# if the parameters are already initialized, we are re-binding
# so automatically copy the already initialized params
modQ._exec_group.set_params(modQ._arg_params, modQ._aux_params)
else:
assert modQ._arg_params is None and modQ._aux_params is None
param_arrays = [
nd.zeros(x[0].shape, dtype=x[0].dtype, ctx=x[0][0].context)
for x in modQ._exec_group.param_arrays
]
modQ._arg_params = {
name: arr
for name, arr in zip(modQ._param_names, param_arrays)
}
aux_arrays = [
nd.zeros(x[0].shape, dtype=x[0].dtype, ctx=x[0][0].context)
for x in modQ._exec_group.aux_arrays
]
modQ._aux_params = {
name: arr
for name, arr in zip(modQ._aux_names, aux_arrays)
}
if shared_module is not None and shared_module.optimizer_initialized:
modQ.borrow_optimizer(shared_module)
