def __init__(self, num_dimensions, flow_config,
dataset_class=None,
vocab=None, vocab_size=-1,
use_decoder=False, decoder_config=None,
default_embed_layer_dims=64,
category_prior=None,
**kwargs):
super().__init__()
self.use_decoder = use_decoder
self.dataset_class = dataset_class
self.D = num_dimensions
self.embed_layer, self.vocab_size = create_embed_layer(vocab, vocab_size, default_embed_layer_dims)
self.num_categories = self.vocab_size
self.prior_distribution = LogisticDistribution(mu=0.0, sigma=1.0) # Prior distribution in encoding flows
self.flow_layers = _create_flows(num_dims=num_dimensions,
embed_dims=self.embed_layer.weight.shape[1],
config=flow_config)
# Create decoder if needed
if self.use_decoder:
self.decoder = create_decoder(num_categories=self.vocab_size,
num_dims=self.D,
config=decoder_config)
# Prior over the categories. If not given, a uniform prior is assumed
if category_prior is None:
category_prior = torch.zeros(self.vocab_size, dtype=torch.float32)
else:
assert category_prior.shape[
0] == self.num_categories, "[!] ERROR: Category prior needs to be of size [%i] but is %s" % (
self.num_categories, str(category_prior.shape))
if isinstance(category_prior, np.ndarray):
category_prior = torch.from_numpy(category_prior)
self.register_buffer("category_prior", F.log_softmax(category_prior, dim=-1))
def __init__(self,
num_dimensions,
flow_config,
dataset_class=None,
vocab=None,
vocab_size=-1,
use_decoder=False,
decoder_config=None,
default_embed_layer_dims=64,
category_prior=None,
**kwargs):
super().__init__()
self.use_decoder = use_decoder
self.dataset_class = dataset_class
self.D = num_dimensions
self.embed_layer, self.vocab_size = create_embed_layer(
vocab, vocab_size, default_embed_layer_dims)
self.num_categories = self.vocab_size
self.prior_distribution = LogisticDistribution(
mu=0.0, sigma=1.0) # Prior distribution in encoding flows
self.flow_layers = _create_flows(
num_dims=num_dimensions,
embed_dims=self.embed_layer.weight.shape[1],
config=flow_config)
# Create decoder if needed
self.decoder = create_decoder(num_categories=self.vocab_size,
num_dims=self.D,
config=decoder_config)