def to_descriptor(self, dimensions=2):
self.active_nodes = 0
descriptor = NeuralDescriptor()
actives = set()
self.actives = set()
# Get only active nodes
for p in self.get_direct_paths():
for n in p:
actives.add(n)
# First add the nodes themselves
for g in self.nodes.genes:
gene = self.nodes.genes[g]
# Don't add inactive nodes
if gene.innovation_number in actives and gene.enabled:
if not gene.io:
self.active_nodes += 1
self.actives.add(str(gene.value))
# Get the node's name (innovation number)
innv = str(gene.innovation_number)
# Get the parameters
parameters = gene.value
filter_no = int(parameters[0])
dropout_rate = float(parameters[1])
weight_scale = float(parameters[2])
kernel_size = int(parameters[3])
max_pool = True if int(parameters[4]) == 1 else False
out_channels = filter_no
# --Define the layers and parameters--
# Convolution layer
conv_layer = nn.Conv2d
if dimensions == 1:
conv_layer = nn.Conv1d
conv_parameters = {
'in_channels': 1000,
'out_channels': out_channels,
'kernel_size': kernel_size
}
descriptor.add_layer(conv_layer,
conv_parameters,
name=innv + 'in')
# Scale the weights
descriptor.add_layer_sequential(ScaleLayer,
{'scale': weight_scale},
name=innv + 'scale')
# Dropout layer
if dimensions == 2:
dout = nn.Dropout2d
else:
dout = nn.Dropout
dout_parameters = {'p': dropout_rate}
descriptor.add_layer_sequential(dout,
dout_parameters,
name=innv + 'dout')
# Max pool layer
if max_pool:
pool = nn.MaxPool2d
if dimensions == 1:
pool = nn.MaxPool1d
pool_parameters = {
'kernel_size': kernel_size,
'stride': kernel_size
}
descriptor.add_layer_sequential(pool,
pool_parameters,
name=innv + 'pool')
# Activation layer
descriptor.add_layer_sequential(nn.ReLU6, {},
name=innv + 'out')
# Add IO layers
descriptor.add_layer(Identity, {}, name='-2out')
descriptor.add_layer(Identity, {}, name='-1in')
descriptor.first_layer = '-2out'
descriptor.last_layer = '-1in'
# Connect the layers
for g in self.connections.genes:
gene = self.connections.genes[g]
from_, to_ = gene.value
# Connect all active
if gene.enabled:
# Only connecitons from/to active nodes should be added
if from_ in actives and to_ in actives:
last_out = str(from_) + 'out'
descriptor.connect_layers(last_out, str(to_) + 'in')
return descriptor
# Instantiate the evaluator
# define the optimizer CLASS and
# any parameters you wish to utilize
# with the optimizer
evaluator = LocalEvaluator(optimizer_class=opt.Adam,
optimizer_params={}, verbose=True)
# Select dataset from:
# - cifar10 (requires 2d conv and pool, i.e. conv = nn.Conv2d, pool = nn.MaxPool2d)
# - fashion-mnist (requires 2d conv and pool, i.e. conv = nn.Conv2d, pool = nn.MaxPool2d)
# - activity_recognition (requires 1d conv and pool, i.e. conv = nn.Conv1d, pool = nn.MaxPool1d)
dataset = 'activity_recognition'
# Instantiate a descriptor
d = NeuralDescriptor()
conf = Configs()
# Define layer presets
conv = nn.Conv1d
# Note that first layer has number of channels
# equal with the dataset
conv_params = {'in_channels': conf.CHANNELS[dataset],
'out_channels': 5, 'kernel_size': 3}
conv_2_params = {'in_channels': 5,
'out_channels': 10, 'kernel_size': 3}
pool = nn.MaxPool1d
pool_params = {'kernel_size': 2, 'stride': 2}
"""
Example usage of BenchmarkEvaluator,
which evaluates the given architecture in NASBench-101
"""
from nord.neural_nets import BenchmarkEvaluator, NeuralDescriptor
# Instantiate the evaluator
evaluator = BenchmarkEvaluator()
# Instantiate a descriptor
d = NeuralDescriptor()
# See the available layers (ops)
# for NASBench-101
layers = evaluator.get_available_ops()
print(layers)
# Add NASBench-101 Layers connected
# sequentially
d.add_layer('input', None, 'in')
d.add_layer_sequential(layers[0], None, 'layer_1')
d.add_layer_sequential(layers[2], None, 'layer_2')
d.add_layer_sequential('output', None, 'out')
# Add Connections
d.connect_layers('layer_1', 'out')
# Get the validation accuracy and training time
val_acc, train_time = evaluator.descriptor_evaluate(d,
acc='validation_accuracy')
def to_descriptor(self, dimensions=2):
assert dimensions <= 2
self.active_nodes = 0
descriptor = NeuralDescriptor()
actives = set()
self.actives = set()
# Get only active nodes
for p in self.get_direct_paths():
for n in p:
actives.add(n)
# First add the nodes themselves
for g in self.nodes.genes:
gene = self.nodes.genes[g]
# Don't add inactive nodes
if gene.innovation_number in actives and gene.enabled:
if not gene.io:
self.active_nodes += 1
self.actives.add(str(gene.value))
# Get the node's name (innovation number)
innv = str(gene.innovation_number)
# Get the parameters
selected_layer, params = layers_list[gene.value[0]], None
if '.' in selected_layer:
selected_layer, params = selected_layer.split('.')
