def apply_transformations(in_model):
src_model = None
src_model_config = in_model.get_config()
k = 1024
for info_txt, check_func, apply_func in zip(info_list[:k], check_transform_list[:k], apply_transform_list[:k]):
if check_func(src_model_config):
if src_model is None:
print('Preparation for the transformation...\n')
src_model = Model.from_config(in_model.get_config(), custom_objects=extra_custom_objects)
transfer_weights(in_model, src_model, {})
check_transform(in_model, src_model)
print(info_txt)
src_model_config = src_model.get_config()
dst_model_config, weight_transfer_rule_dict = apply_func(src_model_config)
dst_model = Model.from_config(dst_model_config, custom_objects=extra_custom_objects)
transfer_weights(src_model, dst_model, weight_transfer_rule_dict)
check_transform(in_model, dst_model)
del src_model
src_model = dst_model
else:
print(f'Nothing to do with {info_txt.split(":")[0]} transform\n')
if src_model is None:
return in_model
return src_model
def create_dropout_predict_function(model, dropout):
"""
Hard-codes a dropout rate to the Dropout layers of a trained model.
When initially training model, Dropout layers must have training=True,
otherwise dropout will not be applied to predictions.
Parameters:
model : trained keras model
dropout : dropout rate to apply to all Dropout layers
Returns:
predict_with_dropout : keras model that will apply dropout when making predictions
"""
# Load the config of the original model
conf = model.get_config()
# Add the specified dropout to all layers
for layer in conf['layers']:
# Dropout layers
if layer["class_name"]=="Dropout":
layer["config"]["rate"] = dropout
# Recurrent layers with dropout
elif "dropout" in layer["config"].keys():
layer["config"]["dropout"] = dropout
# Create a new model with specified dropout
if type(model)==Sequential:
# Sequential
model_dropout = Sequential.from_config(conf)
else:
# Functional
model_dropout = Model.from_config(conf)
model_dropout.set_weights(model.get_weights())
return model_dropout
def from_config(cls, config, custom_objects={}):
if type(custom_objects) is list:
custom_objects = {obj.__name__: obj for obj in custom_objects}
custom_objects.update(_get_cells())
model_config = config.pop('model_config')
model = Model.from_config(model_config, custom_objects)
return cls(model, **config)
def split(model, start, end):
confs = model.get_config()
kept_layers = set()
for i, l in enumerate(confs['layers']):
if i == 0:
confs['layers'][0]['config']['batch_input_shape'] = model.layers[
start].input_shape
if i != start:
# confs['layers'][0]['name'] += str(random.randint(0, 100000000))
confs['layers'][0]['config']['name'] = confs['layers'][0][
'name']
elif i < start or i > end:
continue
kept_layers.add(l['name'])
# filter layers
layers = [l for l in confs['layers'] if l['name'] in kept_layers]
layers[1]['inbound_nodes'][0][0][0] = layers[0]['name']
# set conf
confs['layers'] = layers
confs['input_layers'][0][0] = layers[0]['name']
confs['output_layers'][0][0] = layers[-1]['name']
# create new model
submodel = Model.from_config(confs)
for l in submodel.layers:
orig_l = model.get_layer(l.name)
if orig_l is not None:
l.set_weights(orig_l.get_weights())
return submodel
def adapt_keras_model(keras_model, model_name):
keras_model_config = keras_model.get_config()
keras_model_config['name'] = model_name
# assert len(keras_model_config['input_layers']) == len(
# keras_model_config['output_layers']) == 1, 'Multi IN/OUT is not supported'
if len(keras_model_config['output_layers']) != 1:
print("WARNING :: Multi Output Detected!!!")
