How to replace (or insert) intermediate layer in Keras model? - tensorflow

Question

I have a trained Keras model and I would like:
1) to replace Con2D layer with the same but without bias.
2) to add BatchNormalization layer before first Activation
How can I do this?
def keras_simple_model():
from keras.models import Model
from keras.layers import Input, Dense, GlobalAveragePooling2D
from keras.layers import Conv2D, MaxPooling2D, Activation
inputs1 = Input((28, 28, 1))
x = Conv2D(4, (3, 3), activation=None, padding='same', name='conv1')(inputs1)
x = Activation('relu')(x)
x = Conv2D(4, (3, 3), activation=None, padding='same', name='conv2')(x)
x = Activation('relu')(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='pool1')(x)
x = Conv2D(8, (3, 3), activation=None, padding='same', name='conv3')(x)
x = Activation('relu')(x)
x = Conv2D(8, (3, 3), activation=None, padding='same', name='conv4')(x)
x = Activation('relu')(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='pool2')(x)
x = GlobalAveragePooling2D()(x)
x = Dense(10, activation=None)(x)
x = Activation('softmax')(x)
model = Model(inputs=inputs1, outputs=x)
return model
if __name__ == '__main__':
model = keras_simple_model()
print(model.summary())

Answer 1

The following function allows you to insert a new layer before, after or to replace each layer in the original model whose name matches a regular expression, including non-sequential models such as DenseNet or ResNet.
import re
from keras.models import Model
def insert_layer_nonseq(model, layer_regex, insert_layer_factory,
insert_layer_name=None, position='after'):
# Auxiliary dictionary to describe the network graph
network_dict = {'input_layers_of': {}, 'new_output_tensor_of': {}}
# Set the input layers of each layer
for layer in model.layers:
for node in layer._outbound_nodes:
layer_name = node.outbound_layer.name
if layer_name not in network_dict['input_layers_of']:
network_dict['input_layers_of'].update(
{layer_name: [layer.name]})
else:
network_dict['input_layers_of'][layer_name].append(layer.name)
# Set the output tensor of the input layer
network_dict['new_output_tensor_of'].update(
{model.layers[0].name: model.input})
# Iterate over all layers after the input
model_outputs = []
for layer in model.layers[1:]:
# Determine input tensors
layer_input = [network_dict['new_output_tensor_of'][layer_aux]
for layer_aux in network_dict['input_layers_of'][layer.name]]
if len(layer_input) == 1:
layer_input = layer_input[0]
# Insert layer if name matches the regular expression
if re.match(layer_regex, layer.name):
if position == 'replace':
x = layer_input
elif position == 'after':
x = layer(layer_input)
elif position == 'before':
pass
else:
raise ValueError('position must be: before, after or replace')
new_layer = insert_layer_factory()
if insert_layer_name:
new_layer.name = insert_layer_name
else:
new_layer.name = '{}_{}'.format(layer.name,
new_layer.name)
x = new_layer(x)
print('New layer: {} Old layer: {} Type: {}'.format(new_layer.name,
layer.name, position))
if position == 'before':
x = layer(x)
else:
x = layer(layer_input)
# Set new output tensor (the original one, or the one of the inserted
# layer)
network_dict['new_output_tensor_of'].update({layer.name: x})
# Save tensor in output list if it is output in initial model
if layer_name in model.output_names:
model_outputs.append(x)
return Model(inputs=model.inputs, outputs=model_outputs)
The difference with respect to the simpler case of a purely sequential model is that before iterating over the layers to find the key layer, you first parse the graph and store the input layers of each layer in an auxiliary dictionary. Then, as you iterate over the layers, you also store the new output tensor of each layer, which is used to determine the input layers of each layer, when building the new model.
A use case would be the following, where a Dropout layer is inserted after each activation layer of ResNet50:
from keras.applications.resnet50 import ResNet50
from keras.models import load_model
model = ResNet50()
def dropout_layer_factory():
return Dropout(rate=0.2, name='dropout')
model = insert_layer_nonseq(model, '.*activation.*', dropout_layer_factory)
# Fix possible problems with new model
model.save('temp.h5')
model = load_model('temp.h5')
model.summary()

