def encode(self, inputs, lengths, fr=0):
bsz, max_len = inputs.size()
e_hidden_init = self.e_hidden_init.expand(
2, bsz, self.hidden_dim).contiguous()
e_cell_init = self.e_cell_init.expand(2, bsz,
self.hidden_dim).contiguous()
lens, indices = torch.sort(lengths, 0, True)
if fr and not self.share_vocab:
in_embs = self.embedding_fr(inputs)
else:
in_embs = self.embedding(inputs)
if fr and not self.share_encoder:
if self.dropout > 0:
F.dropout(in_embs, training=self.training)
all_hids, (enc_last_hid, _) = self.lstm_fr(
pack(in_embs[indices], lens.tolist(), batch_first=True),
(e_hidden_init, e_cell_init))
else:
if self.dropout > 0:
F.dropout(in_embs, training=self.training)
all_hids, (enc_last_hid, _) = self.lstm(
pack(in_embs[indices], lens.tolist(), batch_first=True),
(e_hidden_init, e_cell_init))
_, _indices = torch.sort(indices, 0)
all_hids = unpack(all_hids, batch_first=True)[0][_indices]
if self.pool == "max":
embs = model_utils.max_pool(all_hids, lengths, self.gpu)
elif self.pool == "mean":
embs = model_utils.mean_pool(all_hids, lengths, self.gpu)
return embs
def reduce_map(weights, layer_feats, num_split, pool_method):
if pool_method == Pools.AVG:
train_inp = avg_pool(layer_feats, num_split=num_split)
elif pool_method == Pools.MAX:
train_inp = max_pool(layer_feats, num_split=num_split)
else: # pool_method is random
train_inp = randomized_pool(weights, layer_feats, num_split=num_split)
return train_inp
def encode(self, idxs, lengths, fr=0):
if fr and not self.share_vocab:
word_embs = self.embedding_fr(idxs)
else:
word_embs = self.embedding(idxs)
if self.dropout > 0:
F.dropout(word_embs, training=self.training)
if self.pool == "max":
word_embs = model_utils.max_pool(word_embs, lengths, self.args.gpu)
elif self.pool == "mean":
word_embs = model_utils.mean_pool(word_embs, lengths, self.args.gpu)
return word_embs