def get_toy_model():
from learning.models.rws import LayerStack
from learning.models.sbn import SBN, SBNTop
p_layers = [SBN(n_X=25, n_Y=10), SBNTop(n_X=10, )]
q_layers = [SBN(
n_X=10,
n_Y=25,
)]
model = LayerStack(
p_layers=p_layers,
q_layers=q_layers,
)
return model
def test_FromModel():
D = 5
n_vis = D**2
n_hid = 2 * D
# Ground truth params
W_bars = np.zeros([n_hid, D, D])
for d in xrange(D):
W_bars[d, d, :] = 4.
W_bars[D + d, :, d] = 4.
W_bars = W_bars.reshape((n_hid, n_vis))
P_a = -np.log(D / 2 - 1) * np.ones(n_hid)
P_b = -2 * np.ones(n_vis)
# Instantiate model...
p_layers = [SBN(
n_X=n_vis,
n_Y=n_hid,
), SBNTop(n_X=n_hid)]
q_layers = [SBN(
n_X=n_hid,
n_Y=n_vis,
)]
p_layers[0].set_model_param('W', W_bars)
p_layers[0].set_model_param('b', P_b)
p_layers[1].set_model_param('a', P_a)
model = LayerStack(p_layers=p_layers, q_layers=q_layers)
# ...and generate data
n_datapoints = 1000
data = FromModel(model=model, n_datapoints=n_datapoints)
assert data.X.shape == (n_datapoints, n_vis)
yield check_dtype, data
yield check_range, data
n_X=200,
n_hid=[],
final_tanh=True,
log_sigma2_min=log_sigma2_min,
),
DiagonalGaussian(
n_X=100,
n_Y=200,
n_hid=[],
final_tanh=False,
log_sigma2_min=log_sigma2_min,
),
]
model = LayerStack(
p_layers=p_layers,
q_layers=q_layers,
)
trainer = Trainer(
n_samples=10,
learning_rate_p=1e-6,
#learning_rate_q=1e-6,
learning_rate_q=0.0,
#learning_rate_s=1e-6,
learning_rate_s=1e-6,
weight_decay=0.0,
batch_size=25,
dataset=trainset,
model=model,
termination=EarlyStopping(lookahead=10),
epoch_monitors=[