if is_multi_output(keras_model_config):
adapted_model_config = split_output_nodes(keras_model_config)
# adapted_model_config = rename_layer(adapted_model_config, keras_model_config['input_layers'][0][0], 'data')
# adapted_model_config = rename_layer(adapted_model_config, keras_model_config['output_layers'][0][0], 'output')
adapted_keras_model = Model.from_config(
adapted_model_config, custom_objects=extra_custom_objects)
else:
adapted_keras_model = Model.from_config(
keras_model_config, custom_objects=extra_custom_objects)
adapted_keras_model.set_weights(keras_model.get_weights())
return adapted_keras_model
def make_predictions(model, weights, test_gen, lr):
config = model.get_config()
pmodel = Model.from_config(config) # copy of the model
pmodel.set_weights(
weights) #load saved weights with lowest validation loss
pmodel.compile(Adam(lr=lr),
loss='categorical_crossentropy',
metrics=['accuracy'])
print(
'Training has completed. Now loading test set to see how accurate the model is'
)
results = pmodel.evaluate(test_gen, verbose=0)
print('Model accuracy on Test Set is {0:7.2f} %'.format(results[1] * 100))
predictions = pmodel.predict(test_gen, verbose=0)
return predictions
def deepspam_load(path="model"):
global wordmap
global model
print('Loading Model...')
config=json.load(open(path+"/model.config","rt"))
model = Model.from_config(config)
model.load_weights(path+"/model.weights")
####weights=pickle.load(open("model.weights", "rb"))
#model.set_weights(weights)
try:
wordmap=pickle.load(open(path+"/model.wordmap-py3", "rb"))
except:
wordmap=pickle.load(open(path+"/model.wordmap-py2", "rb"))
#wordmap=[]
print("MODEL loaded! (%d words)"%(len(wordmap)))
return wordmap
def reset_pingpong(self):
""" Reset the models for pingpong training """
logger.debug("Resetting models")
# Clear models and graph
self.predictors = dict()
K.clear_session()
# Load Models for current training run
for model in self.networks.values():
model.network = Model.from_config(model.config)
model.network.set_weights(model.weights)
inputs = self.get_inputs()
self.build_autoencoders(inputs)
self.compile_predictors(initialize=False)
logger.debug("Reset models")
def deepspam_load(path="model"):
global wordmap
global model
# global tf_session
# print('Creating TF session...')
# config = tensorflow.ConfigProto(
# intra_op_parallelism_threads=1,
# allow_soft_placement=True
# )
# tf_session = tensorflow.Session(config=config)
# tensorflow.set_session(tf_session)
global tf_session
global tf_graph
print('Loading Model...')
config = json.load(open(path + "/model.config", "rt"))
model = Model.from_config(config)
model.load_weights(path + "/model.weights")
####weights=pickle.load(open("model.weights", "rb"))
#model.set_weights(weights)
try:
wordmap = pickle.load(open(path + "/model.wordmap-py3", "rb"))
except:
wordmap = pickle.load(open(path + "/model.wordmap-py2", "rb"))
#wordmap=[]
print("MODEL loaded! (%d words)" % (len(wordmap)))
# testing
data = np.zeros((1, MAX_SEQUENCE_LENGTH), dtype='int32')
classes = model.predict(data, batch_size=1, verbose=1)
tf_session = K.get_session()
tf_graph = tensorflow.get_default_graph()
tf_graph.finalize()
print("MODEL finalized!")
return wordmap
def from_config(cls, config, custom_objects={}):
if type(custom_objects) is list:
custom_objects = {obj.__name__: obj for obj in custom_objects}
custom_objects.update(_get_cells())
config = config.copy()
model_config = config.pop('model_config')
if model_config is None:
model = None
else:
model = Model.from_config(model_config, custom_objects)
if type(model.input) is list:
input = model.input[0]
initial_states = model.input[1:]
else:
input = model.input
initial_states = None
if type(model.output) is list:
output = model.output[0]
final_states = model.output[1:]
else:
output = model.output
final_states = None
return cls(input, output, initial_states, final_states, **config)
def deepspam_worker(q, path):
print('Loading Model...')
