Dynamic batching not working

[frankplus]

Hi,
when setting --dynamic-batching full it throws an error at the start of the training

  File "/usr/local/bin/parlai", line 8, in <module>
    sys.exit(main())
  File "/usr/local/lib/python3.7/dist-packages/parlai/__main__.py", line 14, in main
    superscript_main()
  File "/usr/local/lib/python3.7/dist-packages/parlai/core/script.py", line 306, in superscript_main
    return SCRIPT_REGISTRY[cmd].klass._run_from_parser_and_opt(opt, parser)
  File "/usr/local/lib/python3.7/dist-packages/parlai/core/script.py", line 89, in _run_from_parser_and_opt
    return script.run()
  File "/usr/local/lib/python3.7/dist-packages/parlai/scripts/train_model.py", line 809, in run
    return self.train_loop.train()
  File "/usr/local/lib/python3.7/dist-packages/parlai/scripts/train_model.py", line 693, in train
    world.parley()
  File "/usr/local/lib/python3.7/dist-packages/parlai/core/worlds.py", line 1097, in parley
    self._task_acts[i]['dyn_batch_idx'] = i
  File "/usr/local/lib/python3.7/dist-packages/parlai/core/message.py", line 28, in __setitem__
    'please use the function `force_set(key, value)`.'.format(key)
RuntimeError: Message already contains key `dyn_batch_idx`. If this was intentional, please use the function `force_set(key, value)`.

This is the command I used:

!parlai train_model \
-t fromfile:parlaiformat \
--fromfile_datapath $DATASET_FILE \
--fromfile-datatype-extension true \
-m transformer/generator \
--save-every-n-secs 300 \
--validation-every-n-secs 600 \
--eval-batchsize 256 \
--dynamic-batching full \
--fp16 True --fp16-impl mem_efficient \
--batchsize 128 \
--embedding-size 512 \
--n-encoder-layers 1 \
--n-decoder-layers 13 \
--ffn-size 2048 \
--dropout 0.1 \
--n-heads 16 \
--learn-positional-embeddings True \
--n-positions 512 \
--variant xlm \
--activation gelu \
--skip-generation True \
--text-truncate 128 \
--label-truncate 128 \
--truncate 128 \
--dict-tokenizer bpe \
--dict-lower True \
-lr 1e-03 \
--optimizer adam \
--lr-scheduler reduceonplateau \
--gradient-clip 0.1 \
-veps 0.25 \
--betas 0.9,0.999 \
--update-freq 1 \
--attention-dropout 0.0 \
--relu-dropout 0.0 \
-vp 15 -stim 60 \
-vme 20000 \
-vmt ppl \
-vmm min \
--save-after-valid True \
--model-file $MODEL 

[stephenroller]

Weird. You don't have that field in your dataset do you?

[frankplus]

Weird. You don't have that field in your dataset do you?

There's no dyn_batch_idx in the dataset. Only standard parlai format:

text:è casa mia .	labels:nessuno mi caccia di casa .
text:voglio mangiare .	labels:ti ho sentito le prime cinque volte .	episode_done:True
...

[stephenroller]

Your fix seems reasonable. Do you mind submitting a PR for it?