# Wandb.watch with PyTorch Lightning not logging
**URL:** https://community.wandb.ai/t/wandb-watch-with-pytorch-lightning-not-logging/2589
**Category:** W&B Help
**Tags:** dashboard, wandb
**Created:** [June 9, 2022, 10:56pm UTC](https://community.wandb.ai/t/wandb-watch-with-pytorch-lightning-not-logging/2589 "2022-06-09T22:56:34Z")
**Posts on this page:** 3
**Page:** 1
### Author:  [@dminn](https://community.wandb.ai/u/dminn)
#### Post date: [June 9, 2022, 10:56pm UTC](https://community.wandb.ai/t/wandb-watch-with-pytorch-lightning-not-logging/2589/1 "2022-06-09T22:56:34Z")
Hi,
I moved all of my calls to wandb from the training loop to PyTorch Lightning (PL)'s `Callback` module. All of my `wandb.log()` calls are working properly, but the gradients and parameter tabs in my wandb dashboard are empty. I checked two threads:
- Wandb.watch with pytorch not logging anything
- When is one supposed to run wandb.watch so that weights and biases tracks params and gradients?
For the first thread, the link to the run has expired and I don’t fully understand the context of the solution "… was using `forward()` instead of ` __call__ ()`".
For the second thread, `wandb.log` is getting called after PL’s Callback hook [`on_train_batch_end`](https://pytorch-lightning.readthedocs.io/en/latest/extensions/callbacks.html#on-train-batch-end), so `wandb.log` should be getting called after a backward pass.
Below is a portion of the code defined in the Callback Module. At the start of training (`on_fit_start`) I initialize the wandb run and call `wandb.watch`. And after a batch is completed, (`on_train_batch_end`) I log all the metrics.
```auto
Class PatentLoggerCallback(Callback):
# Omitted non-relevant code
def on_fit_start(self, trainer, pl_module):
wandb.init(project=self.project,
config=pl_module.hparams,
dir=self.save_dir)
wandb.watch(pl_module,
criterion=torch.nn.functional.binary_cross_entropy_with_logits,
log='all',
log_freq=10,
log_graph=True)
def on_train_batch_end(
self,
trainer: Trainer,
pl_module: LightningModule,
outputs: Sequence,
batch: Sequence,
batch_idx: int,
dataloader_idx: int,
) -> None:
metrics = outputs['metrics']
for metric, value in metrics.items():
wandb.log({f'train/{metric}': value})
```
I would like to have produced a google collab for reproducibility, but there is a lot of code involved. The next best thing I can offer is this [Jupyter Notebook](https://github.com/DennisMinn/patent-phrase-matching/blob/nakama/Workspace.ipynb) that runs through my entire code. I’m not expecting you to clone to repository, but if you do make sure you’re on the “nakama” branch.
---
### Author:  [@dminn](https://community.wandb.ai/u/dminn)
#### Post date: [June 10, 2022, 12:12am UTC](https://community.wandb.ai/t/wandb-watch-with-pytorch-lightning-not-logging/2589/2 "2022-06-10T00:12:42Z")
**RESOLVED**
In my model’s forward pass I wrote `return torch.rand(input_ids.shape[0], requires_grad=True)` to faster debug other `wandb.log` calls by skipping expensive computation. B/c none of the model’s parameters were being updated, there was no backward pass over the model being watched. The only thing that underwent the backward pass was the `torch.rand()` hence no histograms.
---
### Author:  [@system](https://community.wandb.ai/u/system)
#### Post date: [August 9, 2022, 12:13am UTC](https://community.wandb.ai/t/wandb-watch-with-pytorch-lightning-not-logging/2589/3 "2022-08-09T00:13:37Z")
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.