Fix export code after model "hard fork" (#667)

Summary:
Pull Request resolved: #667

After internal version is introduced and transformer uses namedtuple instead of dict, need to update the export code accordingly.

Also we used torch.empty([0]) to represent initial attention instead of None, to unify the data type in TorchScript.

Reviewed By: jhcross

Differential Revision: D18648224

fbshipit-source-id: efea4252675eae98de24029f18790a7cda8ecdbd

Author: cndn

pytorch_translate/ensemble_export.py

@@ -19,8 +19,10 @@
 from caffe2.python.onnx import backend as caffe2_backend
 from caffe2.python.predictor import predictor_exporter
 from fairseq import tasks, utils
+from fairseq.iterative_refinement_generator import DecoderOut
 from fairseq.models import ARCH_MODEL_REGISTRY
-from fairseq.models.model_utils import script_skip_tensor, script_skip_tensor_list
+from fairseq.models.model_utils import script_skip_tensor
+from fairseq.models.transformer import EncoderOut
 from pytorch_translate.beam_decode import BeamDecode
 from pytorch_translate.data import dictionary
 from pytorch_translate.research.knowledge_distillation import (
@@ -2194,42 +2196,43 @@ def generate(self, models, src_tokens, src_lengths, prefix_tokens=None):
 
         # initialize buffers (very model specific, with length prediction or not)
         prev_decoder_out = model.initialize_output_tokens(encoder_out, src_tokens)
-        prev_output_tokens = prev_decoder_out[0].clone()
+        prev_output_tokens = prev_decoder_out.output_tokens.clone()
 
         finalized_tokens_list = [torch.tensor(0) for _ in range(bsz)]
         finalized_scores_list = [torch.tensor(0) for _ in range(bsz)]
         finalized_attns_list = [torch.tensor(0) for _ in range(bsz)]
         finalized_alignments_list = [torch.tensor(0) for _ in range(bsz)]
-        prev_decoder_out[4] = self.max_iter + 1
+        prev_decoder_out._replace(max_step=self.max_iter + 1)
 
         for step in range(self.max_iter + 1):
-            prev_decoder_out[3] = step
+            prev_decoder_out._replace(step=step)
             decoder_out = model.forward_decoder(
                 prev_decoder_out,
                 encoder_out,
                 eos_penalty=self.eos_penalty,
                 max_ratio=self.max_ratio if step == 0 else None,
                 decoding_format=self.decoding_format,
             )
-
             terminated, output_tokens, output_scores, output_attn = is_a_loop(
                 self.pad,
                 prev_output_tokens,
-                decoder_out[0],
-                decoder_out[1],
-                decoder_out[2],
+                decoder_out.output_tokens,
+                decoder_out.output_scores,
+                decoder_out.attn,
+            )
+            decoder_out._replace(
+                output_tokens=output_tokens,
+                output_scores=output_scores,
+                attn=output_attn,
             )
-            decoder_out[0] = output_tokens
-            decoder_out[1] = output_scores
-            decoder_out[2] = output_attn
 
             terminated = last_step(step, self.max_iter, terminated)
             # collect finalized sentences
             finalized_idxs = sent_idxs[terminated]
-            finalized_tokens = decoder_out[0][terminated]
-            finalized_scores = decoder_out[1][terminated]
+            finalized_tokens = decoder_out.output_tokens[terminated]
+            finalized_scores = decoder_out.output_scores[terminated]
             finalized_attn = (
-                None if decoder_out[2] is None else decoder_out[2][terminated]
+                None if decoder_out.attn is None else decoder_out.attn[terminated]
             )
             finalized_tokens_list = finalize_hypos_loop_tokens(
                 finalized_tokens_list,
@@ -2256,17 +2259,30 @@ def generate(self, models, src_tokens, src_lengths, prefix_tokens=None):
             )
 
             # for next step
-            prev_decoder_out = [
-                script_skip_tensor(decoder_out[0], ~terminated),
-                script_skip_tensor(decoder_out[1], ~terminated),
-                decoder_out[2],
-                decoder_out[3],
-                decoder_out[4],
-            ]
-            encoder_out = script_skip_tensor_list(encoder_out, ~terminated)
+            prev_decoder_out = DecoderOut(
+                output_tokens=script_skip_tensor(
+                    decoder_out.output_tokens, ~terminated
+                ),
+                output_scores=script_skip_tensor(
+                    decoder_out.output_scores, ~terminated
+                ),
+                attn=decoder_out.attn,
+                step=decoder_out.step,
+                max_step=decoder_out.max_step,
+                history=None,
+            )
+
+            encoder_out = EncoderOut(
+                encoder_out=script_skip_tensor(encoder_out.encoder_out, ~terminated),
+                encoder_padding_mask=None,
+                encoder_embedding=script_skip_tensor(
+                    encoder_out.encoder_embedding, ~terminated
+                ),
+                encoder_states=None,
+            )
             sent_idxs = script_skip_tensor(sent_idxs, ~terminated)
 
-            prev_output_tokens = prev_decoder_out[0].clone()
+            prev_output_tokens = prev_decoder_out.output_tokens.clone()
 
         return (
             finalized_tokens_list,