def test_get_shallower_nodes():
input_1 = Input((10, ))
input_2 = Input((10, ))
dense_1 = Dense(3)
dense_2 = Dense(4)
dense_3 = Dense(5)
x = dense_1(input_1)
node_1_1 = get_inbound_nodes(dense_1)[0]
y = dense_1(input_2)
node_2_1 = get_inbound_nodes(dense_1)[1]
assert node_1_1 != node_2_1
output_1 = dense_2(x)
node_1_2 = get_inbound_nodes(dense_2)[0]
output_2 = dense_3(y)
node_2_2 = get_inbound_nodes(dense_3)[0]
assert get_shallower_nodes(node_1_1) == [node_1_2]
assert get_shallower_nodes(node_2_1) == [node_2_2]
output_3 = dense_2(y)
node_2_2_2 = get_inbound_nodes(dense_2)[1]
assert get_shallower_nodes(node_2_1) == [node_2_2, node_2_2_2]
def operate(self):
"""Perform all jobs assigned to the surgeon.
"""
# Operate on each node in self.nodes by order of decreasing depth.
sorted_nodes = sorted(
self.nodes,
reverse=True,
key=lambda x: utils.get_node_depth(self.model, x))
for node in sorted_nodes:
# Rebuild submodel up to this node
sub_output_nodes = utils.get_node_inbound_nodes(node)
outputs, output_masks = self._rebuild_graph(
self.model.inputs, sub_output_nodes)
# Perform surgery at this node
kwargs = self._kwargs_map[node]
self._mod_func_map[node](node, outputs, output_masks, **kwargs)
# Finish rebuilding model
output_nodes = []
for output in self.model.outputs:
layer, node_index, tensor_index = output._keras_history
output_nodes.append(get_inbound_nodes(layer)[node_index])
new_outputs, _ = self._rebuild_graph(self.model.inputs, output_nodes)
new_model = self._model_cls(self.model.inputs, new_outputs)
if self._copy:
return utils.clean_copy(new_model, self._custom_objects)
else:
return new_model
def operate(self):
"""Perform all jobs assigned to the surgeon.
"""
# Operate on each node in self.nodes by order of decreasing depth.
sorted_nodes = sorted(
self.nodes,
reverse=True,
key=lambda x: utils.get_node_depth(self.model, x))
for node in sorted_nodes:
# Rebuild submodel up to this node
sub_output_nodes = utils.get_node_inbound_nodes(node)
outputs, output_masks = self._rebuild_graph(
self.model.inputs, sub_output_nodes)
# Perform surgery at this node
kwargs = self._kwargs_map[node]
self._mod_func_map[node](node, outputs, output_masks, **kwargs)
# Finish rebuilding model
output_nodes = [
get_inbound_nodes(self.model.output_layers[i])[node_index] for i,
node_index in enumerate(self.model.output_layers_node_indices)
]
new_outputs, _ = self._rebuild_graph(self.model.inputs, output_nodes)
new_model = Model(self.model.inputs, new_outputs)
if self._copy:
return utils.clean_copy(new_model)
else:
return new_model
def _rebuild_rec(node):
"""Rebuild the graph up to `node` recursively.
Args:
node(Node): Node to rebuild up to.
Returns:
(tuple) containing :
The output tensor of the rebuilt `node`
The output mask of the rebuilt `node`
"""
# TODO: What happens if nodes have multiple output tensors?
# Does that ever happen?
#pdb.set_trace()
print(node, "rebuild node val")
layer = node.outbound_layer
#logging.debug('getting inputs for: {0}'.format(layer.name))
node_output = utils.single_element(node.output_tensors)
# First check for conditions to bottom out the recursion
# Check for replaced tensors before any other checks:
# these are created by the surgery methods.
if node_output in self._replace_tensors:
'''logging.debug('bottomed out at replaced output: {0}'.format(
node_output))'''
output, output_mask = self._replace_tensors[node_output]
#pdb.set_trace()
return output, output_mask
# Next check if the current node has already been rebuilt.
elif node in self._finished_nodes:
#logging.debug('reached finished node: {0}'.format(node))
#pdb.set_trace()
return self._finished_nodes[node]
# Next check if one of the graph_inputs has been reached.
elif node_output in graph_inputs:
#logging.debug('bottomed out at a model input')
output_mask = graph_input_masks[graph_inputs.index(
node_output)]
#pdb.set_trace()
