Error on unpickling TextLogger after deleting directory where text log was saved.

[man2machine]

🐛 Describe the bug
Summary: When reloading a strategy checkpoint using maybe_load_checkpoint, the loggers for that strategy are unpickled, which calls TextLogger._fobj_deserialize. If the strategy saved text logs to a directory that no longer exists (for example in a temporary directory), then when loading the checkpoint you get FileNotFoundError for any file definition of the form "path:<path to file>".

🐜 To Reproduce

os.makedirs("logs", exist_ok=True)
log_fname = "logs/{}.log".format(datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
 
text_logger = TextLogger(file=open(log_fname, 'w'))
eval_plugin = EvaluationPlugin(
    loggers=[text_logger]
)

strategy = Naive(
    model=model,
    optimizer=optimizer,
    criterion=criterion,
    evaluator=eval_plugin,
)

save_checkpoint(
    strategy,
    "checkpoint.pt"
)

shutil.rmtree("logs")

strategy, _ = maybe_load_checkpoint(
    strategy,
    "checkpoint.pt"
)

🐝 Expected behavior
The log directory and file should be created. This is special pickling behavior from avalanche, hence ideally it should either handle all the pickling cases properly or not pickle the text logger.

[AntonioCarta]

Right now loggers add metrics to existing files. The idea is that you want the file to have the complete history. If the file does not exist this is not possible. To me this is expected behavior. However, I'm open to an alternative that starts the log from scratch if the file is not available anymore, as long as it doesn't add too much complexity.