layer, kernel = selected_layer.split('_')
channels = self.channels
kernel = int(kernel)
conv = False
if layer == 'CONV':
conv = True
if params == 'H':
channels = int(channels/2)
layer = nn.Conv2d
parameters = {'in_channels': 1000,
'out_channels': channels,
'kernel_size': kernel,
'stride': self.strides}
if kernel == 151:
layer = Conv2d151
parameters = {'in_channels': 1000,
'out_channels': channels}
elif layer == 'POOL':
layer = nn.MaxPool2d
if params == 'A':
layer = nn.AvgPool2d
parameters = {'kernel_size': kernel,
'stride': kernel}
descriptor.add_layer(layer, parameters, name=innv+'in')
# Activation layer
if conv:
descriptor.add_layer_sequential(
nn.ReLU6, {}, name=innv+'relu')
descriptor.add_layer_sequential(
nn.BatchNorm2d, {'num_features': channels},
name=innv+'batchnorm')
descriptor.add_layer_sequential(
nn.Dropout, {'p': DROPOUT_PROB}, name=innv+'dropout')
descriptor.add_layer_sequential(
Identity, {}, name=innv+'out')
# Add IO layers
descriptor.add_layer(Identity, {}, name=INPUT_NAME)
descriptor.add_layer(Identity, {}, name=OUTPUT_NAME)
descriptor.first_layer = INPUT_NAME
descriptor.last_layer = OUTPUT_NAME
# Connect the layers
for g in self.connections.genes:
gene = self.connections.genes[g]
from_, to_ = gene.value
# Connect all active
if gene.enabled:
# Only connecitons from/to active nodes should be added
if from_ in actives and to_ in actives:
last_out = str(from_)+'out'
to_layer = str(to_)+'in'
if from_ == INPUT:
last_out = INPUT_NAME
elif from_ == OUTPUT:
last_out = OUTPUT_NAME
if to_ == INPUT:
to_layer = INPUT_NAME
elif to_ == OUTPUT:
to_layer = OUTPUT_NAME
descriptor.connect_layers(last_out, to_layer)
return descriptor
def to_descriptor(self, modules_list, dimensions=2):
assert dimensions == 2
self.active_nodes = 0
descriptor = NeuralDescriptor()
actives = set()
self.actives = set()
# Get only active nodes
for p in self.get_direct_paths():
for n in p:
actives.add(n)
# First add the nodes themselves
for g in self.nodes.genes:
gene = self.nodes.genes[g]
# Don't add inactive nodes
if gene.innovation_number in actives and gene.enabled:
if not gene.io:
self.active_nodes += 1
self.actives.add(str(gene.value))
# Get the node's name (innovation number)
innv = str(gene.innovation_number)
# Get the parameters
selected_module = gene.value
module = modules_list[selected_module]
module_descriptor = module.to_descriptor()
module_descriptor.add_suffix('_' + innv)
descriptor.layers.update(module_descriptor.layers)
descriptor.incoming_connections.update(
module_descriptor.incoming_connections)
descriptor.connections.update(
module_descriptor.connections)
# Add IO layers
descriptor.add_layer(Identity, {}, name=INPUT_NAME)
descriptor.add_layer(Identity, {}, name=OUTPUT_NAME)
descriptor.first_layer = INPUT_NAME
descriptor.last_layer = OUTPUT_NAME
# Connect the layers
for g in self.connections.genes:
gene = self.connections.genes[g]
from_, to_ = gene.value
# Connect all active
if gene.enabled:
# Only connecitons from/to active nodes should be added
if from_ in actives and to_ in actives:
from_name = OUTPUT_NAME + '_' + str(from_)
to_name = INPUT_NAME + '_' + str(to_)
if from_ == INPUT:
from_name = INPUT_NAME
elif from_ == OUTPUT:
from_name = OUTPUT_NAME
if to_ == INPUT:
to_name = INPUT_NAME
elif to_ == OUTPUT:
to_name = OUTPUT_NAME
descriptor.connect_layers(from_name, to_name)
return descriptor
from nord.neural_nets import NeuralDescriptor, NeuralNet
# Define layer presets
conv = nn.Conv2d
conv_params = {'in_channels': 3, 'out_channels': 5, 'kernel_size': 3}
conv_2_params = {'in_channels': 5, 'out_channels': 10, 'kernel_size': 5}
pool = nn.MaxPool2d
pool_params = {'kernel_size': 2, 'stride': 2}
pool2_params = {'kernel_size': 2, 'stride': 5}
# # Sequential Example
# # Instantiate a descriptor
d = NeuralDescriptor()
# Add layers sequentially
d.add_layer_sequential(conv, conv_params)
d.add_layer_sequential(conv, conv_2_params)
d.add_layer_sequential(pool, pool_params)
d.add_layer_sequential(pool, pool2_params)
# Instantiate the network
net = NeuralNet(net_descriptor=d,
num_classes=10,
input_shape=(32, 32),
input_channels=3)
# Print it
print(net)
# Plot it