def __init__(self, mlp_opt, cnn_opt, name="deconv"):
"""Constructor.
Args:
mlp_opt: Dictionary. Kwargs for vae_lib.MLP.
cnn_opt: Dictionary. Kwargs for vae_lib.ConvNet2D for the CNN.
name: Optional name.
"""
super().__init__(name=name)
assert cnn_opt["output_channels"][-1] is None, cnn_opt
if "activation" in cnn_opt:
cnn_opt["activation"] = get_act_func(cnn_opt["activation"])
self._cnn_opt = cnn_opt
if mlp_opt and "activation" in mlp_opt:
mlp_opt["activation"] = get_act_func(mlp_opt["activation"])
self._mlp_opt = mlp_opt
self._target_out_shape = None
def __init__(self, cnn_opt, mlp_opt, mode="flatten", name="cnn"):
"""Constructor.
Args:
cnn_opt: Dictionary. Kwargs for the cnn. See vae_lib.ConvNet2D for
details.
mlp_opt: Dictionary. Kwargs for the mlp. See vae_lib.MLP for details.
name: String. Optional name.
"""
super().__init__(name=name)
if "activation" in cnn_opt:
cnn_opt["activation"] = get_act_func(cnn_opt["activation"])
self._cnn_opt = cnn_opt
if "activation" in mlp_opt:
mlp_opt["activation"] = get_act_func(mlp_opt["activation"])
self._mlp_opt = mlp_opt
self._mode = mode
def __init__(
self,
dist=tfd.Normal,
dist_kwargs=None,
scale_act=tf.exp,
scale="stddev",
scale_val=1.0,
loc_act=None,
name="loc_scale_dist",
):
super().__init__(name=name)
self._scale_act = get_act_func(scale_act)
self._loc_act = get_act_func(loc_act)
# supports Normal, Logstic, Laplace, StudentT
self._dist = get_distribution(dist)
self._dist_kwargs = dist_kwargs or {}
assert scale in ["stddev", "var", "prec", "fixed"], scale
self._scale = scale
self._scale_val = scale_val
def __init__(
self,
hidden_sizes=(64, ),
pred_gate_bias=0.0,
corrector_gate_bias=0.0,
activation=tf.nn.elu,
name="predcorr_head",
):
super().__init__(name=name)
self._hidden_sizes = hidden_sizes
self._activation = utils.get_act_func(activation)
self._pred_gate_bias = pred_gate_bias
self._corrector_gate_bias = corrector_gate_bias
def __init__(
self,
cnn_opt,
mlp_opt=None,
coord_type="linear",
coord_freqs=3,
name="broadcast_conv",
):
"""Args:
cnn_opt: dict Kwargs for vae_lib.ConvNet2D for the CNN.
mlp_opt: None or dict If dictionary, then kwargs for snt.nets.MLP. If
None, then the model will not process the latent vector by an mlp.
coord_type: ["linear", "cos", None] type of coordinate channels to
add.
None: add no coordinate channels.
linear: two channels with values linearly spaced from -1. to 1. in
the H and W dimension respectively.
cos: coord_freqs^2 many channels containing cosine basis functions.
coord_freqs: int number of frequencies used to construct the cosine
basis functions (only for coord_type=="cos")
name: Optional name.
"""
super().__init__(name=name)
assert cnn_opt["output_channels"][-1] is None, cnn_opt
if "activation" in cnn_opt:
cnn_opt["activation"] = get_act_func(cnn_opt["activation"])
self._cnn_opt = cnn_opt
if mlp_opt and "activation" in mlp_opt:
mlp_opt["activation"] = get_act_func(mlp_opt["activation"])
self._mlp_opt = mlp_opt
self._target_out_shape = None
self._coord_type = coord_type
self._coord_freqs = coord_freqs
def __init__(
self,
num_components,
component_dist,
mask_activation=None,
name="masked_mixture",
):
"""
Spatial Mixture Model composed of a categorical masking distribution and
a custom pixel-wise component distribution (usually logistic or
gaussian).
Args:
num_components: int Number of mixture components >= 2
component_dist: the distribution to use for the individual components
mask_activation: str or function or None activation function that
should be applied to the mask before the softmax.
name: str
"""
super().__init__(name=name)
self._num_components = num_components
self._dist = component_dist
self._mask_activation = get_act_func(mask_activation)
def __init__(self, name="mlp", **mlp_opt):
super().__init__(name=name)
if "activation" in mlp_opt:
mlp_opt["activation"] = get_act_func(mlp_opt["activation"])
self._mlp_opt = mlp_opt
assert mlp_opt["output_sizes"][-1] is None, mlp_opt