config = json.load(open(path + "/model.config", "rt"))
model = Model.from_config(config)
model.load_weights(path + "/model.weights")
####weights=pickle.load(open("model.weights", "rb"))
#model.set_weights(weights)
#wordmap=[]
# print("MODEL loaded! (%d words)"%(len(wordmap)))
print("worker: started!!!")
while True:
dprint("worker: waiting for job... (%d)" % (q.qsize()))
data, retq = q.get()
if not retq:
break
dprint("worker: new job!")
classes = model.predict(data, batch_size=1, verbose=1)
res = classes[0][0] * 100.0 / (classes[0][0] + classes[0][1])
dprint("worker: result=" + str(res))
retq.put(res)
q.task_done()
print("worker: exiting!!!")
q.task_done()
# ============================================================== #
optimizer = tf.keras.optimizers.Adam(lr=LEARNING_RATE)
optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(
opt=optimizer, loss_scale=loss_scaler)
print('Optimizer Configuration:')
print('\nOptimizer `get_slot_names()`: %s\n' % str(optimizer.get_slot_names()))
autoencoder = Model(inputs, decoded)
autoencoder.compile(optimizer=optimizer, loss='mae', metrics=['mae'])
# ============ Test Model Recreation from config ============ #
model_config = autoencoder.get_config()
autoencoder = Model.from_config(model_config)
autoencoder.compile(optimizer=optimizer,
loss='mae',
metrics=['mae'],
run_eagerly=False)
# ============ Test Model Summary ============ #
autoencoder.summary()
from callback import ProgbarLogger
progbar_callback = ProgbarLogger(count_mode='samples',
stateful_metrics=['mae'])
print("\nEvaluation Before training - At Initialization")
autoencoder.evaluate(x=df_eval_conv,
def _createMobilenetv2Backbone(self, output_stride=16):
mobilenetv2 = MobileNetV2(weights='imagenet',
input_shape=(self.dl_input_shape[1],
self.dl_input_shape[2], 3),
include_top=False)
mobile_config = mobilenetv2.get_config()
mobile_weights = mobilenetv2.get_weights()
#tf.keras.backend.clear_session()
assert output_stride in [
8, 16
], "Only suported output_stride= 8 o 16 for backbone mobilenetv2."
dilatation = 1
for layer in mobile_config['layers']:
if layer['name'] == 'input_1':
layer['config']['batch_input_shape'] = (
None, self.dl_input_shape[-3], self.dl_input_shape[-2],
self.dl_input_shape[-1])
self.logger.info(layer['name'] + ', ' +
str(layer['config']['batch_input_shape']))
if output_stride == 8:
if layer['name'] == 'block_6_depthwise':
layer['config']['strides'] = (1, 1)
dilatation = dilatation * 2 #Replace stride for dilatation
self.logger.info(layer['name'] + ', strides=' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_7_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_8_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_9_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_10_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_11_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_12_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if output_stride in [8, 16]:
if layer['name'] == 'block_13_depthwise':
layer['config']['strides'] = (1, 1)
dilatation = dilatation * 2 #Replace stride for dilatation
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_14_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_15_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
if layer['name'] == 'block_16_depthwise':
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
# tf.keras.backend.clear_session()
new_mobilenetv2 = Model.from_config(mobile_config)
mobile_weights[0] = np.resize(mobile_weights[0],
[3, 3, self.dl_input_shape[-1], 32])
new_mobilenetv2.set_weights(mobile_weights)
self.backbone = Model(
inputs=new_mobilenetv2.inputs,
outputs=new_mobilenetv2.get_layer('block_16_project_BN').output)
self.strideOutput32LayerName = 'block_16_project_BN'
self.strideOutput16LayerName = 'block_12_add'
self.strideOutput8LayerName = 'block_5_add'
self.inputLayerName = mobilenetv2.layers[0].name
def __build_sample_model(self, batch_input_length) -> dict:
"""
Model that predicts a single OHE character.
This model is generated from the modified config file of the self.batch_model.
Returns:
The dictionary of the configuration.
