upsamp_X = T.nnet.abstract_conv.bilinear_upsampling(X, pool_factor)
Y = conv(upsamp_X)
return Y
return upsample
if __name__ == "__main__":
from theano_toolkit.parameters import Parameters
P = Parameters()
image_size = 10
rfield_size = 5
input_size = 1
X = T.as_tensor_variable(
np.random.randn(1, input_size, 4, 4).astype(np.float32))
P.W = conv_weight_init(2, 1, 5)
P.b = np.zeros(2)
W = P.W
b = P.b.dimshuffle('x', 0, 'x', 'x')
upsamp_X_1 = T.zeros(
(X.shape[0], X.shape[1], 2 * X.shape[2], 2 * X.shape[3]))
upsamp_X_1 = T.set_subtensor(upsamp_X_1[:, :, ::2, ::2], X)
upsamp_X_1 = T.inc_subtensor(upsamp_X_1[:, :, 1:-1:2, ::2],
0.5 * (X[:, :, :-1] + X[:, :, 1:]))
upsamp_X_1 = T.inc_subtensor(upsamp_X_1[:, :, ::2, 1:-1:2],
0.5 * (X[:, :, :, :-1] + X[:, :, :, 1:]))
upsamp_X_1 = T.inc_subtensor(
upsamp_X_1[:, :, 1:-1:2, 1:-1:2],
0.25 * (X[:, :, :-1, :-1] + X[:, :, 1:, :-1] + X[:, :, :-1, 1:] +
X[:, :, 1:, 1:]))
