def __init__(self, name, hparams, optimizer='RMS'):
self.name = name
self.hparams = hparams
self.latent_dim = self.hparams.layer_sizes[-1]
# Gaussian prior for each beta_i
self._lambda_prior = self.hparams.lambda_prior
self.mu = [
np.zeros(self.latent_dim) for _ in range(self.hparams.num_actions)
]
self.cov = [(1.0 / self.lambda_prior) * np.eye(self.latent_dim)
for _ in range(self.hparams.num_actions)]
self.precision = [
self.lambda_prior * np.eye(self.latent_dim)
for _ in range(self.hparams.num_actions)
]
self.mu_prior_flag = self.hparams.mu_prior_flag
self.sigma_prior_flag = self.hparams.sigma_prior_flag
self.precision_prior = self.precision[:]
self.mu_prior = np.zeros((self.latent_dim, self.hparams.num_actions))
# Inverse Gamma prior for each sigma2_i
self._a0 = self.hparams.a0
self._b0 = self.hparams.b0
self.a = [self._a0 for _ in range(self.hparams.num_actions)]
self.b = [self._b0 for _ in range(self.hparams.num_actions)]
# Regression and NN Update Frequency
self.update_freq_lr = hparams.training_freq
self.update_freq_nn = hparams.training_freq_network
self.t = 0
self.optimizer_n = optimizer
self.num_epochs = hparams.training_epochs
self.data_h = ContextualDataset(hparams.context_dim,
hparams.num_actions,
intercept=False,
buffer_s=hparams.mem)
self.latent_h = ContextualDataset(self.latent_dim,
hparams.num_actions,
intercept=False,
buffer_s=hparams.mem)
self.bnn = NeuralBanditModel(optimizer, hparams, '{}-bnn'.format(name))
def __init__(self, name, hparams, textflag='no', optimizer='RMS'):
self.first_train = False
self.name = name
self.hparams = hparams
self.latent_dim = self.hparams.layer_sizes[-1]
self.intercept = False
self.pgd_steps = self.hparams.pgd_steps
self.pgd_batch_size = self.hparams.pgd_batch_size
if self.intercept:
self.param_dim = 1 + self.latent_dim
else:
self.param_dim = self.latent_dim
self.EPSILON = 0.00001
# Gaussian prior for each beta_i
self._lambda_prior = self.hparams.lambda_prior
self.before = []
self.after = []
self.mu = [
np.zeros(self.param_dim) for _ in range(self.hparams.num_actions)
]
self.f = [
np.zeros(self.param_dim) for _ in range(self.hparams.num_actions)
]
self.yy = [0 for _ in range(self.hparams.num_actions)]
self.cov = [(1.0 / self.lambda_prior) * np.eye(self.param_dim)
for _ in range(self.hparams.num_actions)]
self.precision = [
self.lambda_prior * np.eye(self.param_dim)
for _ in range(self.hparams.num_actions)
]
self.mu_prior_flag = self.hparams.mu_prior_flag
self.sigma_prior_flag = self.hparams.sigma_prior_flag
self.precision_prior = [
self.lambda_prior * np.eye(self.param_dim)
for _ in range(self.hparams.num_actions)
]
self.mu_prior = np.zeros((self.param_dim, self.hparams.num_actions))
# Inverse Gamma prior for each sigma2_i
self._a0 = self.hparams.a0
self._b0 = self.hparams.b0
self.a = [self._a0 for _ in range(self.hparams.num_actions)]
self.b = [self._b0 for _ in range(self.hparams.num_actions)]
# Regression and NN Update Frequency
self.update_freq_lr = hparams.training_freq
self.update_freq_nn = hparams.training_freq_network
self.t = 0
self.optimizer_n = optimizer
self.num_epochs = hparams.training_epochs
self.data_h = ContextualDataset(hparams.context_dim,
hparams.num_actions,
intercept=False,
buffer_s=hparams.mem)
self.latent_h = ContextualDataset(self.latent_dim,
hparams.num_actions,
intercept=self.intercept,
buffer_s=hparams.mem)
if textflag == 'yes':
self.bnn = TextCNN('adam', self.hparams.num_actions,
self.hparams.batch_size, '{}-bnn'.format(name))
else:
self.bnn = NeuralBanditModel(optimizer, hparams,
'{}-bnn'.format(name))