def test_mean_field_matches_inpainting(cls):
# Tests that running mean field to infer Y given X
# gives the same result as running inpainting with
# Y masked out and all of X observed
# This test uses the MoreConsistent InferenceProcedure
# For SuperWeightDoubling it will fail because mf initializes Y_hat
# with double weights and do_inpainting initializes it with sigmoid(biases)
batch_size = 5
niter = 3
nvis = 4
nhid1 = 5
nhid2 = 10
classes = 3
rng = np.random.RandomState([2012, 11, 3])
vis = BinaryVector(nvis)
vis.set_biases(rng.randn(nvis).astype(floatX))
h1 = BinaryVectorMaxPool(detector_layer_dim=nhid1,
pool_size=1,
layer_name='h1',
irange=1.)
h2 = BinaryVectorMaxPool(detector_layer_dim=nhid2,
pool_size=1,
layer_name='h2',
irange=1.)
y = Softmax(n_classes=classes, irange=1., layer_name='y')
dbm = SuperDBM(batch_size=batch_size,
inference_procedure=cls(),
niter=niter,
visible_layer=vis,
hidden_layers=[h1, h2, y])
X = sharedX(rng.randn(batch_size, nvis))
Y = sharedX(np.zeros((batch_size, classes)))
drop_mask = sharedX(np.zeros((batch_size, nvis)))
drop_mask_Y = sharedX(np.ones((batch_size, )))
q_mf = dbm.mf(X)
Y_hat_mf = q_mf[-1]
V_hat_inpaint, Y_hat_inpaint = dbm.do_inpainting(V=X,
Y=Y,
drop_mask=drop_mask,
drop_mask_Y=drop_mask_Y)
Y_hat_mf, Y_hat_inpaint = function([], [Y_hat_mf, Y_hat_inpaint])()
if not np.allclose(Y_hat_mf, Y_hat_inpaint):
print Y_hat_mf
print Y_hat_inpaint
assert False, cls
def make_super_dbm(n_iter):
return SuperDBM(batch_size = 100,
visible_layer = GaussianConvolutionalVisLayer(
rows = 32,
cols = 32,
channels = 3,
init_beta = 1.,
init_mu = 0.),
hidden_layers = [
ConvMaxPool(output_channels = 256,
kernel_rows = 6,
kernel_cols = 6,
pool_rows = 3,
pool_cols = 3,
irange = .05,
layer_name = 'h0',
init_bias = -3.),
ConvMaxPool(output_channels = 128,
kernel_rows = 4,
kernel_cols = 4,
pool_rows = 3,
pool_cols = 3,
irange = .01,
layer_name = 'h1',
init_bias = -3.),
#ConvMaxPool(output_channels = 3200,
# kernel_rows = 5,
# kernel_cols = 5,
# pool_rows = 1, #really inefficient way of getting a non-pooled sigmoid layer
# pool_cols = 1,
# irange = .01,
# layer_name = 'h2',
# init_bias = -1.),
#ConvMaxPool(output_channels = 1600,
# kernel_rows = 3, #really inefficient way of getting a densely connected sigmoid layer
# kernel_cols = 3,
# pool_rows = 1,
# pool_cols = 1,
# irange = .01,
# layer_name = 'h3',
# init_bias = -1.),
Softmax(n_classes = 10,
irange = .01)
],
niter = n_iter)