"""
self.__batch_input_length = batch_input_length
# Get the configuration of the batch_model
config = self.model.get_config()
# Keep only the "Decoder_Inputs" as single input to the sample_model
config["input_layers"] = [config["input_layers"][0]]
# Find decoder states that are used as inputs in batch_model and remove them
idx_list = []
for idx, layer in enumerate(config["layers"]):
if "Decoder_State_" in layer["name"]:
idx_list.append(idx)
# Pop the layer from the layer list
# Revert indices to avoid re-arranging after deleting elements
for idx in sorted(idx_list, reverse=True):
config["layers"].pop(idx)
# Remove inbound_nodes dependencies of remaining layers on deleted ones
for layer in config["layers"]:
idx_list = []
try:
for idx, inbound_node in enumerate(layer["inbound_nodes"][0]):
if "Decoder_State_" in inbound_node[0]:
idx_list.append(idx)
# Catch the exception for first layer (Decoder_Inputs) that has empty list of inbound_nodes[0]
except:
pass
# Pop the inbound_nodes from the list
# Revert indices to avoid re-arranging
for idx in sorted(idx_list, reverse=True):
layer["inbound_nodes"][0].pop(idx)
# Change the batch_shape of input layer
config["layers"][0]["config"]["batch_input_shape"] = (
batch_input_length,
1,
self.dec_input_shape[-1],
)
# Finally, change the statefulness of the RNN layers
for layer in config["layers"]:
if "Decoder_LSTM_" in layer["name"]:
layer["config"]["stateful"] = True
# layer["config"]["return_sequences"] = True
# Define the sample_model using the modified config file
sample_model = Model.from_config(config)
# Copy the trained weights from the trained batch_model to the untrained sample_model
for layer in sample_model.layers:
# Get weights from the batch_model
weights = self.model.get_layer(layer.name).get_weights()
# Set the weights to the sample_model
sample_model.get_layer(layer.name).set_weights(weights)
if batch_input_length == 1:
self.__sample_model = sample_model
elif batch_input_length > 1:
self.__multi_sample_model = sample_model
return config
def _createVGG16Backbone(self, output_stride=16):
vgg16 = VGG16(weights='imagenet',
input_shape=(self.dl_input_shape[1],
self.dl_input_shape[2], 3),
include_top=False)
assert output_stride in [
8, 16
], "Only suported output_stride= 8 o 16 for backbone vgg16."
mobile_config = vgg16.get_config()
mobile_weights = vgg16.get_weights()
#tf.keras.backend.clear_session()
dilatation = 1
stride_enable = False
for layer in mobile_config['layers']:
if layer['name'] == 'input_1':
layer['config']['batch_input_shape'] = (
None, self.dl_input_shape[-3], self.dl_input_shape[-2],
self.dl_input_shape[-1])
self.logger.info(layer['name'] + ', ' +
str(layer['config']['batch_input_shape']))
if output_stride == 8:
if layer['name'] == 'block4_pool':
layer['config']['strides'] = (1, 1)
layer['config']['pool_size'] = (1, 1)
dilatation = dilatation * 2 #Replace stride for dilatation
self.logger.info(layer['name'] + ', strides=' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['pool_size']))
stride_enable = True
if output_stride in [8, 16]:
if layer['name'] == 'block5_pool':
layer['config']['strides'] = (1, 1)
layer['config']['pool_size'] = (1, 1)
dilatation = dilatation * 2 #Replace stride for dilatation
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['pool_size']))
stride_enable = True
if 'conv' in layer['name'] and stride_enable:
layer['config']['dilation_rate'] = (dilatation, dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
new_vgg16 = Model.from_config(mobile_config)
mobile_weights[0] = np.resize(mobile_weights[0], [
mobile_weights[0].shape[0], mobile_weights[0].shape[1],
self.dl_input_shape[-1], mobile_weights[0].shape[-1]
]) #update input to suport 4 channels
new_vgg16.set_weights(mobile_weights)
x = new_vgg16.output
x = Conv2D(filters=512,
kernel_size=3,
name='block_6_conv1',
padding="same",
