Simplify getting latest checkpoint in examples/albert

[mryab]

From Discord:

Based on huggingface/transformers#10334, I think it's safe for us to rely on transformers for getting the last checkpoint from output_dir and get rid of it in our code. I've drafted #425, will test later this week.

[codecov]

Codecov Report

Merging #425 (b829c7f) into master (595b831) will decrease coverage by 0.13%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #425      +/-   ##
==========================================
- Coverage   84.06%   83.92%   -0.14%     
==========================================
  Files          77       77              
  Lines        7924     7924              
==========================================
- Hits         6661     6650      -11     
- Misses       1263     1274      +11     

Impacted Files	Coverage Δ
hivemind/averaging/matchmaking.py	85.11% <0.00%> (-2.39%)	⬇️
hivemind/averaging/averager.py	87.65% <0.00%> (-0.73%)	⬇️

[mryab]

Tested it with the checkpoint present in the directory

Before:

Dec 27 15:12:14.244 [INFO] Checkpoint dir outputs, contents [PosixPath('outputs/checkpoint-125000'), PosixPath('outputs/checkpoint-124500')]
Dec 27 15:12:14.244 [INFO] Loading model from outputs/checkpoint-125000

After (https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py#L1096):

Dec 27 15:13:20.843 [INFO] max_steps is given, it will override any value given in num_train_epochs
Dec 27 15:13:20.843 [INFO] Using amp fp16 backend
Dec 27 15:13:20.844 [INFO] Loading model from outputs/checkpoint-125000).

[mryab]

Unfortunately, without any checkpoint in the directory, it gives

    raise ValueError(f"No valid checkpoint found in output directory ({args.output_dir})")
ValueError: No valid checkpoint found in output directory (outputs)

So I think I'll close this for now, since removing the code will not make it possible to launch the same script in both cases and expect it to work