Create ML 2.0 - Not enough disk space to train Image Classification model

Question

Created Nov ’20

Replies 5

Boosts 0

Views 2.1k

Participants 6

I am attempting to train an Image Classification model in Create ML 2.0 (53.1) on macOS Big Sur 11.0.1.
My MacBook Air (from year 2019, with Intel processor, 8GB memory) has around 25GB free space on the internal disk at the beginning of the training.

The training dataset is relatively big: around 200.000 images taking 20GB disk space, 800 classes. The images are all stored on an external hard drive.

I successfully imported the dataset in Create ML and could start the training. At the stage of extracting image features, however, a temporary file is getting built up and at around 4% of the feature extracting, the 25GB free space on my internal drive gets eaten up and the training stops with an "Unexpected Error".

Is there any way to change the temporary file location for Create ML to my external hard disk or any other way to make sure that the training can be finished even in case of low internal hard disk space available?

Boost

Answer 1

EasonChang OP

Jan ’21

I got the same problem, what interesting is that my MacBook Air has 23GB space are running the image training well but take days. I bought a MacBook Pro M1 come with 16GB, 256GB SSD, but it doesn't work because of not enough space.

0

Answer 2

sdpartners OP

Jan ’21

The diskspace management problem persists in v53.2.2. Unfortunately there is no transparency to how much space might be consumed or where the temporary files are stored. I've found some of them on my mac and with about 2500 images, I'm seeing a minimum of 30M per 10 iterations (but "About this mac" seems to indicate a lot more, like maybe 3-4x as much consumption, i.e., 100M per 10 iterations).

It's understandable why the disk consumption could be so high, but the fact that there's no ability to estimate it ahead of time is a major headache for large models with many iterations.

0

Answer 3

JeanRomain OP

Mar ’21

I got the same "issue" but it's relative to SSD wear.
I use ML Regressor for big boosted tree.
Each training writes approx 10GB on the drive.
Would it be possible to switch this location to external drive, ram or ramdisk? (without installing macOS on external disk)
Maybe it could be implemented in the future?

0

Answer 4

hefeimao OP

May ’21

I experienced similar issues. After train multiple models, the "cache" consumes 280GB in "Other" of my total 512GB SSD. (m1 mac mini)

Finally what seems to be an easy workaround for me is to simply "restart" the mac. That seems to clear "cache" or the temporary files used in training.

0

Answer 5

rensa OP

Jun ’21

Would also love to know if there's an answer to this. I know I can reclaim the space by rebooting, but if I do that I can no longer train the model, which means there's an iteration limit. I'd also prefer not to wear my Mac's SSD down in a few months when it's relatively expensive to get it replaced.

0