dilation_rate=dilatation,
activation='relu')(x)
x = Conv2D(filters=512,
kernel_size=3,
name='block_6_conv2',
padding="same",
dilation_rate=dilatation,
activation='relu')(x)
x = Conv2D(filters=512,
kernel_size=3,
name='block_6_conv3',
padding="same",
dilation_rate=dilatation,
activation='relu')(x)
self.backbone = Model(inputs=new_vgg16.inputs, outputs=x)
self.strideOutput32LayerName = 'block6_conv3'
self.strideOutput16LayerName = 'block5_conv3'
self.strideOutput8LayerName = 'block4_conv3'
self.inputLayerName = vgg16.layers[0].name
def convert(prevmodel, export_path, export_name):
graph1 = tf.Graph()
with graph1.as_default():
sess1 = tf.Session()
with sess1.as_default():
previous_model = load_model(prevmodel)
previous_model.summary()
config = previous_model.get_config()
weights = previous_model.get_weights()
graph2 = tf.Graph()
with graph2.as_default():
sess2 = tf.Session()
with sess2.as_default():
K.set_learning_phase(0)
try:
model = Sequential.from_config(config)
except:
model = Model.from_config(config)
model.set_weights(weights)
with open(
os.path.join(export_path,
export_name + '.input_output.txt'),
'w') as fid:
model_input_name = model.input.name
#model_input_node_name
model_output_name = model.output.name
print(f"Input name: {model_input_name}")
print(f"Output name: {model_output_name}")
fid.write(f"{model_input_name}\n")
fid.write(f"{model_output_name}\n")
model_output_node_name = model.output.name.split(':')[0]
graph_file = os.path.join(export_path,
export_name + ".graph.pbtxt")
ckpt_file = os.path.join(export_path, export_name + ".ckpt")
saver = tf.train.Saver()
tf.train.write_graph(sess2.graph_def, '', graph_file)
save_path = saver.save(sess2, ckpt_file)
# freeze the graph
frozen_graph_def = tf.graph_util.convert_variables_to_constants(
sess2, # The session is used to retrieve the weights
graph2.as_graph_def(
), # The graph_def is used to retrieve the nodes
model_output_node_name.split(
","
) # The output node names are used to select the usefull nodes
)
# Finally we serialize and dump the output graph to the filesystem
frozen_graph_path = os.path.join(export_path,
export_name + ".frozen.pb")
with tf.gfile.GFile(frozen_graph_path, "wb") as f:
f.write(frozen_graph_def.SerializeToString())
#tf.reset_default_graph()
#frozen_graph_def = freeze_graph(export_path, model_output_node_name)
tf.reset_default_graph()
train_writer = tf.summary.FileWriter(export_path)
train_writer.add_graph(frozen_graph_def)
train_writer.flush()
# optimize for inference
optimized_graph_def = optimize_for_inference_lib.optimize_for_inference(
frozen_graph_def, [model_input_name.split(':')[0]],
[model_output_name.split(':')[0]], tf.float32.as_datatype_enum)
#tf.int32.as_datatype_enum
optimized_graph_path = os.path.join(
export_path, export_name + ".optimized_for_inference.pb")
with tf.gfile.GFile(optimized_graph_path, "wb") as f:
f.write(optimized_graph_def.SerializeToString())
# tflite
# ## from frozen graph
converter = tflite.TocoConverter.from_frozen_graph(
optimized_graph_path, [model_input_name.split(':')[0]],
[model_output_name.split(':')[0]],
{model_input_name.split(':')[0]: [1, 32, 32, 3]})
#
## from keras model file
# only available in tensorflow >=1.11
#converter = tflite.TocoConverter.from_keras_model_file(prevmodel)
converter.post_training_quantize = True
#converter.inference_type = tf.quint8
#converter.inference_input_type = tf.float32
tflite_quantized_model = converter.convert()
optimized_graph_path = os.path.join(export_path, export_name + ".tflite")
open(optimized_graph_path, "wb").write(tflite_quantized_model)
def _createResnetBackbone(self, output_stride=16, depth='101'):
resnet101 = ResNet101(weights='imagenet',
input_shape=(self.dl_input_shape[1],
self.dl_input_shape[2], 3),
include_top=False)
assert output_stride in [
8, 16
], "Only suported output_stride= 8 o 16 for backbone resnet."
