def __init__(self,
n_links,
hidden_size=20,
prior_pi=1.0,
prior_sigma_1=1.0,
prior_sigma_2=0.01):
super(BRNN, self).__init__()
self.hidden_size = hidden_size
self.lstm_1 = BayesianLSTM(n_links,
hidden_size,
prior_pi=prior_pi,
prior_sigma_1=prior_sigma_1,
prior_sigma_2=prior_sigma_2,
posterior_rho_init=1.0,
peephole=False)
self.lstm_2 = BayesianLSTM(hidden_size,
hidden_size,
prior_pi=prior_pi,
prior_sigma_1=prior_sigma_1,
prior_sigma_2=prior_sigma_2,
posterior_rho_init=1.0,
peephole=False)
self.linear = BayesianLinear(hidden_size,
n_links,
prior_pi=prior_pi,
prior_sigma_1=prior_sigma_1,
prior_sigma_2=prior_sigma_2,
posterior_rho_init=3.0)
def test_sequential_cuda(self):
#check if we can create sequential models chaning our Bayesian Linear layers
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
b_lstm = BayesianLSTM(1, 10)
to_feed = torch.ones((10, 10, 1))
predicted = b_lstm(to_feed)
def test_brnn_sharpen_posterior_on_forward_lstm(self):
b_module = BayesianLSTM(3, 5, sharpen=True)
in_tensor = torch.ones(5, 4, 3)
out_tensor = b_module(in_tensor)[0][:, -1, :]
loss = nn.MSELoss()(out_tensor.clone().detach().normal_(), out_tensor)
b_module.forward(in_tensor, sharpen_loss=loss)
def __init__(self, n_features, n_hidden, seq_len, y_length):
super(B_LSTM, self).__init__()
self.n_hidden = n_hidden
self.seq_len = seq_len
self.y_len = y_length
self.lstm = BayesianLSTM(n_features, n_hidden)
self.linear = nn.Linear(in_features=n_hidden, out_features=y_length)
def __init__(self):
super().__init__()
self.blinear1 = BayesianLinear(10, 512)
self.bconv = BayesianConv2d(3,
3,
kernel_size=(3, 3),
padding=1,
bias=True)
self.blstm = BayesianLSTM(10, 2)
def test_frozen_inference(self):
b_lstm = BayesianLSTM(1, 10)
b_lstm.freeze = True
in_data = torch.ones((10, 10, 1))
b_inference_1 = b_lstm(in_data, hidden_states=None)
b_inference_2 = b_lstm(in_data, hidden_states=None)
self.assertEqual((b_inference_1[0] == b_inference_2[0]).all(),
torch.tensor(True))
def test_infer_shape_1_sample(self):
deterministic_lstm = nn.LSTM(1, 10, 1, batch_first=True)
in_data = torch.ones((10, 10, 1))
det_inference = deterministic_lstm(in_data)
b_lstm = BayesianLSTM(1, 10)
b_inference = b_lstm(in_data, hidden_states=None)
self.assertEqual(det_inference[0].shape, b_inference[0].shape)
pass
def test_kl_divergence(self):
#create model, sample weights
#check if kl divergence between apriori and a posteriori is working
b_lstm = BayesianLSTM(1, 10)
to_feed = torch.ones((10, 10, 1))
predicted = b_lstm(to_feed)
complexity_cost = b_lstm.log_variational_posterior - b_lstm.log_prior
self.assertEqual((complexity_cost == complexity_cost).all(),
torch.tensor(True))
def test_variational_inference(self):
#create module, check if inference is variating
deterministic_lstm = nn.LSTM(1, 10, 1, batch_first=True)
in_data = torch.ones((10, 10, 1))
det_inference = deterministic_lstm(in_data)
b_lstm = BayesianLSTM(1, 10)
b_inference_1 = b_lstm(in_data, hidden_states=None)
b_inference_2 = b_lstm(in_data, hidden_states=None)
self.assertEqual((b_inference_1[0] != b_inference_2[0]).any(),
torch.tensor(True))
self.assertEqual((det_inference[0] == det_inference[0]).all(),
torch.tensor(True))
pass
def __init__(self, input_dim, hidden_dim, linear_dim, sequence_length,
output_dim):
'''
input_dim: input dimension (number of stops + dimension of temporal features)
hidden_dim: hidden dimension
linear_dim: linear layer dimension
sequence_length: sequence_length (default: 24*14 hours i.e. two weeks)
output_dim: output dimension (number of stops, default: 10)
'''
super(LSTM_Net, self).__init__()
# define network constants
self.input_dim = input_dim
self.hidden_dim = hidden_dim
self.linear_dim = linear_dim
self.sequence_length = sequence_length
self.output_dim = output_dim
# bayesian LSTM layer
self.lstm1 = BayesianLSTM(in_features=input_dim,
out_features=hidden_dim,
bias=True,
prior_sigma_1=1.0,
prior_sigma_2=4.0,
prior_pi=0.5,
posterior_mu_init=0,
posterior_rho_init=-0.5)
# bayesian linear layer
self.blinear1 = BayesianLinear(
in_features=hidden_dim,
out_features=linear_dim,
bias=True,
prior_sigma_1=1.0,
prior_sigma_2=0.5,
prior_pi=0.5,
posterior_mu_init=0,
posterior_rho_init=-0.5,
)
# linear layer
self.linear1 = nn.Linear(in_features=linear_dim,
out_features=output_dim,
bias=True)
# dropout function
self.dropout = nn.Dropout(p=0.5)
def test_init_bayesian_lstm(self):
b_lstm = BayesianLSTM(10, 10)
self.assertEqual(isinstance(b_lstm, (nn.Module)), True)
self.assertEqual(isinstance(b_lstm, (BayesianModule)), True)
pass
def test_peephole_inference(self):
b_lstm = BayesianLSTM(1, 10, peephole=True)
in_data = torch.ones((10, 10, 1))
b_lstm(in_data)
b_lstm.freeze = True
b_lstm(in_data)
def __init__(self):
super().__init__()
self.lstm1 = BayesianLSTM(3, 5, sharpen=True)
self.gru1 = BayesianGRU(5, 3, sharpen=True)