Unable to add (manage) certain columns in project runs

W&B Help
1

Hi,
I’m fairly new to wandb (and AI in general). I’m currently working with a simple script that runs a ResNet18 for image classification purposes. I’ve been adapting a script that originally logged many infos (train_epoch_loss, valid_accuracy, etc.) and with which I was able to customize the different columns into the project runs panel.

However, after trying really hard to find a way to track GPU usage, I ended up changing the script quite a bit. Everything works fine but now I am unable to customize most of the columns (they stay in the hidden columns no matter what). I have noticed that the icon next to the column name is not the same when I cannot add it to the visible ones (see example below). What am I doing wrong ?

image
image871×717 34.9 KB

I am also including an example of the way I pass the infos in my script:

# Configuration WandB
        config_dict = {
            "batch_size": self.config.batch_size,
            "learning_rate": self.config.learning_rate,
            "momentum": self.config.momentum,
            "seed": self.config.seed,
            "epochs": self.config.epochs
        }
       
        # Ajout des informations GPU si disponible
        if torch.cuda.is_available():
            config_dict.update({
                "gpu_name": torch.cuda.get_device_name(0),
                "gpu_memory_gb": torch.cuda.get_device_properties(0).total_memory / 1024**3,
                "cuda_version": torch.version.cuda,
                "pytorch_version": torch.__version__
            })
       
        wandb.config.update(config_dict)
 
 ...
# Log WandB par epoch avec métriques GPU détaillées
            epoch_metrics = {
                "train_epoch_accuracy": 100. * train_acc,
                "train_epoch_loss": train_loss,
                "valid_accuracy": 100. * valid_acc,
                "valid_loss": valid_loss,
                "epoch": epoch
            }
           
            # Ajout des métriques GPU de fin d'epoch
            if torch.cuda.is_available():
                epoch_metrics.update({
                    "gpu/memory_allocated_epoch_end": torch.cuda.memory_allocated() / 1024**3,
                    "gpu/memory_reserved_epoch_end": torch.cuda.memory_reserved() / 1024**3,
                    "gpu/memory_max_allocated_epoch": torch.cuda.max_memory_allocated() / 1024**3
                })
 
            if wandb.run is not None:
                wandb.log(epoch_metrics)