return node_output, output_mask
# Otherwise recursively call this method on the inbound nodes.
else:
inbound_nodes = utils.get_inbound_nodes(node)
# Recursively rebuild the model up to `node`s inbound nodes to
# obtain its inputs and input masks
inputs, input_masks = zip(
*[_rebuild_rec(n) for n in inbound_nodes])
# Apply masks to the node's layer's weights and call the layer
# on the inputs
new_layer, output_mask = self._apply_delete_mask(
node, input_masks)
output = new_layer(utils.single_element(list(inputs)))
# Record that this node has been rebuild
self._finished_nodes[node] = (output, output_mask)
#logging.debug('layer complete: {0}'.format(layer.name))
#pdb.set_trace()
return output, output_mask
def test_rebuild_submodel(model_2):
output_nodes = []
for output in model_2.outputs:
layer, node_index, tensor_index = output._keras_history
output_nodes.append(get_inbound_nodes(layer)[node_index])
surgeon = Surgeon(model_2)
outputs, _ = surgeon._rebuild_graph(model_2.inputs, output_nodes)
new_model = Model(model_2.inputs, outputs)
assert compare_models(model_2, new_model)
def test_rebuild_submodel(model_2):
output_nodes = [
get_inbound_nodes(model_2.output_layers[i])[node_index]
for i, node_index in enumerate(model_2.output_layers_node_indices)
]
surgeon = Surgeon(model_2)
outputs, _ = surgeon._rebuild_graph(model_2.inputs, output_nodes)
new_model = Model(model_2.inputs, outputs)
assert compare_models(model_2, new_model)
def add_job(self,
job,
layer,
*,
channels=None,
new_layer=None,
node_indices=None):
"""Adds a job for the Surgeon to perform on the model.
Job options are:
'delete_layer': delete `layer` from the model
required keyword arguments: None
'insert_layer': insert `new_layer` before `layer`
required keyword arguments: `new_layer`
'replace_layer': replace `layer` with `new_layer`
required keyword arguments: `new_layer`
'delete_channels': delete `channels` from `layer`
required keyword arguments: `channels`
Jobs can be added in any order. They will be performed in order of
decreasing network depth.
A maximum of one job can be performed per node.
Args:
job(string): job identifier. One of `Surgeon.valid_jobs`.
layer(Layer): A layer from `model` to be modified.
channels(list[int]): A list of channels used for the job.
Used in `delete_channels`.
new_layer(Layer): A new layer used for the job. Used in
`insert_layer` and `replace_layer`.
node_indices(list[int]): (optional) A list of node indices used to
selectively apply the job to a subset of
the layer's nodes. Nodes are selected with:
node[i] = layer.inbound_nodes[node_indices[i]]
"""
# If the model has been copied, identify `layer` in the copied model.
if self._copy:
layer = self.model.get_layer(layer.name)
# Check that layer is in the model
if layer not in self.model.layers:
raise ValueError("layer is not a valid Layer in model.")
layer_node_indices = utils.find_nodes_in_model(self.model, layer)
# If no nodes are specified, all of the layer's inbound nodes which are
# in model are selected.
if not node_indices:
node_indices = layer_node_indices
# Check for duplicate node indices
elif len(node_indices) != len(set(node_indices)):
raise ValueError("`node_indices` contains duplicate values.")
# Check that all of the selected nodes are in the layer
elif not set(node_indices).issubset(layer_node_indices):
raise ValueError("One or more nodes specified by `layer` and "
"`node_indices` are not in `model`.")
# Select the modification function and any keyword arguments.
kwargs = {}
if job == "delete_channels":
# If not all inbound_nodes are selected, the new layer is renamed
# to avoid duplicate layer names.
if set(node_indices) != set(layer_node_indices):
kwargs["layer_name"] = layer.name + "_" + job
kwargs["channels"] = channels
mod_func = self._delete_channels
elif job == "delete_layer":
mod_func = self._delete_layer
elif job == "insert_layer":
kwargs["new_layer"] = new_layer
mod_func = self._insert_layer
elif job == "replace_layer":
kwargs["new_layer"] = new_layer
mod_func = self._replace_layer
else:
raise ValueError(
job + " is not a recognised job. Valid jobs "
"are:\n-",
"\n- ".join(self.valid_jobs),
)
# Get nodes to be operated on for this job
job_nodes = []
for node_index in node_indices:
job_nodes.append(get_inbound_nodes(layer)[node_index])
# Check that the nodes do not already have jobs assigned to them.
if set(job_nodes).intersection(self.nodes):
raise ValueError("Cannot apply several jobs to the same node.")