resnet101_config = resnet101.get_config()
resnet101_weights = resnet101.get_weights()
#tf.keras.backend.clear_session()
output_stride = 8
dilatation = 1
stride_enable = False
for layer in resnet101_config['layers']:
if layer['name'] == 'input_1':
layer['config']['batch_input_shape'] = (
None, self.dl_input_shape[-3], self.dl_input_shape[-2],
self.dl_input_shape[-1])
self.logger.info(layer['name'] + ', ' +
str(layer['config']['batch_input_shape']))
if output_stride == 8 and (layer['name'] == 'conv4_block1_1_conv'
or layer['name']
== 'conv4_block1_0_conv'):
layer['config']['strides'] = (1, 1)
self.logger.info(layer['name'] + ', strides=' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
stride_enable = True
if layer['name'] == 'conv4_block1_1_conv':
dilatation = dilatation * 2 #Replace stride for dilatation
elif output_stride in [8, 16]:
if layer['name'] == 'conv5_block1_1_conv' or layer[
'name'] == 'conv5_block1_0_conv':
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
layer['config']['strides'] = (1, 1)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) + ', ' +
str(layer['config']['dilation_rate']))
stride_enable = True
if layer['name'] == 'conv5_block1_1_conv':
dilatation = dilatation * 2 #Replace stride for dilatation
elif stride_enable and ('_conv' in layer['name']):
if layer['config']['kernel_size'] != (1, 1):
layer['config']['dilation_rate'] = (dilatation,
dilatation)
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) +
', ' +
str(layer['config']['dilation_rate']))
else:
self.logger.info(layer['name'] + ', ' +
str(layer['config']['strides']) +
', ' +
str(layer['config']['dilation_rate']))
self.backbone = Model.from_config(resnet101_config)
resnet101_weights[0] = np.resize(resnet101_weights[0], [
resnet101_weights[0].shape[0], resnet101_weights[0].shape[1],
self.dl_input_shape[-1], resnet101_weights[0].shape[-1]
]) #update input to suport 4 channels
self.backbone.set_weights(resnet101_weights)
self.strideOutput32LayerName = 'conv5_block3_out'
self.strideOutput16LayerName = 'conv4_block23_out'
self.strideOutput8LayerName = 'conv3_block4_out'
self.inputLayerName = resnet101.layers[0].name
def partial_trainability(model_keras, Layer_names, Layer_trainabliities):
"""
This function allows you to get partially trainable layers
function takes
model_keras: keras model you have loaded or created for example using model_zoo.get_model("VGG_small_4",32,3,2)
Layer_names: list containing the names of the layers that should be changed. Of course these names have to be existing in model_keras
Layer_trainability: list containing floats. For each layer that should be changed, provide a value between 0 and 1 (0-layer not trainable at all; 1-layer entirely trainable)
Lets say you use Layer_names=['dense_3'] and Layer_trainability=[0.25]
Lets assume 'dense_3' has 1000 nodes.
Then two new parallell dense layers will be created.
The first one gets 750 nodes, which are set to set to Trainable=False.
The second dense layer has 250 nodes, which are trainable.
The corresponding weights from the initial model are distributed to these
layer accordingly.
