F.interpolate not compatible with bfloat16

[Warvito]

As mentioned in pytorch/pytorch#86679, F.interpolate is not compatible with bfloat.

It is necessary to convert to float32 its usage in

GenerativeModels/generative/networks/nets/autoencoderkl.py

Line 48 in 5f0fd39

x = F.interpolate(x, scale_factor=2.0, mode="nearest")

and

GenerativeModels/generative/networks/nets/diffusion_model_unet.py

Line 518 in 9cd35d0

x = F.interpolate(x, scale_factor=2.0, mode="nearest")

[Warvito]

Related to #133

[danieltudosiu]

You can incapsulate it in a method that you decorate so that amp is not active inside it. Similar with the VQ layer. Best regards, Petru-Daniel (Dan) Tudosiu Biomedical Engineering PhD Student School of Biomedical Engineering & Imaging Sciences King's College London Mobile: +447497517619 Email: ***@***.*** Office Address: 9th Floor, Becket House, 1 Lambeth Palace Rd, South Bank, London, SE1 7EU

…

________________________________ From: Walter Hugo Lopez Pinaya ***@***.***> Sent: Saturday, December 24, 2022 1:22:11 PM To: Project-MONAI/GenerativeModels ***@***.***> Cc: Subscribed ***@***.***> Subject: Re: [Project-MONAI/GenerativeModels] F.interpolate not compatible with bfloat16 (Issue #156) Related to #133<#133> — Reply to this email directly, view it on GitHub<#156 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ACQEEVARJHGM5CRY3QNAVI3WO3MGHANCNFSM6AAAAAATIM4CX4>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Warvito]

Hi @danieltudosiu , here I meant bfloat16. It works fine with amp