colab error, pytorch lighting trainer

[softmurata]

I tried to do pix2pix training scripts on google colab, but I met the following error and I could not start training...

TypeError: on_train_batch_start() missing 1 required positional argument: 'dataloader_idx'

I use pytorch-lightning==1.9.2 and torch1.13.1+cuda11.6
Does this error occur because of newer pytorch-lighting version?
I am looking forward to your reply.

[Mikubill]

You might need to change some code to make it work with higher version of ptl:
(remove dataloader_idx param)

ControlNet/ldm/models/diffusion/ddpm.py

Line 591 in 5b9908e

def on_train_batch_start(self, batch, batch_idx, dataloader_idx):

and

ControlNet/cldm/logger.py

Line 74 in 5b9908e

def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):

Works with ptl=1.9.1 and torch=1.13: https://wandb.ai/nyanko-devs/control-net-test/runs/3d0rf41x

[softmurata]

Oh, Thanks for your reply
I made breakthrough this error part, but I met session crash...., I could not do pix2pix training on colab at free..

[softmurata]

How much memory does pix2pix training require?
I could not work on 16GB GPU... with config.py changed (save_memory=True) and batch_size = 1...

[softmurata]

sorry I could do that. Only download this font
cd ControlNet mkdir data wget https://github.com/prawnpdf/prawn/blob/master/data/fonts/DejaVuSans.ttf -P data
and please go to train phase.
Thanks!!

[MasterScrat]

Would be nice to add these changes to main?

You might need to change some code to make it work with higher version of ptl: (remove dataloader_idx param)

ControlNet/ldm/models/diffusion/ddpm.py

Line 591 in 5b9908e

def on_train_batch_start(self, batch, batch_idx, dataloader_idx):

and

ControlNet/cldm/logger.py

Line 74 in 5b9908e

def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):

Works with ptl=1.9.1 and torch=1.13: https://wandb.ai/nyanko-devs/control-net-test/runs/3d0rf41x