How to define an interval validate callbacks in lightning

[rentainhe]

I want to train my model in 20000 steps and run evaluation on each 1000 steps
I try to define the callback like this:

class IntervalStepValidate(Callback):
    def __init__(self, config):
        self.config = config
        self.total_steps = 20000
        self.validation_interval = 1000

    def on_batch_end(self, trainer, pl_module):
        if self.total_steps % self.validation_interval == 0:
            trainer.run_evaluation()

But I find out that there is no run_evaluation() in the latest version of Pytorch-Lightning :(
How can I update this code to get the function I want?

[rohitgr7]

there is a val_check_interval argument for it inside Trainer. You can set Trainer(val_check_interval=1000).

[rentainhe]

You mean, in general way, if I just want to train on a specific batch-size 512, I can set batch-size=256 and gradient-steps=2 to get the same result. 😃

[rentainhe]

Cuz I didn't see any batch split operation in ViT-pytorch/train.py, I was curious about why it works there lolllll!

[rentainhe]

Sorry! it's my fault! He did batch split operation here:
https://github.com/jeonsworld/ViT-pytorch/blob/460a162767de1722a014ed2261463dbbc01196b6/train.py#L147
😢

[rohitgr7]

haha.. np.. it happens all the time, part of the process :)
lightning doesn't alter the batch_size internally, so yeah you have the set it from your end.

[rentainhe]

Thanks a lot ! ❤️