Pass existing tensorboard SummaryWriter to Trainer PR (#4019)

[jaymody]

Implements feature request for issue #4019.

You should be able to pass in you're own SummaryWriters to Trainer via the tb_writer parameter to the __init__ function:

tb_writer = SummaryWriter(log_dir="my_log_dir")
tb_writer.add_hparams(my_hparams_dict, my_metrics_dict)

trainer = Trainer(
    model = model,
    args = training_args,
    train_dataset = train_dataset,
    tb_writer = tb_writer
)

trainer.train()

[codecov-io]

Codecov Report

Merging #4020 into master will increase coverage by 0.00%.
The diff coverage is 66.66%.

@@           Coverage Diff           @@
##           master    #4020   +/-   ##
=======================================
  Coverage   78.44%   78.45%           
=======================================
  Files         111      111           
  Lines       18518    18520    +2     
=======================================
+ Hits        14527    14529    +2     
  Misses       3991     3991           

Impacted Files	Coverage Δ
src/transformers/trainer.py	43.94% <66.66%> (+0.04%)	⬆️
src/transformers/modeling_tf_utils.py	92.76% <0.00%> (ø)
src/transformers/modeling_utils.py	91.06% <0.00%> (+0.12%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4e817ff...62dbe2d. Read the comment docs.

[jaymody]

Another consideration is that the user might want to add extra details to the log after the training has finished. Right now, the end of the training loop closes the tb_writer, preventing any further logging. Not sure what a reasonable workaround for this is, maybe add a param for the train function that if set to True, will close tb_writer and if False, will keep it open (default: True).

[julien-c]

Looks good, thanks