def __init__(self,
size_input,
size,
depth=1,
size_attn=512,
dropout_p=0.0):
super(SpeechEncoderTopStack, self).__init__()
util.autoassign(locals())
if self.depth > 0:
self.Dropout = nn.Dropout(p=self.dropout_p)
self.RNN = GRUStack(self.size_input, self.size, self.depth)
self.Attn = attention.SelfAttention(self.size, size=self.size_attn)
def __init__(self,
size_feature,
size,
depth=1,
size_attn=512,
dropout_p=0.0):
super(TextEncoderTop, self).__init__()
util.autoassign(locals())
if self.depth > 0:
self.h0 = torch.autograd.Variable(
torch.zeros(self.depth, 1, self.size))
self.Dropout = nn.Dropout(p=self.dropout_p)
self.RNN = nn.GRU(self.size_feature,
self.size,
self.depth,
batch_first=True)
self.Attn = attention.SelfAttention(self.size, size=self.size_attn)
def __init__(self,
size_input,
size,
depth=1,
size_attn=512,
dropout_p=0.0):
super(SpeechEncoderTopBidi, self).__init__()
util.autoassign(locals())
if self.depth > 0:
self.Dropout = nn.Dropout(p=self.dropout_p)
self.RNN = nn.GRU(self.size_input,
self.size,
self.depth,
batch_first=True,
bidirectional=True)
self.Down = nn.Linear(self.size * 2, self.size)
self.Attn = attention.SelfAttention(self.size, size=self.size_attn)
def __init__(self, config):
super(Audio, self).__init__()
util.autoassign(locals())
self.margin_size = config.get('margin_size', 0.2)
# FIXME FIXME ADD gradient clipping!
#self.make_updater = lambda: optim.Adam(self.parameters(), lr=config['lr'])
self.max_norm = config['max_norm']
self.Encode = Encoder(config['size_vocab'],
config['size'],
filter_length=config.get('filter_length', 6),
filter_size=config.get('filter_size', 1024),
stride=config.get('stride', 3),
depth=config.get('depth', 1))
self.Attn = attention.SelfAttention(config['size'],
size=config.get('size_attn', 512))
self.ProjBeg = nn.Linear(config['size'], config['size_target'])
self.ProjEnd = nn.Linear(config['size'], config['size_target'])
def __init__(self, config):
super(Audio, self).__init__()
util.autoassign(locals())
# FIXME FIXME ADD gradient clipping!
#self.make_updater = lambda: optim.Adam(self.parameters(), lr=config['lr'])
self.max_norm = config['max_norm']
self.Encode = Encoder(config['size_vocab'],
config['size'],
filter_length=config.get('filter_length', 6),
filter_size=config.get('filter_size', 1024),
stride=config.get('stride', 3),
depth=config.get('depth', 1),
residual=config.get('residual', False))
self.Attn = attention.SelfAttention(config['size'],
size=config.get('size_attn', 512))
self.Decode1 = Decoder(config['size_vocab'], config['size'])
self.Decode3 = Decoder(config['size_vocab'], config['size'])
def __init__(self, config):
super(Visual, self).__init__()
util.autoassign(locals())
self.margin_size = config.get('margin_size', 0.2)
# FIXME FIXME ADD gradient clipping!
#self.make_updater = lambda: optim.Adam(self.parameters(), lr=config['lr'])
self.max_norm = config['max_norm']
self.Encode = Encoder(config['size_vocab'],
config['size'],
filter_length=config.get('filter_length', 6),
filter_size=config.get('filter_size', 1024),
stride=config.get('stride', 3),
depth=config.get('depth', 1),
recur_depth=config.get('recur_depth', 1),
drop_i=config.get('drop_i', 0.75),
drop_s=config.get('drop_s', 0.25),
residual=config.get('residual', False),
seed=config.get('seed', 1))
self.Attn = attention.SelfAttention(config['size'],
size=config.get('size_attn', 512))
self.ImgEncoder = util.make_linear(config['size_target'],
config['size'])
def __init__(self,
size_vocab,
size,
depth=1,
filter_length=6,
filter_size=64,
stride=2,
size_attn=512,
dropout_p=0.0):
super(SpeechEncoder, self).__init__()
util.autoassign(locals())
self.h0 = torch.autograd.Variable(torch.zeros(self.depth, 1,
self.size))
self.Conv = conv.Convolution1D(self.size_vocab,
self.filter_length,
self.filter_size,
stride=self.stride)
self.Dropout = nn.Dropout(p=self.dropout_p)
self.RNN = nn.GRU(self.filter_size,
self.size,
self.depth,
batch_first=True)
self.Attn = attention.SelfAttention(self.size, size=self.size_attn)
