def get_custom_activations_dict(filepath=None):
"""
Import all implemented custom activation functions so they can be used when
loading a Keras model.
Parameters
----------
filepath : Optional[str]
Path to json file containing additional custom objects.
"""
from snntoolbox.utils.utils import binary_sigmoid, binary_tanh, \
ClampedReLU
# Todo: We should be able to load a different activation for each layer.
# Need to remove this hack:
activation_str = 'relu_Q1.4'
activation = get_quantized_activation_function_from_string(activation_str)
return {'binary_sigmoid': binary_sigmoid,
'binary_tanh': binary_tanh,
# Todo: This should work regardless of the specific attributes of
# the ClampedReLU class used during training.
'clamped_relu': ClampedReLU(),
activation_str: activation,
'precision': precision,
'activity_regularizer': keras.regularizers.l1}
def get_clamped_relu_from_string(activation_str):
from snntoolbox.utils.utils import ClampedReLU
threshold, max_value = map(eval, activation_str.split('_')[-2:])
activation = ClampedReLU(threshold, max_value)
return activation
def get_custom_activations_dict():
"""
Import all implemented custom activation functions so they can be used when
loading a Keras model.
"""
from snntoolbox.utils.utils import binary_sigmoid, binary_tanh, ClampedReLU
# Todo: We should be able to load a different activation for each layer.
# Need to remove this hack:
activation_str = 'relu_Q1.4'
activation = get_quantized_activation_function_from_string(activation_str)
return {'binary_sigmoid': binary_sigmoid,
'binary_tanh': binary_tanh,
'clamped_relu': ClampedReLU(), # Todo: This should work regardless of the specific attributes of the ClampedReLU class used during training.
activation_str: activation}
def get_custom_activations_dict(filepath=None):
"""
Import all implemented custom activation functions so they can be used when
loading a Keras model.
Parameters
----------
filepath : Optional[str]
Path to json file containing additional custom objects.
"""
from snntoolbox.utils.utils import binary_sigmoid, binary_tanh, \
ClampedReLU, LimitedReLU, NoisySoftplus
import keras_metrics as km
# Todo: We should be able to load a different activation for each layer.
# Need to remove this hack:
activation_str = 'relu_Q1.4'
activation = get_quantized_activation_function_from_string(activation_str)
custom_objects = {
'binary_sigmoid': binary_sigmoid,
'binary_tanh': binary_tanh,
# Todo: This should work regardless of the specific attributes of the
# ClampedReLU class used during training.
'clamped_relu': ClampedReLU(),
'LimitedReLU': LimitedReLU,
'relu6': LimitedReLU({'max_value': 6}),
activation_str: activation,
'Noisy_Softplus': NoisySoftplus,
'precision': precision,
'binary_precision': km.binary_precision(label=0),
'binary_recall': km.binary_recall(label=0),
'activity_regularizer': keras.regularizers.l1}
if filepath is not None and filepath != '':
with open(filepath) as f:
kwargs = json.load(f)
for key in kwargs:
if 'LimitedReLU' in key:
custom_objects[key] = LimitedReLU(kwargs[key])
return custom_objects
# model_list[-2]), compile=False)
# # Experiment 3
# label = 'relu_0.1_1_bias_regularizer5'
# nonlinearity = ClampedReLU(0.1, 1.0)
# bias_regularizer = [l2(0.05), l2(0.9), l2(0.05), l2(0.5), l2(0.5), l2(0.5),
# l2(0.5), l2(0.01), l2(0.01)]
# model_list = os.listdir(os.path.join(tensorboard_path, 'relu_0.1_1'))
# print("Initializing with model {}.".format(model_list[-2]))
# model_init = load_model(os.path.join(
# tensorboard_path, 'relu_0.1_1', model_list[-2]),
# {'clamped_relu_0.1_1.0': nonlinearity}, False)
# Experiment 4
label = 'relu_0.1_1_bias_regularizer7'
nonlinearity = ClampedReLU(0.1, 1.0)
bias_regularizer = [l2(0.01), l2(2.0), l2(0.05), l2(3.0), l2(0.5), l2(1.0),
l2(1.0), l2(0.001), l2(0.01)]
model_list = os.listdir(os.path.join(tensorboard_path,
'relu_0.1_1_bias_regularizer6'))
print("Initializing with model {}.".format(model_list[-2]))
model_init = load_model(os.path.join(
tensorboard_path, 'relu_0.1_1_bias_regularizer6', model_list[-2]),
{'clamped_relu_0.1_1.0': nonlinearity}, False)
model = Sequential()
model.add(Conv2D(128, (3, 3), padding='same', activation=nonlinearity,
input_shape=(3, 32, 32), bias_regularizer=bias_regularizer[0]))
model.add(Conv2D(128, (3, 3), padding='same', activation=nonlinearity,
bias_regularizer=bias_regularizer[1]))
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import ModelCheckpoint, TensorBoard
from keras.optimizers import adam
from keras.regularizers import l2
from snntoolbox.utils.utils import ClampedReLU
path = '/home/rbodo/.snntoolbox/data/cifar10/binaryconnect'
dataset_path = '/home/rbodo/.snntoolbox/Datasets/cifar10/pylearn2_gcn_whitened'
tensorboard_path = os.path.join(path, 'training')
batch_size = 64
nb_epoch = 50
# Experiment 1
label = 'relu_schedule_bias_reg'
nonlinearity = [ClampedReLU(1 / (l + 1), 1 / l) for l in range(1, 9)]
bias_regularizer = [
l2(0.05),
l2(0.9),
l2(0.05),
l2(0.5),
l2(0.5),
l2(0.5),
l2(0.5),
l2(0.01),
l2(0.01)
]
model_list = os.listdir(
os.path.join(tensorboard_path, 'relu_0.1_1_bias_regularizer3'))
print("Initializing with model {}.".format(model_list[-2]))
print(dict({(n.__name__, n) for n in nonlinearity}))