# -*- coding: utf-8 -*-
import numpy as np
from pygan.true_sampler import TrueSampler
from pygan.discriminative_model import DiscriminativeModel
from pydbm.loss.mean_squared_error import MeanSquaredError
from pydbm.loss.interface.computable_loss import ComputableLoss
[docs]class FeatureMatching(object):
'''
Value function with Feature matching, which addresses the instability of GANs
by specifying a new objective for the generator that prevents it from overtraining
on the current discriminator(Salimans, T., et al., 2016).
"Instead of directly maximizing the output of the discriminator,
the new objective requires the generator to generate data that matches
the statistics of the real data, where we use the discriminator only to specify
the statistics that we think are worth matching." (Salimans, T., et al., 2016, p2.)
References:
- Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., & Chen, X. (2016). Improved techniques for training gans. In Advances in neural information processing systems (pp. 2234-2242).
- Yang, L. C., Chou, S. Y., & Yang, Y. H. (2017). MidiNet: A convolutional generative adversarial network for symbolic-domain music generation. arXiv preprint arXiv:1703.10847.
'''
def __init__(self, lambda1=1.0, lambda2=0.0, computable_loss=None):
'''
Init.
Args:
lambda1: Weight for results of standard feature matching.
lambda2: Weight for results of difference between generated data points and true samples.
computable_loss: is-a `pydbm.loss.interface.computable_loss.ComputableLoss`.
If `None`, the default value is a `MeanSquaredError`.
Exceptions:
ValueError: When the sum of `lambda1` and `lambda2` is not less than `1.0`.
Those parameters are trade-off parameters.
'''
self.__true_arr = None
if computable_loss is None:
self.__computable_loss = MeanSquaredError()
else:
if isinstance(computable_loss, ComputableLoss) is False:
raise TypeError("The type of `computable_loss` must be `ComputableLoss`.")
self.__computable_loss = computable_loss
self.__loss_list = []
if lambda1 + lambda2 > 1:
raise ValueError("The sum of `lambda1` and `lambda2` must be less than `1.0`. Those parameters are trade-off parameters.")
self.__lambda1 = lambda1
self.__lambda2 = lambda2
[docs] def compute_delta(
self,
true_sampler,
discriminative_model,
generated_arr
):
'''
Compute generator's reward.
Args:
true_sampler: Sampler which draws samples from the `true` distribution.
discriminative_model: Discriminator which discriminates `true` from `fake`.
generated_arr: `np.ndarray` generated data points.
Returns:
`np.ndarray` of Gradients.
'''
if isinstance(true_sampler, TrueSampler) is False:
raise TypeError("The type of `true_sampler` must be `TrueSampler`.")
if isinstance(discriminative_model, DiscriminativeModel) is False:
raise TypeError("The type of `discriminative_model` must be `DiscriminativeModel`.")
self.__true_arr = true_sampler.draw()
grad_arr1 = np.zeros_like(generated_arr)
grad_arr2 = np.zeros_like(generated_arr)
loss1, loss2 = 0.0, 0.0
if self.__lambda1 > 0.0:
_generated_arr = discriminative_model.feature_matching_forward(generated_arr)
_true_arr = discriminative_model.feature_matching_forward(self.__true_arr)
grad_arr1 = self.__computable_loss.compute_delta(
_generated_arr,
_true_arr
)
grad_arr1 = discriminative_model.feature_matching_backward(grad_arr1)
grad_arr1 = grad_arr1.reshape(generated_arr.shape)
loss1 = self.__computable_loss.compute_loss(
_generated_arr,
_true_arr
)
if self.__lambda2 > 0.0:
grad_arr2 = self.__computable_loss.compute_delta(
generated_arr,
self.__true_arr
)
loss2 = self.__computable_loss.compute_loss(
generated_arr,
self.__true_arr
)
grad_arr = (grad_arr1 * self.__lambda1) + (grad_arr2 * self.__lambda2)
loss = (loss1 * self.__lambda1) + (loss2 * self.__lambda2)
self.__loss_list.append(loss)
return grad_arr
[docs] def set_readonly(self, value):
''' setter '''
raise TypeError("This property must be read-only.")
[docs] def get_true_arr(self):
''' getter '''
return self.__true_arr
true_arr = property(get_true_arr, set_readonly)
[docs] def get_computable_loss(self):
''' getter '''
return self.__computable_loss
computable_loss = property(get_computable_loss, set_readonly)
[docs] def get_loss_list(self):
''' getter '''
return self.__loss_list
[docs] def set_loss_list(self, value):
''' setter '''
return self.__loss_list
_loss_list = property(get_loss_list, set_loss_list)
[docs] def get_loss_arr(self):
''' getter '''
return np.array(self.__loss_list)
loss_arr = property(get_loss_arr, set_readonly)