Trainer's fit method fails when the only logger is the custom logger from the official documentation

[pgagarinov]

🐛 Bug

Trainer's fit method fails when the only logger is the custom logger from the official documentation

To Reproduce

1. Define a custom logger as per https://pytorch-lightning.readthedocs.io/en/latest/logging.html#make-a-custom-logger
   logger = MyLogger()
2. Define a lightning module as per https://pytorch-lightning.readthedocs.io/en/latest/new-project.html#step-1-define-lightningmodule
3. Load MNIST and call training as per https://pytorch-lightning.readthedocs.io/en/latest/new-project.html#step-2-fit-with-lightning-trainer with ONE MODIFICATION - specify Trainer(gpus=1, logger = logger)

Expected behavior

Successful execution

Environment

• CUDA:
  - GPU:
  - GeForce RTX 2070
  - available: True
  - version: 10.2
• Packages:
  - numpy: 1.19.2
  - pyTorch_debug: False
  - pyTorch_version: 1.6.0
  - pytorch-lightning: 1.0.3
  - tqdm: 4.50.2
• System:
  - OS: Linux
  - architecture:
  - 64bit
  - ELF
  - processor:
  - python: 3.8.5
  - version: Proposal for help #1 SMP PREEMPT Sat Oct 17 11:06:19 UTC 2020

Additional context

Stack trace:

`
PU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

---

TypeError Traceback (most recent call last)
in
5 # trainer = pl.Trainer(gpus=8) (if you have GPUs)
6 trainer = pl.Trainer(gpus=1, logger = logger)
----> 7 trainer.fit(autoencoder, train_loader)

~/.conda/envs/dlschool0/lib/python3.8/site-packages/pytorch_lightning/trainer/trainer.py in fit(self, model, train_dataloader, val_dataloaders, datamodule)
437 self.call_hook('on_fit_start')
438
--> 439 results = self.accelerator_backend.train()
440 self.accelerator_backend.teardown()
441

~/.conda/envs/dlschool0/lib/python3.8/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py in train(self)
49
50 # set up training routine
---> 51 self.trainer.train_loop.setup_training(model)
52
53 # train or test

~/.conda/envs/dlschool0/lib/python3.8/site-packages/pytorch_lightning/trainer/training_loop.py in setup_training(self, model)
144 # --------------------------
145 # on pretrain routine start
--> 146 self.trainer.on_pretrain_routine_start(ref_model)
147 if self.trainer.is_function_implemented('on_pretrain_routine_start'):
148 ref_model.on_pretrain_routine_start()

~/.conda/envs/dlschool0/lib/python3.8/site-packages/pytorch_lightning/trainer/callback_hook.py in on_pretrain_routine_start(self, model)
120 """Called when the train begins."""
121 for callback in self.callbacks:
--> 122 callback.on_pretrain_routine_start(self, model)
123
124 def on_pretrain_routine_end(self, model):

~/.conda/envs/dlschool0/lib/python3.8/site-packages/pytorch_lightning/callbacks/model_checkpoint.py in on_pretrain_routine_start(self, trainer, pl_module)
159 When pretrain routine starts we build the ckpt dir on the fly
160 """
--> 161 self.__resolve_ckpt_dir(trainer, pl_module)
162
163 def on_validation_end(self, trainer, pl_module):

~/.conda/envs/dlschool0/lib/python3.8/site-packages/pytorch_lightning/callbacks/model_checkpoint.py in __resolve_ckpt_dir(self, trainer, pl_module)
405 version, name = trainer.accelerator_backend.broadcast((version, trainer.logger.name))
406
--> 407 ckpt_path = os.path.join(
408 save_dir, name, version, "checkpoints"
409 )

~/.conda/envs/dlschool0/lib/python3.8/posixpath.py in join(a, *p)
88 path += sep + b
89 except (TypeError, AttributeError, BytesWarning):
---> 90 genericpath._check_arg_types('join', a, *p)
91 raise
92 return path

~/.conda/envs/dlschool0/lib/python3.8/genericpath.py in _check_arg_types(funcname, *args)
150 hasbytes = True
151 else:
--> 152 raise TypeError(f'{funcname}() argument must be str, bytes, or '
153 f'os.PathLike object, not {s.class.name!r}') from None
154 if hasstr and hasbytes:

TypeError: join() argument must be str, bytes, or os.PathLike object, not 'method'
`

[github-actions]

Hi! thanks for your contribution!, great first issue!

[edenlightning]

Hey! Can you please send a colab so we can reproduce with your custom logger?

[pgagarinov]

@edenlightning
Sorry for not making this clear. The logger I used is the one from the official documentation i.e. this is not "my custom logger"):

[rohitgr7]

@pgagarinov I think this is just a template. You need to write the method definitions.

[pgagarinov]

@rohitgr7 Based on the documentation it is a valid logger all by itself as it is not obligatory for log_* methods to do anything. The "experiment" method could benefit from returning something but the fact is that the execution fails even before "experiment" method is called.

If there are some hidden assumptions about the way "a valid" logger should look like - I would say those need to be documented and the example on "Custom Logger" page better be changed...

[rohitgr7]

version, name = trainer.accelerator_backend.broadcast((version, trainer.logger.name))

breaks here.
looks like some decorators on the methods in the docs are missing. wanna do a PR and fix the docs? For reference, you can have a look at TensorboardLogger. @pgagarinov