def __init__(self,
input,
params=None,
rng=np.random.RandomState(),
zsize=100):
self.input = input
h_input = input
h = FullyConnected(input=h_input,
n_in=zsize,
n_out=4 * 4 * 1024,
W=params[0] if params is not None else None,
b=params[1] if params is not None else None,
rng=rng)
h_out = relu(batchnorm(h.output.reshape((input.shape[0], 1024, 4, 4))))
conv1 = ConvLayer(h_out,
4,
8,
1024,
512,
rng=rng,
W=params[2] if params is not None else None)
conv1_out = relu(batchnorm(conv1.output))
conv2 = ConvLayer(conv1_out,
8,
16,
512,
256,
rng=rng,
W=params[3] if params is not None else None)
conv2_out = relu(batchnorm(conv2.output))
conv3 = ConvLayer(conv2_out,
16,
32,
256,
128,
rng=rng,
W=params[4] if params is not None else None)
conv3_out = relu(batchnorm(conv3.output))
conv4 = ConvLayer(conv3_out,
32,
64,
128,
3,
rng=rng,
W=params[5] if params is not None else None)
conv4_out = T.tanh(conv4.output)
self.output = conv4_out
self.params = h.params + conv1.params + conv2.params + \
conv3.params + conv4.params
def __init__(self, input, params=None, rng=np.random.RandomState()):
self.input = input
conv1 = ConvLayer(input,
64,
32,
3,
128,
rng=rng,
W=params[0] if params is not None else None)
conv1_out = lrelu(conv1.output)
conv2 = ConvLayer(conv1_out,
32,
16,
128,
256,
rng=rng,
W=params[1] if params is not None else None)
conv2_out = lrelu(batchnorm(conv2.output))
conv3 = ConvLayer(conv2_out,
16,
8,
256,
512,
rng=rng,
W=params[2] if params is not None else None)
conv3_out = lrelu(batchnorm(conv3.output))
conv4 = ConvLayer(conv3_out,
8,
4,
512,
1024,
rng=rng,
W=params[3] if params is not None else None)
conv4_out = lrelu(batchnorm(conv4.output))
h_input = conv4_out.flatten(2)
h = FullyConnected(input=h_input,
n_in=1024 * 4 * 4,
n_out=1,
W=params[4] if params is not None else None,
b=params[5] if params is not None else None,
rng=rng)
h_out = h.output
self.output = h.output
self.params = conv1.params + conv2.params + \
conv3.params + conv4.params + h.params