Repository navigation
Expand file tree
/
Copy pathEncoderDecoder.py
More file actions
43 lines (38 loc) · 1.67 KB
/
Copy pathEncoderDecoder.py
File metadata and controls
43 lines (38 loc) · 1.67 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
from typing import Optional
import paddle
import paddle.nn as nn
from paddle import Tensor
from Decoder import Decoder
from Embedding import TransformerEmbedding, PositionalEncoding
from Encoder import Encoder
from utils import PAD_ID
class EncoderDecoder(nn.Layer):
def __init__(self, vocab_size: int, d_model: int = 512):
super(EncoderDecoder, self).__init__()
self.layers_nums = 3
self.embedding = nn.Sequential(
TransformerEmbedding(vocab_size),
PositionalEncoding()
)
self.encoder = Encoder(self.layers_nums)
self.decoder = Decoder(self.layers_nums)
self.linear = nn.Linear(d_model, vocab_size)
self.soft_max = nn.Softmax()
# ignore_index 必须等于 utils 里实际填进 label 的 PAD_ID, 两处一致才能跳过 padding
self.loss_fct = nn.CrossEntropyLoss(ignore_index=PAD_ID)
def forward(self, x, label, true_label: Optional[Tensor] = None, src_mask=None, tgt_mask=None):
input_embedding = self.embedding(x)
label_embedding = self.embedding(label)
encoder_output = self.encoder(input_embedding, src_mask)
decoder_output = self.decoder(label_embedding, encoder_output, src_mask, tgt_mask)
logits = self.linear(decoder_output)
res_dict = {}
if true_label is not None:
loss = self.loss_fct(logits.reshape((-1, logits.shape[-1])),
true_label.reshape((-1,)))
res_dict['loss'] = loss
result = self.soft_max(logits)
max_index = paddle.argmax(result, axis=-1)
res_dict['logits'] = result
res_dict['index'] = max_index
return res_dict