def call(self, inputs, weights, training: tf.constant):
"""
Apply rb sparsity mask to given weights.
:param inputs: Target weights to sparsify.
:param weights: Operation weights contains
`mask` and param `trainable`.
:param training: True if operation called in training mode
else False
"""
true_fn = lambda: apply_mask(inputs, self._calc_rb_binary_mask(weights))
false_fn = lambda: apply_mask(inputs, binary_mask(weights['mask']))
return smart_cond(training,
true_fn=lambda: smart_cond(weights['trainable'],
true_fn=true_fn, false_fn=false_fn),
false_fn=false_fn)
def test_loss_value(self, mask_value, ref_loss, frozen):
sw, layer, op_name = get_RBSParsityWeigth_and_layer(frozen=frozen)
op_weights = layer.ops_weights[op_name]
mask = op_weights['mask']
if mask_value is not None:
mask.assign(tf.fill(mask.shape, mask_value))
assert sw.loss(op_weights) == ref_loss
w = tf.ones(1)
assert apply_mask(w, st_binary_mask(mask)) == ref_loss
sw.freeze(op_weights)
assert sw(w, op_weights, tf.constant(True, tf.bool)) == ref_loss
def _collect_weights_descriptions(self) -> List[WeightDescription]:
weights_descriptions = []
excluded_names = []
# Collect description for sparse weights i.e. weights for which
# sparsity algorithm was applied.
for wrapped_layer, weight_attr, op in get_nncf_operations(
self._model, self._operation_names):
weight = wrapped_layer.layer_weights[weight_attr]
operation_weights = wrapped_layer.get_operation_weights(op.name)
binary_mask = op.get_binary_mask(operation_weights)
sparse_weight = apply_mask(weight, binary_mask)
weights_descriptions.append(
WeightDescription(
weight.name,
_get_standardized_weight_shape(weight.shape.as_list()),
tf.math.count_nonzero(sparse_weight).numpy().item(),
is_sparse=True))
# Exclude this name because it has been processed.
excluded_names.append(weight.name)
# Exclude these names because they were added to the model
# by the sparsity algorithm.
excluded_names.extend([w.name for w in operation_weights.values()])
if isinstance(op, BinaryMaskWithWeightsBackup):
excluded_names.append(op.bkup_var.name)
# Collect descriptions for rest weights.
unique_weights = {
id(w): w
for w in self._model.weights if w.name not in excluded_names
}.values()
for weight in unique_weights:
weights_descriptions.append(
WeightDescription(weight.name,
_get_standardized_weight_shape(
weight.shape.as_list()),
tf.math.count_nonzero(weight).numpy().item(),
is_sparse=False))
return weights_descriptions
def call(self, inputs, weights, _):
self.bkup_var.assign(
tf.where(weights['mask'] > 0.5, inputs, self.bkup_var))
return apply_mask(self.bkup_var, weights['mask'])
def call(self, inputs, weights, _):
return apply_mask(inputs, weights['mask'])