# Add the modification function and keyword arguments to the
# self._mod_func_map and _kwargs_map dictionaries for later retrieval.
for node in job_nodes:
self._mod_func_map[node] = mod_func
self._kwargs_map[node] = kwargs
self.nodes.extend(job_nodes)
def _apply_delete_mask(self, node, inbound_masks):
"""Apply the inbound delete mask and return the outbound delete mask
When specific channels in a layer or layer instance are deleted, the
mask propagates information about which channels are affected to
downstream layers.
If the layer contains weights, those which were previously connected
to the deleted channels are deleted and outbound masks are set to None
since further downstream layers aren't affected.
If the layer does not contain weights, its output mask is calculated to
reflect any transformations performed by the layer to ensure that
information about the deleted channels is propagated downstream.
Arguments:
node(Node): The node where the delete mask is applied.
inbound_masks: Mask(s) from inbound node(s).
Returns:
new_layer: Pass through `layer` if it has no weights, otherwise a
new `Layer` object with weights corresponding to the
inbound mask deleted.
outbound_mask: Mask corresponding to `new_layer`.
"""
# if delete_mask is None or all values are True, it does not affect
# this layer or any layers above/downstream from it
layer = node.outbound_layer
if all(mask is None for mask in inbound_masks):
new_layer = layer
outbound_mask = None
return new_layer, outbound_mask
# If one or more of the masks are None, replace them with ones.
if any(mask is None for mask in inbound_masks):
inbound_masks = [
np.ones(shape[1:], dtype=bool)
if inbound_masks[i] is None else inbound_masks[i]
for i, shape in enumerate(node.input_shapes)
]
# If the layer is shared and has already been affected by this
# operation, use the cached new layer.
if (len(get_inbound_nodes(layer)) > 1
and layer in self._replace_layers_map.keys()):
return self._replace_layers_map[layer]
output_shape = utils.single_element(node.output_shapes)
input_shape = utils.single_element(node.input_shapes)
data_format = getattr(layer, "data_format", "channels_last")
inbound_masks = utils.single_element(inbound_masks)
# otherwise, delete_mask.shape should be: layer.input_shape[1:]
layer_class = layer.__class__.__name__
if layer_class == "InputLayer":
raise RuntimeError("This should never get here!")
elif layer_class == "Dense":
if np.all(inbound_masks):
new_layer = layer
else:
weights = layer.get_weights()
weights[0] = weights[0][np.where(inbound_masks)[0], :]
config = layer.get_config()
config["weights"] = weights
new_layer = type(layer).from_config(config)
outbound_mask = None
elif layer_class == "Flatten":
outbound_mask = np.reshape(inbound_masks, [
-1,
])
new_layer = layer
elif layer_class in ("Conv1D", "Conv2D", "Conv3D"):
if np.all(inbound_masks):
new_layer = layer
else:
if data_format == "channels_first":
inbound_masks = np.swapaxes(inbound_masks, 0, -1)
# Conv layer: trim down inbound_masks to filter shape
k_size = layer.kernel_size
index = [slice(None, 1, None) for _ in k_size]
inbound_masks = inbound_masks[tuple(index + [slice(None)])]
weights = layer.get_weights()
# Delete unused weights to obtain new_weights
# Each deleted channel was connected to all of the channels
# in layer; therefore, the mask must be repeated for each
# channel.
# `delete_mask`'s size: size(weights[0])
delete_mask = np.tile(
inbound_masks[..., np.newaxis],
list(k_size) + [1, weights[0].shape[-1]],
)
new_shape = list(weights[0].shape)
new_shape[-2] = -1 # Weights always have channels_last
weights[0] = np.reshape(weights[0][delete_mask], new_shape)
# Instantiate new layer with new_weights
config = layer.get_config()
config["weights"] = weights
new_layer = type(layer).from_config(config)
outbound_mask = None
elif layer_class in (
"Cropping1D",
"Cropping2D",
"Cropping3D",
"MaxPooling1D",
"MaxPooling2D",
"MaxPooling3D",
"AveragePooling1D",
"AveragePooling2D",
"AveragePooling3D",
):
index = [slice(None, x, None) for x in output_shape[1:]]
if data_format == "channels_first":
index[0] = slice(None)
elif data_format == "channels_last":
index[-1] = slice(None)
else:
raise ValueError("Invalid data format")
outbound_mask = inbound_masks[tuple(index)]
new_layer = layer
elif layer_class in (
