problem with interactive.py about sca

[teslacool]

When inferring a sentence, we do not touch the lm. First use nmt's encoder to generate hidden states and then work on nmt's decoder. You should look carefully at the fairseq decoding steps.

And now i test a sentence and fixed a bug. You can use interactive.py like original fairseq without any other command parameters.

[nicolabertoldi]

@teslacool
could you please give me an example how to run properly the command ./interactive.py

assuming that I have the following checkpoint models:
for srclm: ./lm_sl/checkpoint_best.pt
for tgtlm: ./lm_tl/checkpoint_best.pt
for transformer: ./engine/checkpoint_best.pt

I got these checkpoints by running the following preprocessing and training commands:

fairseq-preprocess --source-lang sl --target-lang tl --trainpref ./encoded_corpora//train --validpref ./encoded_corpora//dev --destdir ./data_generated --joined-dictionary

fairseq-train --task language_modeling --arch transformer_lm --lr-scheduler inverse_sqrt --lr-shrink 0.1 --warmup-updates 4000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer adam --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 4096 --update-freq 8 --seed 1 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --save-interval-updates 1000 --keep-interval-updates 10 --no-epoch-checkpoints --attention-dropout 0.1 --dropout 0.3 --criterion label_smoothed_cross_entropy --save-dir ./lm_sl/ ./data_generated_sl

fairseq-train --task language_modeling --arch transformer_lm --lr-scheduler inverse_sqrt --lr-shrink 0.1 --warmup-updates 4000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer adam --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 4096 --update-freq 8 --seed 1 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --save-interval-updates 1000 --keep-interval-updates 10 --no-epoch-checkpoints --attention-dropout 0.1 --dropout 0.3 --criterion label_smoothed_cross_entropy --save-dir ./lm_tl/ ./data_generated_tl

python3 ../train.py ./data_generated --task lm_translation --arch transformer_iwslt_de_en --share-decoder-input-output-embed --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --lr-scheduler inverse_sqrt --warmup-init-lr 1e-07 --warmup-updates 4000 --lr 0.0009 --min-lr 1e-09 --dropout 0.3 --weight-decay 0.0 --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --max-tokens 2084 --update-freq 8 --save-dir ./engine --save-interval-updates 1000 --seed 200 --tradeoff 0.15 --load-lm --load-srclm-file ./lm_sl/checkpoint_best.pt --load-tgtlm-file ./lm_tl/checkpoint_best.pt --lmdecoder-ffn-embed-dim 2048 --no-epoch-checkpoints --keep-interval-updates 10

Note that all these scripts succedeed.

[teslacool]

run

echo "Danke dir ." | python interactive.py data-bin/iwslt14.tokenized.de-en     --path  checkpoints/transformer/checkpoint_best.pt --buffer-size 1024     --batch-size 128 --beam 5 --remove-bpe  | grep ^H | cut -f3-

you will get

thank you .

if your model is deen task.

[nicolabertoldi]

@teslacool

after your fix, it works

echo "ciao ciao ciao" | python3 ../interactive.py --remove-bpe --raw-text --path ./engine/checkpoint_best.pt --task lm_translation --src-no-lm --tgt-no-lm --load-srclm-file ./lm_sl/checkpoint_best.pt  --load-tgtlm-file ./lm_tl/checkpoint_best.pt  ./data_generated 

thank you

[nicolabertoldi]

@teslacool

I have a few more questions:

• why do I need do specify the data-bin directory? Only for the dictionary? or is there another reason?

[teslacool]

yes, you need dict to get id for embedding matrix.

[nicolabertoldi]

• I specify --remove-bpe, but my output still contains the _ symbols; why does this happen?

[teslacool]

remove bpe is to remove @@

[nicolabertoldi]

so it is correct that my output looks like

I_ am_ a_ newbie_ ._

[nicolabertoldi]

actually I get

I_ am_ a_ new bie_ ._

[teslacool]

it looks like not correct. I do not know why your sentences consist of words following _.

Because I and am are not a single word, the symbol _ was not added by bpe operation.

[nicolabertoldi]

@teslacool

sorry, my fault.

I close the issue by now,
but probably I will have more questions in short time.