Skip to content
This repository was archived by the owner on Aug 1, 2023. It is now read-only.

Flow side support for separate decoders / Fix max_iter at inference #659

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Flow side support for separate decoders / Fix max_iter at inference #659

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 11 additions & 20 deletions pytorch_translate/ensemble_export.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2073,7 +2073,7 @@ def finalize_hypos_loop_attns(

class IterativeRefinementGenerateAndDecode(torch.jit.ScriptModule):
def __init__(
self, checkpoint_files, src_dict_filename, tgt_dict_filename, max_iter=2
self, checkpoint_files, src_dict_filename, tgt_dict_filename, max_iter=1
):
super().__init__()
self.models, _, tgt_dict = load_models_from_checkpoints(
Expand Down Expand Up @@ -2209,21 +2209,16 @@ def generate(self, models, src_tokens, src_lengths, prefix_tokens=None):
decoding_format=self.decoding_format,
)

if self.adaptive:
# terminate if looping.
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[0] = output_tokens
decoder_out[1] = output_scores
decoder_out[2] = output_attn

else:
terminated = torch.zeros_like(decoder_out[0]).bool()
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[0] = output_tokens
decoder_out[1] = output_scores
decoder_out[2] = output_attn

terminated = last_step(step, self.max_iter, terminated)
# collect finalized sentences
Expand Down Expand Up @@ -2257,10 +2252,6 @@ def generate(self, models, src_tokens, src_lengths, prefix_tokens=None):
finalized_attn,
)

# check if all terminated
if terminated.sum() == terminated.size(0):
break

# for next step
prev_decoder_out = [
script_skip_tensor(decoder_out[0], ~terminated),
Expand Down