pretrained model mode (ADV/TB) to skip training

[dgerin]

pretrained model mode (ADV/TB) to skip training

[dgerin]

My first post is incorrect, such mod is done with --load-model="path to model.pt" --inference-only, so my problem now becomes : Where can we find pretrained ADV and TB dlrm model.

[mnaumovfb]

At the moment, we do not provide pre-trained models for Criteo Kaggle Display Advertising Challenge and Terabyte datasets. We have considered this in the past #40, so I'll take this as another request for it.

You should be able to train a model from scratch in a few hours when interfacing with Criteo Kaggle Display Advertising Challenge dataset. However, the Terabyte dataset would take significantly longer to train.

[dgerin]

mnaumovfb, thanks for this confirmation.
I'll be able to train ADV myself but it will be much more difficult
for me to do the same with TB. I just finish day_0 downloading and it takes
48 GB after gzip, 2 hour to load splits with data API, and finally I don't have descent GPU
for training. It would be very nice to provide a trained .pt for TB ? Maybe I'm quite marginal when I say that I'm only interested in inference accuracy, but I don't understand the reason why
such pretrained models are not provided, especially when it requires such efforts.

[mnaumovfb]

We will work on providing pre-trained models. Give us a bit of time for it.

I'll let you know when they become available.

[dgerin]

Super mnaumovfb, thank you.
Should I close the issue ?

[mnaumovfb]

Let's keep it open for now. We will close once the pre-trained models become available.

[mnaumovfb]

I have added a link to pre-trained models to the README. They are available under CC-BY-NC license.

[dgerin]

Thank you mnaumovfb.