AttributeError with LightningModule forward without Trainer

[lucmos]

🐛 Bug

LightningModule with Lightning 1.4 assumes to always have the self.trainer not None. No issues with 1.3

To Reproduce

import pytorch_lightning as pl


class MyModule(pl.LightningModule):
    def forward(self, *args, **kwargs):
        self.log_dict({"key": 0})


print(pl.__version__)

model = MyModule()
model()

Outputs:

1.4.0
Traceback (most recent call last):
  File "/home/luca/Projects/CookieTesting/my-new-project/src/my_new_project/commands/mymodule.py", line 11, in <module>
    model()
  File "/home/luca/miniconda3/envs/my-new-project/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/luca/Projects/CookieTesting/my-new-project/src/my_new_project/commands/mymodule.py", line 6, in forward
    self.log_dict({"key": 0})
  File "/home/luca/miniconda3/envs/my-new-project/lib/python3.8/site-packages/pytorch_lightning/core/lightning.py", line 507, in log_dict
    self.log(
  File "/home/luca/miniconda3/envs/my-new-project/lib/python3.8/site-packages/pytorch_lightning/core/lightning.py", line 407, in log
    results = self.trainer._results
AttributeError: 'NoneType' object has no attribute '_results'

Expected behavior

If the Trainer and/or logger are not defined, the log and log_dict calls should be only ignored.

Environment

* CUDA:
        - GPU:
        - available:         False
        - version:           10.2
* Packages:
        - numpy:             1.21.2
        - pyTorch_debug:     False
        - pyTorch_version:   1.9.0
        - pytorch-lightning: 1.4.0
        - tqdm:              4.62.3
* System:
        - OS:                Linux
        - architecture:
                - 64bit
                - ELF
        - processor:         x86_64
        - python:            3.8.11
        - version:           #40~20.04.1-Ubuntu SMP Sat Sep 18 02:14:19 UTC 2021

Additional context

The problem is caused by this line:

https://github.com/PyTorchLightning/pytorch-lightning/blob/ab069876cb19bb9de0179f74c6f83764876a73ff/pytorch_lightning/core/lightning.py#L398

Related PR: #7891

@carmocca

[rohitgr7]

self.log_dict({"key": 0})

log_dict is something specific to lightning which is handled by trainer and loggers. So I guess a trainer is required here?

[lucmos]

My use case is to re-use the same LightningModule code for training (with the trainer and loggers) and inference/predictions (without the trainer and loggers)

[carmocca]

Previous answer on this:

#8509 (comment)

[rohitgr7]

seems reasonable. just curious, if you are logging inside forward function then how are you differentiating the log keys since forward will be used for training/validating/testing I assume?
you can actually prevent it if you want by simply not calling log_dict during inference using some sort of boolean argument.

[carmocca]

Actually, calling self.log or self.log_dict inside forward is highly discouraged. forward should be kept a pure function to compute the network's outputs. Please call it in one of the *_step functions instead! That's exactly what they are for.

Closing this as using forward for this is not a good practice. Discussion can be continued in #8509

if you are logging inside forward function then how are you differentiating the log keys since forward will be used for training/validating/testing I assume?

It will use the *_step method name that called forward