def cloneLayerFromLayer(pLayer):
if isinstance(pLayer, Convolution1D):
return Convolution1D.from_config(pLayer.get_config())
elif isinstance(pLayer, Convolution2D):
return Convolution2D.from_config(pLayer.get_config())
elif isinstance(pLayer, Convolution3D):
return Convolution3D.from_config(pLayer.get_config())
# Max-Pooling:
elif isinstance(pLayer, MaxPooling1D):
return MaxPooling2D.from_config(pLayer.get_config())
elif isinstance(pLayer, MaxPooling2D):
return MaxPooling2D.from_config(pLayer.get_config())
elif isinstance(pLayer, MaxPooling3D):
return MaxPooling3D.from_config(pLayer.get_config())
# Average-Pooling
elif isinstance(pLayer, AveragePooling1D):
return AveragePooling1D.from_config(pLayer.get_config())
elif isinstance(pLayer, AveragePooling2D):
return AveragePooling2D.from_config(pLayer.get_config())
elif isinstance(pLayer, AveragePooling3D):
return AveragePooling3D.from_config(pLayer.get_config())
#
elif isinstance(pLayer, Flatten):
return Flatten.from_config(pLayer.get_config())
elif isinstance(pLayer, Merge):
return Merge.from_config(pLayer.get_config())
elif isinstance(pLayer, Activation):
return Activation.from_config(pLayer.get_config())
elif isinstance(pLayer, Dropout):
return Dropout.from_config(pLayer.get_config())
#
elif isinstance(pLayer, Dense):
return Dense.from_config(pLayer.get_config())
return None
def generate_conv_model(model):
model = model.get_layer(index=1)
inp = (model.inputs[0].shape.dims[1].value,
model.inputs[0].shape.dims[2].value,
model.inputs[0].shape.dims[3].value)
H = Input(inp)
inp = H
for layer_idx in range(1, len(model.layers)):
layer = model.get_layer(index=layer_idx)
config = layer.get_config()
if isinstance(layer, MaxPooling2D):
H = MaxPooling2D.from_config(config)(H)
if isinstance(layer, Dropout):
H = Dropout.from_config(config)(H)
if isinstance(layer, Activation):
H = Activation.from_config(config)(H)
if isinstance(layer, BatchNormalization):
weights = layer.get_weights()
H = BatchNormalization(weights=weights)(H)
elif isinstance(layer, Conv2D):
weights = layer.get_weights()
config['filters'] = weights[1].shape[0]
H = Conv2D(activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
data_format=config['data_format'],
dilation_rate=config['dilation_rate'],
filters=config['filters'],
kernel_constraint=config['kernel_constraint'],
kernel_regularizer=config['kernel_regularizer'],
kernel_size=config['kernel_size'],
name=config['name'],
padding=config['padding'],
strides=config['strides'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(H)
return Model(inp, H)
def build_model(self, layers, compile={}):
"""根據配置項構建 `self.model`。
Args:
layers ([dict]): 層定義集合。集合中每一項為一層的定義。
層定義包含 `type`:(dense或lstm) 用來定義層的類型。
層定義中其他屬性參見 `Dense`_ 和 `LSTM`_ 構造函數參數定義。
compile (dict): 訓練配置模型定義。定義可用屬性參見 `compile`_ 函數定義。
Returns:
.. _Dense:
https://keras.io/zh/layers/core/#dense
.. _LSTM:
https://keras.io/zh/layers/recurrent/#lstm
.. _compile:
https://keras.io/zh/models/model/#compile
"""
for layer in layers:
t = layer.pop('type')
if t == 'dense':
# https://keras.io/zh/layers/core/
self.__model.add(Dense.from_config(layer))
elif t == 'lstm':
# https://keras.io/zh/layers/recurrent/#lstm
self.__model.add(LSTM.from_config(layer))
elif t == 'dropout':
# https://keras.io/zh/layers/recurrent/#Dropout
self.__model.add(Dropout.from_config(layer))
elif t == 'cudnnlstm':
# https://keras.io/zh/layers/recurrent/#Dropout
self.__model.add(CuDNNLSTM.from_config(layer))
# https://keras.io/zh/models/model/#compile
self.__model.compile(**compile)
def read_model(src_model):
