def __init__(self, features, latent_dim, encoder, decoder, deterministic_path_drop_rate=0.5):
super().__init__()
self.embedder = nn.Embedding(MAX_VALUE, features)
self.encoder = ResidualAttentionEncoder(**encoder)
self._latent_encoder = nn.ModuleList([
ResidualAttentionEncoder(**encoder),
NormalNode(features, latent_dim)]
)
self.z_to_c = nn.Linear(latent_dim, latent_dim*155)
self.decoder = CondtionalResidualAttentionEncoder(**decoder)
self.logits = FeedForwardGELU(features, MAX_VALUE)
self.drop = nn.Dropout(deterministic_path_drop_rate)
def __init__(self,
features,
c_features,
attention_layer,
max_len=200,
n_layers=3):
super().__init__()
self.layers = nn.ModuleList(
map(lambda x: AttentionLayer(**attention_layer), range(n_layers)))
positional_encoding = position_encoding_init(max_len, features)
self.c_layers = nn.ModuleList(
map(lambda x: FeedForwardGELU(c_features, features * 2),
range(n_layers)))
self.p2x = nn.Linear(features, features * 2)
self.register_buffer('positional_encoding', positional_encoding)
def __init__(self, features, latent_dim, encoder, decoder, num_flows=3):
"""
features - number of features in the model
latent_dim - the latent dimension of the model
encoder - dictionary containing instantiation parameters for ResidualAttentionEncoder module
decoder - dictionary containing instantiation parameters for CondtionalResidualAttentionEncoder module
num_flows - the number of flows for the TriangularSylvesterFlow module
"""
super().__init__()
self.embedder = nn.Embedding(MAX_VALUE, features)
self.encoder = ResidualAttentionEncoder(**encoder)
self._latent_encoder = nn.ModuleList([
ResidualAttentionEncoder(**encoder),
TriangularSylvesterFlow(features, latent_dim, num_flows)
])
self.z_to_c = nn.Linear(latent_dim, latent_dim * 155)
self.decoder = CondtionalResidualAttentionEncoder(**decoder)
self.logits = FeedForwardGELU(features, MAX_VALUE)
self.n_features = features
def __init__(self,
features,
latent_dim,
encoder,
decoder,
deterministic_path_drop_rate=0.5,
num_flows=3):
super().__init__()
self.embedder = nn.Embedding(MAX_VALUE, features)
self.encoder = ResidualAttentionEncoder(**encoder)
self._latent_encoder = nn.ModuleList([
ResidualAttentionEncoder(**encoder),
TriangularSylvesterFlow(features, latent_dim, num_flows)
])
self.z_to_c = nn.Linear(latent_dim, latent_dim * 155)
self.decoder = CondtionalResidualAttentionEncoder(**decoder)
self.logits = FeedForwardGELU(features, MAX_VALUE)
self.drop = nn.Dropout(deterministic_path_drop_rate)
self.n_features = features
def __init__(self, features, c_features, attention_layer, max_len=200, n_layers=3):
"""
features - the number of features per parameter
c_features - the number of side conditioning features per batch item
attention_layer - a dictionary containing instantiation parameters for the AttentionLayer module
max_len - the maximum needed size of the positional encodings
n_layers - number of layers for the module to use
"""
super().__init__()
self.layers = nn.ModuleList(
map(lambda x: AttentionLayer(**attention_layer), range(n_layers))
)
positional_encoding = position_encoding_init(max_len, features)
self.c_layers = nn.ModuleList(
map(lambda x: FeedForwardGELU(c_features, features*2), range(n_layers))
)
self.p2x = nn.Linear(features, features * 2)
self.register_buffer('positional_encoding', positional_encoding)