"UpSampling1D",
"UpSampling2D",
"UpSampling3D",
"ZeroPadding1D",
"ZeroPadding2D",
"ZeroPadding3D",
):
# Get slice of mask with all singleton dimensions except
# channels dimension
index = [slice(1)] * (len(input_shape) - 1)
tile_shape = list(output_shape[1:])
if data_format == "channels_first":
index[0] = slice(None)
tile_shape[0] = 1
elif data_format == "channels_last":
index[-1] = slice(None)
tile_shape[-1] = 1
else:
raise ValueError("Invalid data format")
channels_vector = inbound_masks[tuple(index)]
# Tile this slice to create the outbound mask
outbound_mask = np.tile(channels_vector, tile_shape)
new_layer = layer
elif layer_class in (
"GlobalMaxPooling1D",
"GlobalMaxPooling2D",
"GlobalAveragePooling1D",
"GlobalAveragePooling2D",
):
# Get slice of mask with all singleton dimensions except
# channels dimension
index = [0] * (len(input_shape) - 1)
if data_format == "channels_first":
index[0] = slice(None)
elif data_format == "channels_last":
index[-1] = slice(None)
else:
raise ValueError("Invalid data format")
channels_vector = inbound_masks[tuple(index)]
# Tile this slice to create the outbound mask
outbound_mask = channels_vector
new_layer = layer
elif layer_class in (
"Dropout",
"Activation",
"SpatialDropout1D",
"SpatialDropout2D",
"SpatialDropout3D",
"ActivityRegularization",
"Masking",
"LeakyReLU",
"ELU",
"ThresholdedReLU",
"GaussianNoise",
"GaussianDropout",
"AlphaDropout",
"Rename",
) or (layer_class == "TensorFlowOpLayer"
and layer.node_def.op == "ResizeBilinear"):
# Pass-through layers
outbound_mask = inbound_masks
new_layer = layer
elif layer_class == "Reshape":
outbound_mask = np.reshape(inbound_masks, layer.target_shape)
new_layer = layer
elif layer_class == "Permute":
outbound_mask = np.transpose(inbound_masks,
[x - 1 for x in layer.dims])
new_layer = layer
elif layer_class == "RepeatVector":
outbound_mask = np.repeat(np.expand_dims(inbound_masks, 0),
layer.n,
axis=0)
new_layer = layer
elif layer_class == "Embedding":
# Embedding will always be the first layer so it doesn't need
# to consider the inbound_delete_mask
if inbound_masks is not None:
raise ValueError("Channels cannot be deleted bedore Embedding "
"layers because they change the number of "
"channels.")
outbound_mask = None
new_layer = layer
elif layer_class in ("Add", "Multiply", "Average", "Maximum"):
# The inputs must be the same size
if not utils.all_equal(inbound_masks):
ValueError(
"{0} layers must have the same size inputs. All "
"inbound nodes must have the same channels deleted".format(
layer_class))
outbound_mask = inbound_masks[1]
new_layer = layer
elif layer_class == "Concatenate":
axis = layer.axis
if layer.axis < 0:
axis = axis % len(layer.input_shape[0])
# Below: axis=axis-1 because the mask excludes the batch dimension
outbound_mask = np.concatenate(inbound_masks, axis=axis - 1)
new_layer = layer
elif layer_class in ("SimpleRNN", "GRU", "LSTM"):
if np.all(inbound_masks):
new_layer = layer
else:
weights = layer.get_weights()
weights[0] = weights[0][np.where(inbound_masks[0, :])[0], :]
config = layer.get_config()
config["weights"] = weights
new_layer = type(layer).from_config(config)
outbound_mask = None
elif layer_class == "BatchNormalization":
outbound_mask = inbound_masks
# Get slice of mask with all singleton dimensions except
# channels dimension
index = [0] * (len(input_shape))
index[layer.axis] = slice(None)
index = index[1:]
# TODO: Maybe use channel indices everywhere instead of masks?
channel_indices = np.where(inbound_masks[tuple(index)] == False)[0]
weights = [
np.delete(w, channel_indices, axis=-1)
for w in layer.get_weights()
]
new_layer = BatchNormalization.from_config(layer.get_config())
new_input_shape = list(input_shape)
new_input_shape[new_layer.axis] -= len(channel_indices)
new_layer.build(new_input_shape)
new_layer.set_weights(weights)
else:
# Not implemented:
# - Lambda
# - SeparableConv2D
# - Conv2DTranspose
# - LocallyConnected1D
# - LocallyConnected2D
# - TimeDistributed
# - Bidirectional
# - Dot
# - PReLU
# Warning/error needed for Reshape if channels axis is split
raise ValueError('"{0}" layers are currently '
"unsupported.".format(layer_class))
if len(get_inbound_nodes(layer)) > 1 and new_layer != layer:
self._replace_layers_map[layer] = (new_layer, outbound_mask)
return new_layer, outbound_mask