## args
## list of kernels to remove.
model = load_model(src_model)
inp = (model.inputs[0].shape.dims[1].value,
model.inputs[0].shape.dims[2].value,
model.inputs[0].shape.dims[3].value)
H = Input(inp)
inp = H
for i in range(len(model.layers)):
layer = model.get_layer(index=i)
config = layer.get_config()
if isinstance(layer, MaxPooling2D):
H = MaxPooling2D.from_config(config)(H)
if isinstance(layer, Dropout):
H = Dropout.from_config(config)(H)
if isinstance(layer, Activation):
H = Activation.from_config(config)(H)
elif isinstance(layer, Conv2D):
weights = layer.get_weights()
config['trainable'] = True
H = Conv2D(
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
data_format=config['data_format'],
dilation_rate=config['dilation_rate'],
filters=config['filters'],
kernel_constraint=config['kernel_constraint'],
# config=config['config'],
# scale=config['scale'],
kernel_regularizer=config['kernel_regularizer'],
kernel_size=config['kernel_size'],
name=config['name'],
padding=config['padding'],
strides=config['strides'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(H)
elif isinstance(layer, Flatten):
H = Flatten()(H)
elif isinstance(layer, Dense):
weights = layer.get_weights()
config['trainable'] = True
config['name'] = config['name'] + "1"
H = Dense(units=config['units'],
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
kernel_constraint=config['kernel_constraint'],
kernel_regularizer=config['kernel_regularizer'],
kernel_initializer='glorot_uniform',
name=config['name'],
trainable=config['trainable'],
use_bias=config['use_bias'])(H)
## it returns the model changed
return Model(inp, H)
def PrunWeight(model, model_name, x_prune, y_prune, x_test, y_test, pruning_rate, compile_info , fine_tune):
############ Calculating weight limit for pruning #####
############ We do not consider biases in the pruning process #####
parameters = []
conv_layers_weights = []
for layer in model.layers:
if layer.get_config()['name'].find("conv") != -1:
conv_layers_weights.append(layer.get_weights())
for _, layer_weights in enumerate(conv_layers_weights):
parameters.append(K.flatten(K.abs(layer_weights[0])))
dense_layers_weights = []
for layer in model.layers:
if layer.get_config()['name'].find("dense") != -1:
dense_layers_weights.append(layer.get_weights())
for _, layer_weights in enumerate(dense_layers_weights):
parameters.append(K.flatten(K.abs(layer_weights[0])))
parameters = K.concatenate(parameters)
parameters = sorted(K.get_value(parameters).tolist())
weight_limit = parameters[int(pruning_rate*len(parameters))]
print("Pruning weight threshhold : ", weight_limit)
##################################################################
dense_layers_weights = []
conv_filter_weights = []
batch_norm_params = []
kernel_masks_for_dense_and_conv_layers = []
model_tensors_dict = {}
input_height,input_width,input_channels = model.input.shape[1:]
pruned_model_input = Input(shape=(int(input_height),int(input_width),int(input_channels)))
if model.layers[0].name.find('input') == -1:
model_tensors_dict[str(model.layers[0].input.name)] = pruned_model_input
else:
model_tensors_dict[str(model.layers[0].output.name)] = pruned_model_input
Flow = pruned_model_input
for _,layer in enumerate(model.layers):
if layer.get_config()['name'].find("conv2d") != -1:
kernel_mask = K.cast(weight_limit <= K.abs(layer.get_weights()[0]) ,'float32')
kernel_masks_for_dense_and_conv_layers.append(kernel_mask)
Flow = MaskedConv2D(filters=layer.get_config()['filters'], kernel_size=layer.get_config()['kernel_size'],kernel_initializer=layer.get_config()['kernel_initializer'],
kernel_regularizer= layer.get_config()['kernel_regularizer'], strides=layer.get_config()['strides'],
padding=layer.get_config()['padding'], activation=layer.get_config()['activation'], use_bias=layer.get_config()['use_bias'], Masked=True , kernel_mask_val=kernel_mask)(model_tensors_dict[str(layer.input.name)])
conv_filter_weights.append(layer.get_weights())
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("dense") != -1:
kernel_mask = K.cast(weight_limit <= K.abs(layer.get_weights()[0]) ,'float32')
kernel_masks_for_dense_and_conv_layers.append(kernel_mask)
Flow = MaskedDense(units=layer.get_config()['units'], activation=layer.get_config()['activation'],
use_bias=layer.get_config()['use_bias'], kernel_initializer = layer.get_config()['kernel_initializer'],
Masked=True , kernel_mask_val=kernel_mask)(model_tensors_dict[str(layer.input.name)])
dense_layers_weights.append(layer.get_weights())
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("activation") != -1:
Flow = Activation.from_config(layer.get_config())(model_tensors_dict[str(layer.input.name)])
