def create_objects(status):
status.Data_gen=GeneratorFromCIFAR.GeneratorFromCIFAR(
status.Comp, status.S, cuda=status.cuda)
status.Data_gen.dataConv2d()
dataGen=status.Data_gen
x = dataGen.size[1]
y = dataGen.size[2]
max_layers=status.max_layer
max_filters=status.max_filter
def condition(DNA):
return max_filter(max_layer(DNA,max_layers),max_filters)
version=status.typos_version
center=((-1,1,3,x,y),
(0,3, 15, 3 , 3),
(0,18, 15, 3, 3),
(0,33, 15, x, y),
(1, 15,10),
(2,),
(3,-1,0),
(3,0,1),(3,-1,1),
(3,1,2),(3,0,2),(3,-1,2),
(3,2,3),
(3,3,4))
selector=status.Selector_creator(condition=condition,
directions=version)
status.Selector=selector
creator=status.Creator
selector.update(center)
actions=selector.get_predicted_actions()
space=DNA_Graph(center,1,(x,y),condition,actions,
version,creator)
if status.Alai:
stream=TorchStream(status.Data_gen,status.log_size,
min_size=status.min_log_size,
Alai=status.Alai)
else:
stream=TorchStream(status.Data_gen,status.log_size,
min_size=status.min_log_size)
status.stream=stream
Phase_space=DNA_Phase_space(space,
stream=stream)
Dynamics=Dynamic_DNA(space,Phase_space,status.dx,
Creator=creator,Selector=selector,
update_velocity=velocity_updater,
update_space=space_updater,version=version)
Phase_space.create_particles(status.n)
Phase_space.beta=status.beta
Phase_space.alpha=status.alpha
Phase_space.influence=status.influence
status.Dynamics=Dynamics
status.objects=Dynamics.objects
def create_objects(status, loaded_network):
settings = status.settings
status.Alai = Alai(min=status.dt_min,
max=status.dt_Max,
max_time=status.restart_period)
if loaded_network == True:
status.Alai.time = 14077 - 14076
status.Alai.reset_count = 1
status.Data_gen = GeneratorFromCIFAR.GeneratorFromCIFAR(
status.Comp,
status.S,
cuda=status.cuda,
threads=status.threads,
dataAugmentation=settings.enable_augmentation,
transforms_mode=settings.transformations_compose)
status.Data_gen.dataConv2d()
dataGen = status.Data_gen
x = dataGen.size[1]
y = dataGen.size[2]
condition = status.condition
mutations = status.mutations
version = status.version
center = status.Center
num_actions = status.num_actions
selector = status.Selector_creator(condition=condition,
directions=version,
num_actions=num_actions,
mutations=mutations)
status.Selector = selector
creator = status.Creator
selector.update(center)
actions = selector.get_predicted_actions()
space = DNA_Graph(center,
1, (x, y),
condition,
actions,
version,
creator,
selector=status.Selector,
num_morphisms=5)
if status.Alai:
stream = TorchStream(status.Data_gen,
status.log_size,
min_size=status.min_log_size,
Alai=status.Alai,
status=status)
else:
stream = TorchStream(status.Data_gen,
status.log_size,
min_size=status.min_log_size,
status=status)
status.stream = stream
Phase_space = DNA_Phase_space(space, stream=stream, status=status)
Dynamics = Dynamic_DNA(
space,
Phase_space,
status.dx,
Creator=creator,
Selector=selector,
update_velocity=velocity_updater,
update_space=space_updater,
version=version,
mutation_coefficient=status.mutation_coefficient,
clear_period=status.clear_period,
update_force_field=status.update_force_field,
)
Phase_space.create_particles(status.n)
Phase_space.beta = status.beta
Phase_space.alpha = status.alpha
Phase_space.influence = status.influence
status.Dynamics = Dynamics
status.objects = Dynamics.objects
def TestMemoryManager():
epochs = 0.2
batch_size = 64
def dropout_function(base_p, total_layers, index_layer, isPool=False):
value = 0
if index_layer != 0 and isPool == False:
value = base_p
if index_layer == total_layers - 2:
value = base_p
print("conv2d: ", index_layer, " - dropout: ", value, " - isPool: ",
isPool)
return value
settings = ExperimentSettings.ExperimentSettings()
settings.momentum = 0.9
settings.dropout_value = 0.05
settings.weight_decay = 0.0005
settings.enable_activation = True
settings.enable_last_activation = True
