Saving of models

[lucidrains]

Saving of models

[lucidrains]

@dave7895 models are periodically saved at intervals of 1k iterations, and you can find it under ./models/{project name}. they should be never overwritten, and you can even go back in time and load previous models along the way with --load-from {checkpoint num}

[dave7895]

I understand. Sorry for my possibly badly formulated question. I meant: If I train 990 iterations and then interrupt the script, are all the 990 iterations wasted or are they in the most recent model and if I continue there are basically 1990 iterations in this checkpoint added?

[lucidrains]

@dave7895 ahh, they are wasted, it's too expensive to save every iteration. however, you can change the frequency at which it saves with a --save-every flag

[dave7895]

Yeah, okay I can understand this. Did this change with a more recent release as I hadn't updated my package and so the --save-every flag did not work. But I did not know this and so I trained it about 15k iterations but always interrupted before the checkpoint as I did not know it was at 10k and not 1k. These are the ones named automatically 0.jpg, so the ones before the first checkpoint.

That's why I am confused because I parallely train on colab and there the ones from before the 1st checkpoint look like this:

[lucidrains]

oh sorry I think I misunderstood. If you run the same command with the same project name (or leave at default), it will pick up where you left off

[lucidrains]

Oh and yup, save-every was added in recent updates!

[dave7895]

Thanks for the clarification. Where does this get saved then?

[lucidrains]

@dave7895 it's saved to the ./models folder where you executed the command

[dave7895]

Ok so just updating and then duplicating with only one evolving at the 1k mark? Is this happening when the images are updated?

[lucidrains]

@dave7895 yup! it happens every time the images are updated

[dave7895]

Ok. Thanks for your time.