| 309 | |
| 310 | |
| 311 | class TextEncoderBiGRUCo(nn.Module): |
| 312 | def __init__(self, word_size, pos_size, hidden_size, output_size, device): |
| 313 | super(TextEncoderBiGRUCo, self).__init__() |
| 314 | self.device = device |
| 315 | |
| 316 | self.pos_emb = nn.Linear(pos_size, word_size) |
| 317 | self.input_emb = nn.Linear(word_size, hidden_size) |
| 318 | self.gru = nn.GRU(hidden_size, hidden_size, batch_first=True, bidirectional=True) |
| 319 | self.output_net = nn.Sequential( |
| 320 | nn.Linear(hidden_size * 2, hidden_size), |
| 321 | nn.LayerNorm(hidden_size), |
| 322 | nn.LeakyReLU(0.2, inplace=True), |
| 323 | nn.Linear(hidden_size, output_size) |
| 324 | ) |
| 325 | |
| 326 | self.input_emb.apply(init_weight) |
| 327 | self.pos_emb.apply(init_weight) |
| 328 | self.output_net.apply(init_weight) |
| 329 | # self.linear2.apply(init_weight) |
| 330 | # self.batch_size = batch_size |
| 331 | self.hidden_size = hidden_size |
| 332 | self.hidden = nn.Parameter(torch.randn((2, 1, self.hidden_size), requires_grad=True)) |
| 333 | |
| 334 | # input(batch_size, seq_len, dim) |
| 335 | def forward(self, word_embs, pos_onehot, cap_lens): |
| 336 | num_samples = word_embs.shape[0] |
| 337 | |
| 338 | pos_embs = self.pos_emb(pos_onehot) |
| 339 | inputs = word_embs + pos_embs |
| 340 | input_embs = self.input_emb(inputs) |
| 341 | hidden = self.hidden.repeat(1, num_samples, 1) |
| 342 | |
| 343 | cap_lens = cap_lens.data.tolist() |
| 344 | emb = pack_padded_sequence(input_embs, cap_lens, batch_first=True) |
| 345 | |
| 346 | gru_seq, gru_last = self.gru(emb, hidden) |
| 347 | |
| 348 | gru_last = torch.cat([gru_last[0], gru_last[1]], dim=-1) |
| 349 | |
| 350 | return self.output_net(gru_last) |
| 351 | |
| 352 | |
| 353 | class MotionEncoderBiGRUCo(nn.Module): |