def forward(self, x):
batch_size = x.size(0)
lens = torch.randint(1, self.max_len + 1, (batch_size,)).long()
mask = utils.mask(lens, self.max_len)
x = self.in_linear(x).view(batch_size, self.max_len, -1)
x = x.masked_fill(1 - mask.unsqueeze(-1), float("nan"))
return self.out_linear(self.pool(x, lens))
def forward(self, x):
batch_size, max_len = x.size(0), self.max_len
x = self.in_linear(x).view(batch_size, max_len, -1)
lens = torch.randint(1, max_len + 1, (batch_size,)).long()
lens = lens.to(x)
mask = utils.mask(lens, max_len)
x = x.masked_fill(1 - mask.unsqueeze(-1), float("nan"))
o, _, h = self.rnn(x, lens)
x = torch.cat([o.contiguous().view(batch_size, -1), h], 1)
return self.out_linear(x)
def forward(self, x):
batch_size = x.size(0)
qrys = self.qry_linear(x).view(batch_size, self.num_qrys, -1)
keys = self.key_linear(x).view(batch_size, self.max_keys, -1)
num_keys = torch.randint(1, self.max_keys + 1, (batch_size,))
num_keys = num_keys.long()
mask = utils.mask(num_keys, self.max_keys).byte()
keys = keys.masked_fill(1 - mask.unsqueeze(-1), float("nan"))
att = self.att(qrys, keys, num_keys)
return self.out_linear(att.view(batch_size, -1))
def forward(self, x, lens=None):
if lens is None:
return x.mean(1)
mask = utils.mask(lens, x.size(1))
x = x.masked_fill(1 - mask.unsqueeze(-1), 0)
return x.sum(1) / lens.unsqueeze(-1).float()
def forward(self, x, lens=None):
if lens is not None:
mask = utils.mask(lens, x.size(1))
x = x.masked_fill(1 - mask.unsqueeze(-1), 1)
return x.prod(1)
def forward(self, x, lens=None):
if lens is not None:
mask = utils.mask(lens, x.size(1))
x = x.masked_fill(1 - mask.unsqueeze(-1), float("-inf"))
return x.max(1)[0]
def forward(self, x, lens=None):
if lens is None:
return x.mean(1)
mask = utils.mask(lens, x.size(1)).unsqueeze(-1).float()
return (x * mask).sum(1) / lens.unsqueeze(-1).float()
def forward(self, x, lens=None):
if lens is not None:
mask = utils.mask(lens, x.size(1)).unsqueeze(-1)
x *= mask.float()
return self.pool(x)