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("max_pooling") != -1:
Flow = MaxPooling2D.from_config(layer.get_config())(model_tensors_dict[str(layer.input.name)])
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("average_pooling") != -1:
Flow = AveragePooling2D.from_config(layer.get_config())(model_tensors_dict[str(layer.input.name)])
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("dropout") != -1:
Flow = Dropout.from_config(layer.get_config())(model_tensors_dict[str(layer.input.name)])
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("flatten") != -1:
Flow = Flatten.from_config(layer.get_config())(model_tensors_dict[str(layer.input.name)])
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("add") != -1:
input_tensors_list = []
for idx in range(len(layer.input)):
input_tensors_list.append(model_tensors_dict[layer.input[idx].name])
Flow = add(input_tensors_list)
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("batch_normalization") != -1:
batch_norm_params.append(layer.get_weights())
Flow = BatchNormalization.from_config(layer.get_config())(model_tensors_dict[str(layer.input.name)])
model_tensors_dict[str(layer.output.name)] = Flow
elif layer.get_config()['name'].find("input") != -1:
pass
pruned_model = Model(pruned_model_input, Flow)
########################## setting the weight s of layers #############################
for layer in pruned_model.layers:
if layer.get_config()['name'].find("dense") != -1:
pruned_weights = [dense_layers_weights[0][0]*K.get_value(kernel_masks_for_dense_and_conv_layers[0])]
if layer.get_config()['use_bias']:
pruned_weights.append(dense_layers_weights[0][1])
layer.set_weights(pruned_weights)
del kernel_masks_for_dense_and_conv_layers[0]
del dense_layers_weights[0]
elif layer.get_config()['name'].find("conv2d") != -1:
pruned_weights = [conv_filter_weights[0][0]*K.get_value(kernel_masks_for_dense_and_conv_layers[0])]
if layer.get_config()['use_bias']:
pruned_weights.append(conv_filter_weights[0][1])
layer.set_weights(pruned_weights)
del kernel_masks_for_dense_and_conv_layers[0]
del conv_filter_weights[0]
elif layer.get_config()['name'].find("batch") != -1:
layer.set_weights(batch_norm_params[0])
del batch_norm_params[0]
############################### Fine-tuning ############################################
pruned_model.compile(loss=compile_info['loss'],
optimizer=compile_info['optimizer'],
metrics=compile_info['metrics'])
if not fine_tune:
return pruned_model
else:
early_stopping = EarlyStopping(monitor='val_acc', patience=2,verbose=0)
callbacks = [early_stopping]
# fine-tuning the network.
pruned_model.fit(x_prune, y_prune,
batch_size=256,
epochs=10,
validation_data=(x_test, y_test),
shuffle=True,
callbacks=callbacks,
verbose=0
)
return pruned_model
def rebuild_model(self, kernel_list=None):
## args
## list of kernels to remove.
inp = (self.model.inputs[0].shape.dims[1].value,
self.model.inputs[0].shape.dims[2].value,
self.model.inputs[0].shape.dims[3].value)
H = Input(inp)
inp = H
for i in range(len(self.model.layers)):
layer = self.model.get_layer(index=i)
config = layer.get_config()
if isinstance(layer, MaxPooling2D):
H = MaxPooling2D.from_config(config)(H)
if isinstance(layer, Dropout):
H = Dropout.from_config(config)(H)
if isinstance(layer, Activation):
H = Activation.from_config(config)(H)
elif isinstance(layer, Conv2D):
weights = layer.get_weights()
if i == self.target_layer:
weights[0] = np.delete(weights[0], kernel_list, axis=3)
if (len(weights) == 2):
weights[1] = np.delete(weights[1], kernel_list, 0)
else:
if i == self.changed_layer:
weights[0] = np.delete(weights[0], kernel_list, axis=2)
config['filters'] = weights[0].shape[3]
H = Conv2D(
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
data_format=config['data_format'],
dilation_rate=config['dilation_rate'],
filters=config['filters'],
kernel_constraint=config['kernel_constraint'],
# config=config['config'],
# scale=config['scale'],
kernel_regularizer=config['kernel_regularizer'],
kernel_size=config['kernel_size'],
name=config['name'],
padding=config['padding'],
strides=config['strides'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(H)
elif isinstance(layer, Flatten):
H = Flatten()(H)
elif isinstance(layer, Dense):
weights = layer.get_weights()
if i == self.changed_layer:
shape = self.model.layers[i - 1].input_shape
new_weights = np.zeros(
(shape[1] * shape[2] * (shape[3] - len(kernel_list)),
weights[0].shape[1]))
for j in range(weights[0].shape[1]):
new_weights[:, j] = np.delete(
weights[0][:, j].reshape(
(shape[1], shape[2], shape[3])), kernel_list,
2).reshape(-1)
weights[0] = new_weights
config['units'] = weights[0].shape[1]
