def __init__(self, config):
super(conv_VQ_encoder, self).__init__()
conv_init = config['conv_init']
conv = config['conv']
att = config['att']
VQ = config['VQ']
self.max_len = config['max_len']
# initial conv over the full feature dimension
self.Conv = nn.Conv1d(in_channels=conv_init['in_channels'],
out_channels=conv_init['out_channels'],
kernel_size=conv_init['kernel_size'],
stride=conv_init['stride'],
padding=conv_init['padding'])
# residual conv blocks
self.res_convs = nn.ModuleList()
for x in range(conv['n_layers']):
self.res_convs.append(
res_conv(in_ch=conv['in_channels'][x],
out_ch=conv['out_channels'][x],
ks=conv['kernel_size'][x],
stride=conv['stride'][x]))
# VQ layers
self.VQ = nn.ModuleList()
for x in range(VQ['n_layers']):
self.VQ.append(VQ_EMA_layer(VQ['n_embs'][x], VQ['emb_dim'][x]))
# final pooling over the full time dimension
self.att = multi_attention(in_size=att['in_size'],
hidden_size=att['hidden_size'],
n_heads=att['heads'])
# application order of the vq and conv layers. list with 1 bool per
# VQ/res_conv layer.
self.app_order = config['app_order']
def __init__(self, config):
super(audio_rnn_encoder, self).__init__()
conv = config['conv']
rnn = config['rnn']
VQ = config['VQ']
att = config['att']
self.max_len = rnn['max_len']
#self.norm = nn.LayerNorm(conv['in_channels'])
self.Conv = nn.Conv1d(in_channels=conv['in_channels'],
out_channels=conv['out_channels'],
kernel_size=conv['kernel_size'],
stride=conv['stride'],
padding=conv['padding'])
self.RNN = nn.ModuleList()
for x in range(len(rnn['n_layers'])):
self.RNN.append(
nn.LSTM(input_size=rnn['input_size'][x],
hidden_size=rnn['hidden_size'][x],
num_layers=rnn['n_layers'][x],
batch_first=rnn['batch_first'],
bidirectional=rnn['bidirectional'],
dropout=rnn['dropout']))
# VQ layers
self.VQ = nn.ModuleList()
for x in range(VQ['n_layers']):
self.VQ.append(VQ_EMA_layer(VQ['n_embs'][x], VQ['emb_dim'][x]))
self.att = multi_attention(in_size=att['in_size'],
hidden_size=att['hidden_size'],
n_heads=att['heads'])
# application order of the vq and conv layers. list with 1 bool per
# VQ/res_conv layer.
self.app_order = config['app_order']
def __init__(self, config, log = True):
super(nwp_rnn_att, self).__init__()
embed = config['embed']
rnn = config['rnn']
lin = config['lin']
att = config ['att']
self.max_len = config['max_len']
self.embed = nn.Embedding(num_embeddings = embed['n_embeddings'],
embedding_dim = embed['embedding_dim'],
sparse = embed['sparse'],
padding_idx = embed['padding_idx'])
self.RNN = nn.GRU(input_size = rnn['in_size'],
hidden_size = rnn['hidden_size'],
num_layers = rnn['n_layers'],
batch_first = rnn['batch_first'],
bidirectional = rnn['bidirectional'],
dropout = rnn['dropout'])
self.att = multi_attention(in_size = rnn['hidden_size'],
hidden_size = att['hidden_size'],
n_heads = att['heads'])
self.linear = nn.Sequential(nn.Linear(rnn['hidden_size'],
lin['hidden_size']
),
nn.ReLU(),
nn.Linear(lin['hidden_size'],
embed['n_embeddings']
)
)
if log:
self.log(embed, rnn, lin, att)
def __init__(self, config):
super(audio_rnn_encoder, self).__init__()
conv = config['conv']
rnn= config['rnn']
att = config ['att']
self.Conv = nn.Conv1d(in_channels = conv['in_channels'],
out_channels = conv['out_channels'], kernel_size = conv['kernel_size'],
stride = conv['stride'], padding = conv['padding'])
self.RNN = nn.GRU(input_size = rnn['input_size'], hidden_size = rnn['hidden_size'],
num_layers = rnn['num_layers'], batch_first = rnn['batch_first'],
bidirectional = rnn['bidirectional'], dropout = rnn['dropout'])
self.att = multi_attention(in_size = att['in_size'], hidden_size = att['hidden_size'], n_heads = att['heads'])
def __init__(self, config):
super(text_rnn_encoder, self).__init__()
embed = config['embed']
rnn= config['rnn']
att = config ['att']
self.embed = nn.Embedding(num_embeddings = embed['num_chars'],
embedding_dim = embed['embedding_dim'], sparse = embed['sparse'],
padding_idx = embed['padding_idx'])
self.RNN = nn.GRU(input_size = rnn['input_size'], hidden_size = rnn['hidden_size'],
num_layers = rnn['num_layers'], batch_first = rnn['batch_first'],
bidirectional = rnn['bidirectional'], dropout = rnn['dropout'])
self.att = multi_attention(in_size = att['in_size'], hidden_size = att['hidden_size'], n_heads = att['heads'])