Source code for pygan.discriminativemodel.nn_model
# -*- coding: utf-8 -*-
import numpy as np
from logging import getLogger
from pygan.discriminative_model import DiscriminativeModel
from pydbm.nn.neural_network import NeuralNetwork
from pydbm.nn.nn_layer import NNLayer
from pydbm.optimization.opt_params import OptParams
from pydbm.verification.interface.verificatable_result import VerificatableResult
from pydbm.loss.interface.computable_loss import ComputableLoss
from pydbm.cnn.layerablecnn.convolution_layer import ConvolutionLayer
from pydbm.synapse.nn_graph import NNGraph
# Loss function.
from pydbm.loss.mean_squared_error import MeanSquaredError
# Adam as a optimizer.
from pydbm.optimization.optparams.adam import Adam
# Verification.
from pydbm.verification.verificate_function_approximation import VerificateFunctionApproximation
[docs]class NNModel(DiscriminativeModel):
'''
Neural Network as a Discriminator.
'''
def __init__(
self,
batch_size,
nn_layer_list,
learning_rate=1e-05,
learning_attenuate_rate=0.1,
attenuate_epoch=50,
computable_loss=None,
opt_params=None,
verificatable_result=None,
nn=None,
feature_matching_layer=0
):
'''
Init.
Args:
batch_size: Batch size in mini-batch.
nn_layer_list: `list` of `NNLayer`.
learning_rate: Learning rate.
learning_attenuate_rate: Attenuate the `learning_rate` by a factor of this value every `attenuate_epoch`.
attenuate_epoch: Attenuate the `learning_rate` by a factor of `learning_attenuate_rate` every `attenuate_epoch`.
Additionally, in relation to regularization,
this class constrains weight matrixes every `attenuate_epoch`.
computable_loss: is-a `ComputableLoss`.
This parameters will be refered only when `nn` is `None`.
opt_params: is-a `OptParams`.
This parameters will be refered only when `nn` is `None`.
verificatable_result: is-a `VerificateFunctionApproximation`.
This parameters will be refered only when `nn` is `None`.
nn: is-a `NeuralNetwork` as a model in this class.
If not `None`, `self.__nn` will be overrided by this `nn`.
If `None`, this class initialize `NeuralNetwork`
by default hyper parameters.
feature_matching_layer: Key of layer number for feature matching forward/backward.
'''
if nn is None:
if computable_loss is None:
computable_loss = MeanSquaredError()
if isinstance(computable_loss, ComputableLoss) is False:
raise TypeError()
if verificatable_result is None:
verificatable_result = VerificateFunctionApproximation()
if isinstance(verificatable_result, VerificatableResult) is False:
raise TypeError()
if opt_params is None:
opt_params = Adam()
opt_params.weight_limit = 1e+10
opt_params.dropout_rate = 0.0
if isinstance(opt_params, OptParams) is False:
raise TypeError()
nn = NeuralNetwork(
# The `list` of `ConvolutionLayer`.
nn_layer_list=nn_layer_list,
# The number of epochs in mini-batch training.
epochs=200,
# The batch size.
batch_size=batch_size,
# Learning rate.
learning_rate=learning_rate,
# Loss function.
computable_loss=computable_loss,
# Optimizer.
opt_params=opt_params,
# Verification.
verificatable_result=verificatable_result,
# Pre-learned parameters.
pre_learned_path_list=None,
# Others.
learning_attenuate_rate=learning_attenuate_rate,
attenuate_epoch=attenuate_epoch
)
self.__nn = nn
self.__batch_size = batch_size
self.__learning_rate = learning_rate
self.__attenuate_epoch = attenuate_epoch
self.__learning_attenuate_rate = learning_attenuate_rate
self.__q_shape = None
self.__loss_list = []
self.__feature_matching_layer = feature_matching_layer
self.__epoch_counter = 0
logger = getLogger("pygan")
self.__logger = logger
[docs] def inference(self, observed_arr):
'''
Draws samples from the `true` distribution.
Args:
observed_arr: `np.ndarray` of observed data points.
Returns:
`np.ndarray` of inferenced.
'''
if observed_arr.ndim != 2:
observed_arr = observed_arr.reshape((observed_arr.shape[0], -1))
return self.__nn.inference(observed_arr)
[docs] def learn(self, grad_arr, fix_opt_flag=False):
'''
Update this Discriminator by ascending its stochastic gradient.
Args:
grad_arr: `np.ndarray` of gradients.
fix_opt_flag: If `False`, no optimization in this model will be done.
Returns:
`np.ndarray` of delta or gradients.
'''
if grad_arr.ndim != 2:
grad_arr = grad_arr.reshape((grad_arr.shape[0], -1))
delta_arr = self.__nn.back_propagation(grad_arr)
if fix_opt_flag is False:
if ((self.__epoch_counter + 1) % self.__attenuate_epoch == 0):
self.__learning_rate = self.__learning_rate * self.__learning_attenuate_rate
self.__nn.optimize(self.__learning_rate, self.__epoch_counter)
self.__epoch_counter += 1
return delta_arr
[docs] def feature_matching_forward(self, observed_arr):
'''
Forward propagation in only first or intermediate layer
for so-called Feature matching.
Args:
observed_arr: `np.ndarray` of observed data points.
Returns:
`np.ndarray` of outputs.
'''
if observed_arr.ndim != 2:
observed_arr = observed_arr.reshape((observed_arr.shape[0], -1))
if self.__feature_matching_layer == 0:
return self.__nn.nn_layer_list[0].forward_propagate(observed_arr)
else:
for i in range(self.__feature_matching_layer):
observed_arr = self.__nn.nn_layer_list[i].forward_propagate(observed_arr)
return observed_arr
[docs] def feature_matching_backward(self, grad_arr):
'''
Back propagation in only first or intermediate layer
for so-called Feature matching.
Args:
observed_arr: `np.ndarray` of observed data points.
Returns:
`np.ndarray` of outputs.
'''
if grad_arr.ndim != 2:
grad_arr = grad_arr.reshape((grad_arr.shape[0], -1))
if self.__feature_matching_layer == 0:
return np.dot(grad_arr, self.__nn.nn_layer_list[0].graph.weight_arr.T)
else:
nn_layer_list = self.__nn.nn_layer_list[:self.__feature_matching_layer][::-1]
for i in range(len(nn_layer_list)):
grad_arr = np.dot(grad_arr, nn_layer_list[i].graph.weight_arr.T)
return grad_arr
[docs] def get_nn(self):
''' getter '''
return self.__nn
[docs] def set_nn(self, value):
''' setter '''
raise TypeError("This property must be read-only.")
nn = property(get_nn, set_nn)