Answer 2

You can use the following functions:
from keras.models import Model
def replace_intermediate_layer_in_keras(model, layer_id, new_layer):
layers = [l for l in model.layers]
x = layers[0].output
for i in range(1, len(layers)):
if i == layer_id:
x = new_layer(x)
else:
x = layers[i](x)
new_model = Model(input=layers[0].input, output=x)
return new_model
def insert_intermediate_layer_in_keras(model, layer_id, new_layer):
layers = [l for l in model.layers]
x = layers[0].output
for i in range(1, len(layers)):
if i == layer_id:
x = new_layer(x)
x = layers[i](x)
new_model = Model(input=layers[0].input, output=x)
return new_model
Example:
from keras.layers import Conv2D, BatchNormalization
model = keras_simple_model()
print(model.summary())
model = replace_intermediate_layer_in_keras(
model, 3,
Conv2D(
4, (3, 3),
activation=None,
padding='same',
name='conv2_repl',
use_bias=False
)
)
print(model.summary())
model = insert_intermediate_layer_in_keras(
model, 4, BatchNormalization()
)
print(model.summary())
There are some limitation on replacements due to layer shapes etc.

Answer 3

This was how i did it:
import keras
from keras.models import Model
from tqdm import tqdm
from keras import backend as K
def make_list(X):
if isinstance(X, list):
return X
return [X]
def list_no_list(X):
if len(X) == 1:
return X[0]
return X
def replace_layer(model, replace_layer_subname, replacement_fn,
**kwargs):
"""
args:
model :: keras.models.Model instance
replace_layer_subname :: str -- if str in layer name, replace it
replacement_fn :: fn to call to replace all instances
> fn output must produce shape as the replaced layers input
returns:
new model with replaced layers
quick examples:
want to just remove all layers with 'batch_norm' in the name:
> new_model = replace_layer(model, 'batch_norm', lambda **kwargs : (lambda u:u))
want to replace all Conv1D(N, m, padding='same') with an LSTM (lets say all have 'conv1d' in name)
> new_model = replace_layer(model, 'conv1d', lambda layer, **kwargs: LSTM(units=layer.filters, return_sequences=True)
"""
model_inputs = []
model_outputs = []
tsr_dict = {}
model_output_names = [out.name for out in make_list(model.output)]
for i, layer in enumerate(model.layers):
### Loop if layer is used multiple times
for j in range(len(layer._inbound_nodes)):
### check layer inp/outp
inpt_names = [inp.name for inp in make_list(layer.get_input_at(j))]
outp_names = [out.name for out in make_list(layer.get_output_at(j))]
### setup model inputs
if 'input' in layer.name:
for inpt_tsr in make_list(layer.get_output_at(j)):
model_inputs.append(inpt_tsr)
tsr_dict[inpt_tsr.name] = inpt_tsr
continue
### setup layer inputs
inpt = list_no_list([tsr_dict[name] for name in inpt_names])
### remake layer
if replace_layer_subname in layer.name:
print('replacing '+layer.name)
x = replacement_fn(old_layer=layer, **kwargs)(inpt)
else:
x = layer(inpt)
### reinstantialize outputs into dict
for name, out_tsr in zip(outp_names, make_list(x)):
### check if is an output
if name in model_output_names:
model_outputs.append(out_tsr)
tsr_dict[name] = out_tsr
return Model(model_inputs, model_outputs)
I have a custom layer (taken from someone online) called BatchNormalizationFreeze, so an example of usage is this:
new_model = model_replacement(model, 'batch_normal', lambda **kwargs : BatchNormalizationFreeze()(x))
If youre gonna do multiple layers just replace the replacement function with a psuedo model that does them all at once

Answer 4

Unfortunately replacing a layer is no small feat for models that do not follow the sequential pattern. For sequential patterns it is OK to just x = layer(x) and replace with new_layer when you see fit as in the previous answer.
However, for models that do not have a classic sequential pattern (say you have a simple "concatenation" of two columns) you have to actually "parse" the graph and use your "new_layer" (or layers) in the right places. Hope this is not too discouraging and happy graph parsing and reconstructing :)