"""
#Get the config of the current model
model_config = model_keras.get_config()
#Deep-Copy the config of the original model
model_config_new = copy.deepcopy(model_config)
if type(model_config_new) == dict:
print("Model used functional API of Keras")
elif type(model_config_new) == list:
print(
"Model used sequential API of Keras and will now be converted to functional API"
)
#api = "Sequential"
#Convert to functional API
input_layer = Input(batch_shape=model_keras.layers[0].input_shape)
prev_layer = input_layer
for layer in model_keras.layers:
prev_layer = layer(prev_layer)
model_keras = Model([input_layer], [prev_layer])
#Now we have functional API :)
#Get the model config for the converted model
model_config = model_keras.get_config()
model_config_new = copy.deepcopy(model_config)
else:
print("Unknown format for model config")
#return
Layer_nontrainability = list(1 - np.array(Layer_trainabliities))
del Layer_trainabliities
#Original names of the layers
Layer_names_orig = [
model_keras.layers[i].name for i in range(len(model_keras.layers))
]
#Now we are starting to loop over all the requested changes:
for i in range(len(Layer_names)):
nontrainability = Layer_nontrainability[i]
layer_name = Layer_names[i]
layer_names_list = [
model_config_new["layers"][it]['name']
for it in range(len(model_config_new["layers"]))
]
layer_index = int(
np.where(np.array(layer_names_list) == layer_name)[0])
layer_config = model_config_new['layers'][layer_index]
layer_type = layer_config["class_name"]
if layer_type == "Dense":
split_property = "units" #'units' are the number of nodes in dense layers
elif layer_type == "Conv2D":
split_property = "filters"
nr_nodes = layer_config["config"][split_property]
nr_const = int(np.round(nontrainability * nr_nodes))
#get a config for the non-trainable part
layer_config_const = copy.deepcopy(layer_config)
layer_name_const = layer_config_const['config']["name"] + "_1"
layer_config_const["config"][split_property] = nr_const
layer_config_const["config"]["trainable"] = False #not trainable
layer_config_const["config"]["name"] = layer_name_const #rename
layer_config_const["name"] = layer_name_const #rename
#get a config for the rest (trainable part)
layer_config_rest = copy.deepcopy(layer_config)
#this part will only exist if nr_nodes-nr_const>0:
if nr_nodes - nr_const > 0:
layer_name_rest = layer_config_rest["config"]["name"] + "_2"
layer_config_rest["config"][split_property] = nr_nodes - nr_const
layer_config_rest["config"]["name"] = layer_name_rest #rename
layer_config_rest["name"] = layer_name_rest #rename
#Assemble a config for a corresponding concatenate layer
inbound_1 = [layer_name_const, 0, 0, {}]
inbound_2 = [layer_name_rest, 0, 0, {}]
inbound_nodes = [inbound_1, inbound_2]
layer_name = layer_config['config'][
"name"] #Call it like the initial layer that has been there. This will allow other layers to correctly conncet like before #'concatenate_'+layer_config["config"]["name"]
config_conc = {"axis": -1, "name": layer_name, "trainable": True}
conc_dict = {
"class_name": 'Concatenate',
"config": config_conc,
"inbound_nodes": [inbound_nodes],
"name": layer_name
}
#insert these layers into the config at Index: layer_index
layerlist = model_config_new["layers"]
#Replace existing layer with the const part
layerlist[layer_index] = layer_config_const
#After that insert the rest part
layerlist.insert(layer_index + 1, layer_config_rest)
#After that insert the concatenate layer
layerlist.insert(layer_index + 2, conc_dict)
#Update the config with the new layers
model_config_new["layers"] = layerlist
#Build the model using this updated config
model_keras_new = Model.from_config(model_config_new)
#Compilation might not be required.
model_keras_new.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
iterator = 0
for layer_name in Layer_names_orig: #for all layers of the original model
layer = model_keras.get_layer(layer_name)
if layer_name not in Layer_names: #if this layer was not subjected to change in the new model...
