def __init__(self, image_shape, hidden_channels, K, L, actnorm_scale,
flow_permutation, flow_coupling, LU_decomposed, y_classes,
learn_top, y_condition):
super().__init__()
self.flow = FlowNet(image_shape=image_shape,
hidden_channels=hidden_channels,
K=K,
L=L,
actnorm_scale=actnorm_scale,
flow_permutation=flow_permutation,
flow_coupling=flow_coupling,
LU_decomposed=LU_decomposed)
self.y_classes = y_classes
self.y_condition = y_condition
self.learn_top = learn_top
# learned prior
if learn_top:
C = self.flow.output_shapes[-1][1]
self.learn_top_fn = Conv2dZeros(C * 2, C * 2)
if y_condition:
C = self.flow.output_shapes[-1][1]
self.project_ycond = LinearZeros(y_classes, 2 * C)
self.project_class = LinearZeros(C, y_classes)
self.register_buffer(
"prior_h",
torch.zeros([
1, self.flow.output_shapes[-1][1] * 2,
self.flow.output_shapes[-1][2], self.flow.output_shapes[-1][3]
]))
def __init__(self, image_shape, hidden_channels, K, L, actnorm_scale,
flow_permutation, flow_coupling, LU_decomposed, y_classes,
learn_top, y_condition, extra_condition, sp_condition, d_condition, yd_condition):
super().__init__()
self.flow = FlowNet(image_shape=image_shape,
hidden_channels=hidden_channels,
K=K,
L=L,
actnorm_scale=actnorm_scale,
flow_permutation=flow_permutation,
flow_coupling=flow_coupling,
LU_decomposed=LU_decomposed,
extra_condition=extra_condition,
sp_condition=sp_condition,
num_classes=y_classes)
self.y_classes = y_classes
if y_condition or d_condition or yd_condition:
self.y_condition = True
print("conditional version", self.y_condition)
else:
self.y_condition=False
self.learn_top = learn_top
print("extra condtion", extra_condition)
print("split prior condition", sp_condition)
# learned prior
if learn_top:
C = self.flow.output_shapes[-1][1]
self.learn_top_fn = Conv2dZeros(C * 2, C * 2)
if self.y_condition:
C = self.flow.output_shapes[-1][1]
print("prior", 2*C)
self.project_ycond = LinearZeros(y_classes, 2 * C)
self.project_class = LinearZeros(C, y_classes)
self.register_buffer("prior_h",
torch.zeros([1,
self.flow.output_shapes[-1][1] * 2,
self.flow.output_shapes[-1][2],
self.flow.output_shapes[-1][3]]))
def __init__(self, image_shape, hidden_channels, K, L, actnorm_scale,
flow_permutation, flow_coupling, LU_decomposed,
flow_embed_dim, y_classes, learn_top, y_condition):
super().__init__()
self.flow = FlowNet(image_shape=image_shape,
hidden_channels=hidden_channels,
K=K,
L=L,
actnorm_scale=actnorm_scale,
flow_permutation=flow_permutation,
flow_coupling=flow_coupling,
LU_decomposed=LU_decomposed,
flow_embed_dim=flow_embed_dim)
self.y_classes = y_classes
self.y_condition = y_condition
self.learn_top = learn_top
# learned prior
if learn_top:
C = self.flow.output_shapes[-1][1]
self.learn_top_fn = Conv2dZeros(C * 2, C * 2)
if y_condition:
C = self.flow.output_shapes[-1][1]
self.project_ycond = LinearZeros(y_classes, 2 * C)
self.project_class = LinearZeros(C, y_classes)
self.register_buffer(
"prior_h",
torch.zeros([
1, self.flow.output_shapes[-1][1] * 2,
self.flow.output_shapes[-1][2], self.flow.output_shapes[-1][3]
]))
self.num_param = sum(p.numel() for p in self.parameters()
if p.requires_grad)
print("num_param: {}".format(self.num_param))