settings.enable_track_stats = True
settings.version = directions_version.CONVEX_VERSION
settings.eps_batchorm = 0.001
settings.dropout_function = dropout_function
settings.ricap = Augmentation.Ricap(beta=0.3)
dataGen = GeneratorFromCIFAR.GeneratorFromCIFAR(2,
batch_size,
threads=0,
dataAugmentation=True)
dataGen.dataConv2d()
memoryManager = MemoryManager.MemoryManager()
mutation_manager = MutationManager.MutationManager(
directions_version=settings.version)
dna = test_DNAs.DNA_base
e = 50000 / batch_size
e = math.ceil(e)
print("e: ", e)
print("total iterations: ", int(epochs * e))
dt_array = Alaising(1.2, 99, int(epochs * e))
input("press to continue: before load network")
network = nw_dendrites.Network(
dna,
cuda_flag=True,
momentum=settings.momentum,
weight_decay=settings.weight_decay,
enable_activation=settings.enable_activation,
enable_track_stats=settings.enable_track_stats,
dropout_value=settings.dropout_value,
dropout_function=settings.dropout_function,
enable_last_activation=settings.enable_last_activation,
version=settings.version,
eps_batchnorm=settings.eps_batchorm)
input("press to continue: before training network")
network.training_custom_dt(dataGenerator=dataGen,
dt_array=dt_array,
ricap=settings.ricap,
evalLoss=True)
network.generate_accuracy(dataGen)
print("net acc: ", network.get_accuracy())
input("press to continue: before save network")
memoryManager.saveTempNetwork(network)
input("press to continue: after save network")
input("press to continue: before load temp network")
network_loaded = memoryManager.loadTempNetwork(dna, settings)
input("press to continue: after load temp network")
network_loaded.generate_accuracy(dataGen)
print("loaded acc: ", network_loaded.get_accuracy())
input("press to continue: before mutate network (add filters layer 1)")
dna_mutate = direction_dna.increase_filters(1, network_loaded.dna)
network_mutate = mutation_manager.execute_mutation(network_loaded,
dna_mutate)
input("press to continue: after mutate network")
input("press to continue: before delete old network")
memoryManager.deleteNetwork(network_loaded)
input("press to continue: after delete old network")
network_mutate.generate_accuracy(dataGen)
print("mutated acc: ", network_mutate.get_accuracy())
input("press to conitnue: before training mutate network")
network_mutate.training_custom_dt(dataGenerator=dataGen,
dt_array=dt_array,
ricap=settings.ricap,
evalLoss=True)
input("press to conitnue: after training mutate network")
network_mutate.generate_accuracy(dataGen)
print("mutate net acc: ", network_mutate.get_accuracy())
input("press to continue: before save network")
memoryManager.saveTempNetwork(network_mutate)
input("press to continue: after save network")
input("press to continue: before load network")
network_loaded = memoryManager.loadTempNetwork(dna_mutate, settings)
input("press to continue: after load network")
network_loaded.generate_accuracy(dataGen)
print("loaded acc: ", network_loaded.get_accuracy())
input("press to continue: before mutate network (add layer pool 1)")
dna_mutate_2 = direction_dna.add_pool_layer(1, network_loaded.dna)
network_mutate = mutation_manager.execute_mutation(network_loaded,
dna_mutate_2)
input("press to continue: after mutate network")
input("press to continue: before delete old network")
memoryManager.deleteNetwork(network_loaded)
input("press to continue: after delete old network")
network_mutate.generate_accuracy(dataGen)
print("mutated acc: ", network_mutate.get_accuracy())
input("press to conitnue: before training mutate network")
network_mutate.training_custom_dt(dataGenerator=dataGen,
dt_array=dt_array,
ricap=settings.ricap,
evalLoss=True)
input("press to conitnue: after training mutate network")
network_mutate.generate_accuracy(dataGen)
print("mutate net acc: ", network_mutate.get_accuracy())
input("press to continue: before save network")
memoryManager.saveTempNetwork(network_mutate)
