def _get_mask_assign_ops(self):
# Make sure the assignment ops have not already been added to the list
if self._assign_ops:
raise ValueError(
'Assign op list not empty. _get_mask_assign_ops() called twice?')
masks = get_masks()
weights = get_weights()
thresholds = get_thresholds()
if len(masks) != len(thresholds):
raise ValueError(
'Number of masks %s and number of thresholds %s mismatch' %
(len(masks), len(thresholds)))
for index, mask in enumerate(masks):
threshold = thresholds[index]
weight = weights[index]
is_partitioned = isinstance(weight, variables.PartitionedVariable)
if is_partitioned:
weight = weight.as_tensor()
new_threshold, new_mask = self._maybe_update_block_mask(weight, threshold)
self._assign_ops.append(
pruning_utils.variable_assign(threshold, new_threshold))
self._assign_ops.append(
pruning_utils.partitioned_variable_assign(mask, new_mask)
if is_partitioned else pruning_utils.variable_assign(mask, new_mask))
def _get_mask_assign_ops(self):
# Make sure the assignment ops have not already been added to the list
if self._assign_ops:
raise ValueError(
'Assign op list not empty. _get_mask_assign_ops() called twice?')
masks = get_masks()
weights = get_weights()
thresholds = get_thresholds()
if len(masks) != len(thresholds):
raise ValueError(
'Number of masks %s and number of thresholds %s mismatch' %
(len(masks), len(thresholds)))
for index, mask in enumerate(masks):
threshold = thresholds[index]
weight = weights[index]
is_partitioned = isinstance(weight, variables.PartitionedVariable)
if is_partitioned:
weight = weight.as_tensor()
new_threshold, new_mask = self._maybe_update_block_mask(weight, threshold)
self._assign_ops.append(
pruning_utils.variable_assign(threshold, new_threshold))
self._assign_ops.append(
pruning_utils.partitioned_variable_assign(mask, new_mask)
if is_partitioned else pruning_utils.variable_assign(mask, new_mask))