Daniel Voigt Godoy - Deep Learning with PyTorch Step-by-Step A Beginner’s Guide-leanpub

22 self.decoder.init_keys(encoder_states)
23
24 def decode(self, shifted_target_seq,
25 source_mask=None, target_mask=None):
26 # Decodes / generates a sequence using the shifted
27 # (masked) target sequence - used in TRAIN mode
28 outputs = self.decoder(shifted_target_seq,
29 source_mask=source_mask,
30 target_mask=target_mask)
31 return outputs
32
33 def predict(self, source_seq, source_mask):
34 # Decodes / generates a sequence using one input
35 # at a time - used in EVAL mode
36 inputs = source_seq[:, -1:]
37 for i in range(self.target_len):
38 out = self.decode(inputs,
39 source_mask,
40 self.trg_masks[:, :i+1, :i+1])
41 out = torch.cat([inputs, out[:, -1:, :]], dim=-2)
42 inputs = out.detach()
43 outputs = inputs[:, 1:, :]
44 return outputs
45
46 def forward(self, X, source_mask=None):
47 # Sends the mask to the same device as the inputs
48 self.trg_masks = self.trg_masks.type_as(X).bool()
49 # Slices the input to get source sequence
50 source_seq = X[:, :self.input_len, :]
51 # Encodes source sequence AND initializes decoder
52 self.encode(source_seq, source_mask)
53 if self.training:
54 # Slices the input to get the shifted target seq
55 shifted_target_seq = X[:, self.input_len-1:-1, :]
56 # Decodes using the mask to prevent cheating
57 outputs = self.decode(shifted_target_seq,
58 source_mask,
59 self.trg_masks)
60 else:
61 # Decodes using its own predictions
62 outputs = self.predict(source_seq, source_mask)
63
760 | Chapter 9 — Part II: Sequence-to-Sequence