#Straight forward: get weights of the orignal model
weights = layer.get_weights()
#And put them into the new model
model_keras_new.get_layer(layer_name).set_weights(weights)
else: #Otherwise the layer was changed
layer_type = layer.__class__.__name__
layer_config = layer.get_config()
weights = layer.get_weights() #Get the original weights
trainability = Layer_nontrainability[iterator]
iterator += 1
if layer_type == "Dense":
split_property = "units" #'units' are the number of nodes in dense layers
elif layer_type == "Conv2D":
split_property = "filters"
nr_nodes = layer_config[split_property]
nr_const = int(np.round(trainability * nr_nodes))
#Constant part
if layer_type == "Dense":
#Put the former weights into the layer
weights_const = list(np.copy(weights))
weights_const[0] = weights_const[0][:, 0:nr_const]
weights_const[1] = weights_const[1][0:nr_const]
elif layer_type == "Conv2D":
#Put the former weights into the layer
weights_const = list(np.copy(weights))
weights_const[0] = weights_const[0][:, :, :, 0:nr_const]
weights_const[1] = weights_const[1][0:nr_const]
else:
print("Unknown layer type")
#return
layer_name_const = layer_name + "_1" #rename
model_keras_new.get_layer(layer_name_const).set_weights(
weights_const)
#Rest trainable part
#this part will only exist if nr_nodes-nr_const>0:
if nr_nodes - nr_const > 0:
if layer_type == "Dense":
#Get the weights
weights_rest = list(np.copy(weights))
weights_rest[0] = weights_rest[0][:, nr_const:]
weights_rest[1] = weights_rest[1][nr_const:]
if layer_type == "Conv2D":
#Get the weights
weights_rest = list(np.copy(weights))
weights_rest[0] = weights_rest[0][:, :, :, nr_const:]
weights_rest[1] = weights_rest[1][nr_const:]
layer_name_rest = layer_name + "_2" #rename
model_keras_new.get_layer(layer_name_rest).set_weights(
weights_rest)
return model_keras_new
def build_pruned_model(model, new_model_param, layer_types, num_new_neurons,
num_new_filters, comp):
"""
The new number of neurons and filters are changed in the model config.
Load the new weight matrices into the model.
Args:
model: Model which should be pruned
new_model_param: Stores the new weights of the model
layer_types: The types of all layers of the model
num_new_neurons: Number of neurons of the dense layers
num_new_filters: Number of filters of the conv layers
Return:
pruned_model: New model after pruning all dense and conv layers
"""
model_config = model.get_config()
'''
For functional model first layer is the input layer.
For sequential model the first layer is the layer after the input layer
'''
a = 1
if layer_types[0] == 'InputLayer':
a = 0
for i in range(0, len(model_config['layers']) - 3):
if model_config['layers'][i + a][
'class_name'] == "Dense": #i+1 because first layer of model is the inputlayer
print("Dense")
model_config['layers'][i +
a]['config']['units'] = num_new_neurons[i]
elif model_config['layers'][i + a]['class_name'] == "Conv2D":
print("Conv2D")
model_config['layers'][i +
a]['config']['filters'] = num_new_filters[i]
elif model_config['layers'][i + a]['class_name'] == "Reshape":
temp_list = list(
model_config['layers'][i + a]['config']['target_shape'])
cur_layer = i
cur_filters = num_new_filters[cur_layer]
#Get number of filters of last Conv layer
if cur_filters == 0:
while cur_filters == 0:
cur_layer -= 1
cur_filters = num_new_filters[cur_layer]
temp_list[2] = cur_filters
temp_tuple = tuple(temp_list)
model_config['layers'][i +
a]['config']['target_shape'] = temp_tuple
else:
print("No Dense or Conv2D")
print("Before pruning:")
model.summary()
if "Sequential" in str(model):
pruned_model = Sequential.from_config(model_config)
elif "Functional" in str(model):
pruned_model = Model.from_config(model_config)
print("After pruning:")
pruned_model.summary()
for i in range(0, len(pruned_model.layers)):
if len(new_model_param[i]) != 0:
pruned_model.layers[i].set_weights(new_model_param[i])
else:
None
pruned_model.compile(**comp)
return pruned_model