H = Dense(units=config['units'],
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
kernel_constraint=config['kernel_constraint'],
kernel_regularizer=config['kernel_regularizer'],
name=config['name'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(H)
## it returns the model changed
return Model(inp, H)
def get_partial_model(self):
inp = (self.prunedmodel.inputs[0].shape.dims[1].value,
self.prunedmodel.inputs[0].shape.dims[2].value,
self.prunedmodel.inputs[0].shape.dims[3].value)
inp = Input(inp)
i = 0
while (i <= self.changed_layer):
layer = self.prunedmodel.get_layer(index=i)
config = layer.get_config()
if isinstance(layer, MaxPooling2D):
H = MaxPooling2D(pool_size=config['pool_size'],
strides=config['strides'],
name=config['name'])(H)
if isinstance(layer, Dropout):
H = Dropout.from_config(config)(H)
if isinstance(layer, Activation):
H = Activation.from_config(config)(H)
elif isinstance(layer, Conv2D):
weights = self.prunedmodel.layers[i].get_weights()
if i == 1:
H = Conv2D(
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
data_format=config['data_format'],
dilation_rate=config['dilation_rate'],
filters=config['filters'],
kernel_constraint=config['kernel_constraint'],
# config=config['config'],
# scale=config['scale'],
kernel_regularizer=config['kernel_regularizer'],
kernel_size=config['kernel_size'],
name=config['name'],
padding=config['padding'],
strides=config['strides'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(inp)
else:
H = Conv2D(
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
data_format=config['data_format'],
dilation_rate=config['dilation_rate'],
filters=config['filters'],
kernel_constraint=config['kernel_constraint'],
# config=config['config'],
# scale=config['scale'],
kernel_regularizer=config['kernel_regularizer'],
kernel_size=config['kernel_size'],
name=config['name'],
padding=config['padding'],
strides=config['strides'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(H)
elif isinstance(layer, Flatten):
H = Flatten()(H)
elif isinstance(layer, Dense):
weights = layer.get_weights()
H = Dense(units=config['units'],
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
kernel_constraint=config['kernel_constraint'],
kernel_regularizer=config['kernel_regularizer'],
name=config['name'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(H)
i += 1
layer = self.prunedmodel.get_layer(index=i)
if isinstance(layer, MaxPooling2D):
config = layer.get_config()
H = MaxPooling2D(pool_size=config['pool_size'],
strides=config['strides'],
name=config['name'])(H)
i += 1
return Model(inp, H), (i - 1)
def rebuild_net(model=None, layer_filters=[]):
n_discarded_filters = 0
total_filters = 0
model = model
inp = (model.inputs[0].shape.dims[1].value,
model.inputs[0].shape.dims[2].value,
model.inputs[0].shape.dims[3].value)
H = Input(inp)
inp = H
idx_previous = []
for i in range(1, len(model.layers)):
layer = model.get_layer(index=i)
config = layer.get_config()
if isinstance(layer, MaxPooling2D):
H = MaxPooling2D.from_config(config)(H)
if isinstance(layer, Dropout):
H = Dropout.from_config(config)(H)
if isinstance(layer, Activation):
H = Activation.from_config(config)(H)
if isinstance(layer, BatchNormalization):
weights = layer.get_weights()
weights[0] = np.delete(weights[0], idx_previous)
weights[1] = np.delete(weights[1], idx_previous)
weights[2] = np.delete(weights[2], idx_previous)
weights[3] = np.delete(weights[3], idx_previous)
H = BatchNormalization(weights=weights)(H)
elif isinstance(layer, Conv2D):
weights = layer.get_weights()
n_filters = weights[0].shape[3]
total_filters = total_filters + n_filters
idxs = [item for item in layer_filters if item[0] == i][0][1]
weights[0] = np.delete(weights[0], idxs, axis=3)
weights[1] = np.delete(weights[1], idxs)
n_discarded_filters += len(idxs)
if len(idx_previous) != 0:
weights[0] = np.delete(weights[0], idx_previous, axis=2)
config['filters'] = weights[1].shape[0]
H = Conv2D(
activation=config['activation'],
activity_regularizer=config['activity_regularizer'],
bias_constraint=config['bias_constraint'],
bias_regularizer=config['bias_regularizer'],
data_format=config['data_format'],
dilation_rate=config['dilation_rate'],
filters=config['filters'],
kernel_constraint=config['kernel_constraint'],
# config=config['config'],
# scale=config['scale'],
kernel_regularizer=config['kernel_regularizer'],
kernel_size=config['kernel_size'],
name=config['name'],
padding=config['padding'],
strides=config['strides'],
trainable=config['trainable'],
use_bias=config['use_bias'],
weights=weights)(H)
idx_previous = idxs
print('Percentage of discarded filters {}'.format(n_discarded_filters /
float(total_filters)))
return Model(inp, H)