input("press to continue: after save network")
def Test_Mutacion():
memoryManager = MemoryManager.MemoryManager()
augSettings = AugmentationSettings.AugmentationSettings()
list_transform = {
augSettings.randomHorizontalFlip: True,
augSettings.translate: True,
}
PARENT_DNA = ((-1, 1, 3, 32, 32), (0, 3, 64, 3, 3), (0, 64, 128, 3, 3, 2),
(0, 192, 256, 3, 3, 2), (0, 256, 256, 13, 13), (1, 256, 10),
(2, ), (3, -1, 0), (3, 0, 1), (3, 1, 2), (3, 0, 2),
(3, 2, 3), (3, 3, 4), (3, 4, 5))
MUTATE_DNA = direction_dna.spread_convex_dendrites(1, PARENT_DNA)
print("MUTATED DNA: ", MUTATE_DNA)
transform_compose = augSettings.generateTransformCompose(
list_transform, False)
dataGen = GeneratorFromCIFAR.GeneratorFromCIFAR(
2,
64,
threads=0,
dataAugmentation=True,
transforms_mode=transform_compose)
dataGen.dataConv2d()
version = directions_version.POOL_VERSION
mutation_manager = MutationManager.MutationManager(
directions_version=version)
parent_network = nw_dendrites.Network(dna=PARENT_DNA,
cuda_flag=True,
momentum=0.9,
weight_decay=0,
enable_activation=True,
enable_track_stats=True,
dropout_value=0.2,
dropout_function=None,
version=version)
parent_network.training_cosine_dt(dataGenerator=dataGen,
max_dt=0.001,
min_dt=0.001,
epochs=1,
restart_dt=1)
parent_network.generate_accuracy(dataGen)
print("original acc: ", parent_network.get_accuracy())
mutate_network = mutation_manager.execute_mutation(parent_network,
MUTATE_DNA)
mutate_network.generate_accuracy(dataGen)
print("mutated acc: ", mutate_network.get_accuracy())
mutate_network.training_cosine_dt(dataGenerator=dataGen,
max_dt=0.001,
min_dt=0.001,
epochs=1,
restart_dt=1)
mutate_network.generate_accuracy(dataGen)
print("mutated acc after training: ", mutate_network.get_accuracy())
def Test_Convex():
augSettings = AugmentationSettings.AugmentationSettings()
list_transform = {
augSettings.randomHorizontalFlip: True,
augSettings.translate: True,
}
version = directions_version.CONVEX_VERSION
mutation_manager = MutationManager.MutationManager(
directions_version=version)
transform_compose = augSettings.generateTransformCompose(
list_transform, False)
dataGen = GeneratorFromCIFAR.GeneratorFromCIFAR(
2,
128,
threads=0,
dataAugmentation=True,
transforms_mode=transform_compose)
dataGen.dataConv2d()
DNA = ((-1, 1, 3, 32, 32), (0, 3, 32, 3, 3), (0, 32, 64, 3, 3, 2),
(0, 64, 128, 3, 3, 2), (0, 128, 256, 3, 3), (0, 256, 128, 2, 2),
(0, 128, 128, 3, 3, 2), (0, 256, 256, 3, 3), (0, 384, 256, 3, 3),
(0, 256, 128, 3, 3), (0, 384, 128, 3, 3), (0, 128, 128, 3, 3),
(0, 256, 256, 3, 3), (0, 256, 128, 8, 8), (1, 128, 10), (2, ),
(3, -1, 0), (3, 0, 1), (3, 1, 2), (3, 2, 3), (3, 3, 4), (3, 4, 5),
(3, 3, 6), (3, 5, 7), (3, 6, 7), (3, 7, 8), (3, 8, 9), (3, 3, 9),
(3, 4, 10), (3, 10, 11), (3, 9, 11), (3, 11, 12), (3, 13, 14))
MUTATE_DNA = ((-1, 1, 3, 32, 32), (0, 3, 32, 3, 3), (0, 32, 64, 3, 3, 2),
(0, 64, 128, 3, 3, 2), (0, 128, 256, 3, 3),
(0, 256, 128, 2, 2), (0, 128, 128, 3, 3, 2),
(0, 256, 256, 3, 3), (0, 128, 128, 3, 3),
(0, 384, 256, 3, 3), (0, 256, 128, 3, 3), (0, 384, 128, 3,
3),
(0, 128, 128, 3, 3), (0, 256, 256, 3, 3), (0, 256, 128, 8,
8), (1, 128, 10),
(2, ), (3, -1, 0), (3, 0, 1), (3, 1, 2), (3, 2, 3),
(3, 3, 4), (3, 4, 5), (3, 3, 6), (3, 5, 7), (3, 7, 8),
(3, 6, 8), (3, 8, 9), (3, 9, 10), (3, 3, 10), (3, 4, 11),
(3, 11, 12), (3, 10, 12), (3, 12, 13), (3, 14, 15))
((-1, 1, 3, 32, 32), (0, 3, 32, 3, 3), (0, 32, 64, 3, 3, 2), (0, 64, 128,
3, 3, 2),
(0, 128, 256, 3, 3), (0, 256, 256, 8, 8), (1, 128, 10), (2, ), (3, -1, 0),
(3, 0, 1), (3, 1, 2), (3, 2, 3), (3, 3, 4), (3, 5, 6))
parent_network = nw_dendrites.Network(dna=DNA,
cuda_flag=True,
momentum=0.9,
weight_decay=0,
enable_activation=True,
enable_track_stats=True,
dropout_value=0,
dropout_function=None,
version=version)
print("starting mutation")
mutate_network = mutation_manager.execute_mutation(parent_network,
MUTATE_DNA)