Free VRAM after training.

[ExtReMLapin]

Hello,

I symply train my model using

        trainer = trainer_class(cfg)
        trainer.resume_or_load(resume=False)
        trainer.train()
        del trainer

And even after python interpreter left the function, VRAM usage doesn't go down.
Obviously, if I don't call the code block, it doesn't go up.

Is there a way to free VRAM ?

Or is it some kind of internal tensorflow memory managment, which mean it "reserve" vram and internally dispatch memory blocks to whatever calls it from python ?

Thanks.

[ExtReMLapin]

Solution : torch.cuda.empty_cache()

#4007