def __init__(self, config, log = True):
super(nwp_transformer_2lin, self).__init__()
embed = config['embed']
tf= config['tf']
self.is_cuda = config['cuda']
self.max_len = tf['max_len']
self.embed = nn.Embedding(num_embeddings = embed['n_embeddings'],
embedding_dim = embed['embedding_dim'],
sparse = embed['sparse'],
padding_idx = embed['padding_idx'])
# prepares the positional embeddings
self.pos_emb = self.pos_embedding(tf['max_len'],embed['embedding_dim'])
self.TF_enc = transformer_encoder(in_size = tf['in_size'],
fc_size = tf['fc_size'],
n_layers = tf['n_layers'],
h = tf['h'])
self.linear = nn.Sequential(nn.Linear(tf['in_size'],
tf['in_size']
),
nn.ReLU(),
nn.Linear(tf['in_size'],
embed['n_embeddings']
)
)
if log:
self.log(embed, tf)
def __init__(self, config, log = True):
super(nwp_transformer, self).__init__()
embed = config['embed']
tf= config['tf']
self.is_cuda = config['cuda']
self.max_len = tf['max_len']
self.embed = nn.Embedding(num_embeddings = embed['n_embeddings'],
embedding_dim = embed['embedding_dim'],
sparse = embed['sparse'],
padding_idx = embed['padding_idx'])
# prepares the positional embeddings
self.pos_emb = self.pos_embedding(tf['max_len'], embed['embedding_dim'])
self.TF_enc = transformer_encoder(in_size = tf['in_size'],
fc_size = tf['fc_size'],
n_layers = tf['n_layers'],
h = tf['h'])
# linear layer has no extra configurations, it just maps directly
# from the transformer output to the number of embeddings
self.linear = nn.Linear(tf['in_size'], embed['n_embeddings'])
if log:
self.log(embed, tf)
def __init__(self, config):
super(translator_transformer, self).__init__()
embed = config['embed']
tf = config['tf']
self.is_cuda = config['cuda']
self.max_len = tf['max_len']
# create the embedding layer
self.embed = nn.Embedding(num_embeddings = embed['num_chars'],
embedding_dim = embed['embedding_dim'], sparse = embed['sparse'],
padding_idx = embed['padding_idx'])
# create the positional embeddings
self.pos_emb = self.pos_embedding(tf['max_len'], embed['embedding_dim'])
# create the (stacked) transformer
self.TF_enc = transformer_encoder(in_size = tf['input_size'], fc_size = tf['fc_size'],
n_layers = tf['n_layers'], h = tf['h'])
self.TF_dec = transformer_decoder(in_size = tf['input_size'], fc_size = tf['fc_size'],
n_layers = tf['n_layers'], h = tf['h'])
self.linear = nn.Linear(embed['embedding_dim'], embed['num_chars'])