Distillation Joint Translate Bug about indictrans2 HOT 3 OPEN

@PranjalChitale @VarunGumma
Also, even if I bypass this issue fairseq-interactive isn't taking in the input file.

Following is the input command:
fairseq-interactive ${ckpt_dir}/final_bin \ --distributed-world-size 1 --memory-efficient-fp16 \ --path ${ckpt_dir}/models/checkpoint_best.pt \ --task translation \ --source-lang SRC --target-lang TGT \ --batch-size 256 --buffer-size 2500 --beam 5 \ --num-workers 24 \ --skip-invalid-size-inputs-valid-test \ --input $outfname.bpe > $outfname.log 2>&1

In the above syntax, --input parameter has the valid outfname.bpe file but in the logs I am unable to check this as input. I am attaching the cfg.interactive, the one logged by the script, input should not be equal to '-'.

"interactive":{ "_name":"None", "buffer_size":2500, "input":"-", "force_override_max_positions":"None"}

from indictrans2.

The issue described above is due to the IndicNLP resources not being installed and the path not being set correctly.

Please refer to this link for guidance.

Additionally, because the preprocessing failed, the outfname.bpe file was not created successfully.

Regarding the README for the distillation branch, it may have been accidentally removed in the previous commit.

An up-to-date README will be added soon.

from indictrans2.

@PranjalChitale As mentioned above, I was able to solve the initial issue of IndicNLP, and after having the correct outfname.bpe (verified the file) I faced the above mentioned issue where the fairseq-interactive wasn't considering in the given input.

from indictrans2.