def _delete_channels(self,
node,
inputs,
input_masks,
channels=None,
layer_name=None):
"""Delete selected channels of node.outbound_layer. Add it to the graph.
"""
old_layer = node.outbound_layer
old_layer_output = utils.single_element(node.output_tensors)
# Create a mask to propagate the deleted channels to downstream layers
new_delete_mask = self._make_delete_mask(old_layer, channels)
if len(set(channels)) == getattr(old_layer,
utils.get_channels_attr(old_layer)):
self._replace_tensors[old_layer_output] = (None, new_delete_mask)
return None
# If this layer has already been operated on, use the cached copy of
# the new layer. Otherwise, apply the inbound delete mask and
# delete channels to obtain the new layer
if old_layer in self._new_layers_map.keys():
new_layer = self._new_layers_map[old_layer]
else:
temp_layer, new_mask = self._apply_delete_mask(node, input_masks)
# This call is needed to initialise input_shape and output_shape
temp_layer(utils.single_element(inputs))
new_layer = self._delete_channel_weights(temp_layer, channels)
if layer_name:
new_layer.name = layer_name
self._new_layers_map[old_layer] = new_layer
new_output = new_layer(utils.single_element(inputs))
# Replace the original layer's output with the modified layer's output
self._replace_tensors[old_layer_output] = (new_output, new_delete_mask)
def layer_test_helper_flatten_1d(layer, channel_index):
# This should test that the output is the correct shape so it should pass
# into a Dense layer rather than a Conv layer.
# The weighted layer is the previous layer,
# Create model
main_input = Input(shape=list(random.randint(10, 20, size=2)))
x = Conv1D(3, 3)(main_input)
x = layer(x)
x = Flatten()(x)
main_output = Dense(5)(x)
model = Model(inputs=main_input, outputs=main_output)
# Delete channels
del_layer_index = 1
next_layer_index = 4
del_layer = model.layers[del_layer_index]
surgeon = Surgeon(model)
surgeon.add_job("delete_channels", del_layer, channels=channel_index)
new_model = surgeon.operate()
new_w = new_model.layers[next_layer_index].get_weights()
# Calculate next layer's correct weights
flat_sz = np.prod(layer.get_output_shape_at(0)[1:])
channel_count = getattr(del_layer, utils.get_channels_attr(del_layer))
channel_index = [i % channel_count for i in channel_index]
delete_indices = [
x + i for i in range(0, flat_sz, channel_count) for x in channel_index
]
correct_w = model.layers[next_layer_index].get_weights()
correct_w[0] = np.delete(correct_w[0], delete_indices, axis=0)
assert weights_equal(correct_w, new_w)
def layer_test_helper_2d_global(layer, channel_index, data_format):
# This should test that the output is the correct shape so it should pass
# into a Dense layer rather than a Conv layer.
# The weighted layer is the previous layer,
# Create model
main_input = Input(shape=list(random.randint(10, 20, size=3)))
x = Conv2D(3, [3, 3], data_format=data_format)(main_input)
x = layer(x)
main_output = Dense(5)(x)
model = Model(inputs=main_input, outputs=main_output)
# Delete channels
del_layer_index = 1
next_layer_index = 3
del_layer = model.layers[del_layer_index]
new_model = operations.delete_channels(model, del_layer, channel_index)
new_w = new_model.layers[next_layer_index].get_weights()
# Calculate next layer's correct weights
channel_count = getattr(del_layer, utils.get_channels_attr(del_layer))
channel_index = [i % channel_count for i in channel_index]
correct_w = model.layers[next_layer_index].get_weights()
correct_w[0] = np.delete(correct_w[0], channel_index, axis=0)
assert weights_equal(correct_w, new_w)
def test_delete_channels_flatten(channel_index, data_format):
# Create model
main_input = Input(shape=list(random.randint(4, 10, size=3)))
x = Conv2D(3, [3, 3], data_format=data_format)(main_input)
x = Flatten()(x)
main_output = Dense(5)(x)
model = Model(inputs=main_input, outputs=main_output)
# Delete channels
layer_index = 1
next_layer_index = 3
layer = model.layers[layer_index]
new_model = operations.delete_channels(model, layer, channel_index)
new_w = new_model.layers[next_layer_index].get_weights()
# Calculate next layer's correct weights
flat_sz = np.prod(layer.output_shape[1:])
channel_count = getattr(layer, utils.get_channels_attr(layer))
channel_index = [i % channel_count for i in channel_index]
if data_format == 'channels_first':
delete_indices = [x*flat_sz//channel_count + i for x in channel_index
for i in range(0, flat_sz//channel_count, )]
elif data_format == 'channels_last':
delete_indices = [x + i for i in range(0, flat_sz, channel_count)
for x in channel_index]
else:
raise ValueError
correct_w = model.layers[next_layer_index].get_weights()
correct_w[0] = np.delete(correct_w[0], delete_indices, axis=0)
assert weights_equal(correct_w, new_w)
def test_delete_channels_merge_concatenate(channel_index, data_format):
# This should test that the output is the correct shape so it should pass
# into a Dense layer rather than a Conv layer.
# The weighted layer is the previous layer,
# Create model
if data_format == "channels_first":
axis = 1
elif data_format == "channels_last":
axis = -1
else:
raise ValueError
input_shape = list(random.randint(10, 20, size=3))
input_1 = Input(shape=input_shape)
input_2 = Input(shape=input_shape)
x = Conv2D(3, [3, 3], data_format=data_format, name="conv_1")(input_1)
y = Conv2D(3, [3, 3], data_format=data_format, name="conv_2")(input_2)
x = Concatenate(axis=axis, name="cat_1")([x, y])
x = Flatten()(x)
main_output = Dense(5, name="dense_1")(x)
model = Model(inputs=[input_1, input_2], outputs=main_output)
old_w = model.get_layer("dense_1").get_weights()
# Delete channels
layer = model.get_layer("cat_1")
del_layer = model.get_layer("conv_1")
surgeon = Surgeon(model, copy=True)
surgeon.add_job("delete_channels", del_layer, channels=channel_index)
new_model = surgeon.operate()
new_w = new_model.get_layer("dense_1").get_weights()
# Calculate next layer's correct weights
flat_sz = np.prod(layer.get_output_shape_at(0)[1:])
channel_count = getattr(del_layer, utils.get_channels_attr(del_layer))
channel_index = [i % channel_count for i in channel_index]
if data_format == "channels_first":
delete_indices = [
x * flat_sz // 2 // channel_count + i for x in channel_index
for i in range(
0,
flat_sz // 2 // channel_count,
)
]
elif data_format == "channels_last":
delete_indices = [
x + i for i in range(0, flat_sz, channel_count * 2)
for x in channel_index
]
else:
raise ValueError
correct_w = model.get_layer("dense_1").get_weights()
correct_w[0] = np.delete(correct_w[0], delete_indices, axis=0)
assert weights_equal(correct_w, new_w)
def recursive_test_helper(layer, channel_index):
main_input = Input(shape=[32, 10])
x = layer(main_input)
x = GRU(4, return_sequences=False)(x)
main_output = Dense(5)(x)
model = Model(inputs=main_input, outputs=main_output)
# Delete channels
del_layer_index = 1
next_layer_index = 2
del_layer = model.layers[del_layer_index]
new_model = operations.delete_channels(model, del_layer, channel_index)
new_w = new_model.layers[next_layer_index].get_weights()
# Calculate next layer's correct weights
channel_count = getattr(del_layer, utils.get_channels_attr(del_layer))
channel_index = [i % channel_count for i in channel_index]
correct_w = model.layers[next_layer_index].get_weights()
correct_w[0] = np.delete(correct_w[0], channel_index, axis=0)
assert weights_equal(correct_w, new_w)
def _delete_channel_weights(self, layer, channel_indices):
"""Delete channels from layer and remove the corresponding weights.
Arguments:
layer: A layer whose channels are to be deleted
channel_indices: The indices of the channels to be deleted.
Returns:
A new layer with the channels and corresponding weights deleted.
"""
layer_config = layer.get_config()
channels_attr = utils.get_channels_attr(layer)
channel_count = layer_config[channels_attr]
# Check inputs
if any([i + 1 > channel_count for i in channel_indices]):
raise ValueError(
'Channels_index value(s) out of range. '
'This layer only has {0} channels.'.format(channel_count))
print('Deleting {0}/{1} channels from layer: {2}'.format(
len(channel_indices), channel_count, layer.name))
# numpy.delete ignores negative indices in lists: wrap indices
channel_indices = [i % channel_count for i in channel_indices]
# Reduce layer channel count in config.
layer_config[channels_attr] -= len(channel_indices)
# Delete weights corresponding to deleted channels from config.
# Except for recurrent layers, the weights' channels dimension is last.
# Each recurrent layer type has a different internal weights layout.
if layer.__class__.__name__ == 'SimpleRNN':
weights = [
np.delete(w, channel_indices, axis=-1)
for w in layer.get_weights()
]
weights[1] = np.delete(weights[1], channel_indices, axis=0)
elif layer.__class__.__name__ == 'GRU':
# Repeat the channel indices for all internal GRU weights.
channel_indices_gru = [
layer.units * m + i for m in range(3) for i in channel_indices
]
weights = [
np.delete(w, channel_indices_gru, axis=-1)
for w in layer.get_weights()
]
weights[1] = np.delete(weights[1], channel_indices, axis=0)
elif layer.__class__.__name__ == 'LSTM':
# Repeat the channel indices for all interal LSTM weights.
channel_indices_lstm = [
layer.units * m + i for m in range(4) for i in channel_indices
]
weights = [
np.delete(w, channel_indices_lstm, axis=-1)
for w in layer.get_weights()
]
weights[1] = np.delete(weights[1], channel_indices, axis=0)
else:
weights = [
np.delete(w, channel_indices, axis=-1)
for w in layer.get_weights()
]
layer_config['weights'] = weights
# Create new layer from the modified configuration and return it.
return type(layer).from_config(layer_config)
