Sharing the checkpoint about roma

[Parskatt]

See here https://github.com/Parskatt/RoMa/blob/main/roma/models/model_zoo/__init__.py

[Parskatt]

I think it should load automatically when you load from the model zoo? You can find the raw urls hardcoded in the repo.

[Cuistiano]

I get it. Thanks!

[YJonmo]

Hi there,

Thanks for this great work. Is there TensorRT version of the weights available somewhere?

Cheers,
Yaqub

[Parskatt]

Hi there,

Thanks for this great work. Is there TensorRT version of the weights available somewhere?

Cheers,
Yaqub

There is not. How can I make that conversion, is there a simple way?

[YJonmo]

There are several ways of doing it:

1. convert the pth to onnx and then onnx to trt format. For example danielgatis/rembg#193

2. convert the pth directly to trt format: https://forums.developer.nvidia.com/t/how-to-directly-convert-a-trained-pytorch-model-into-tensorrt-model-there-are-already-pth-files-ready-after-training/240929/2

[Parskatt]

Cool. Ill give it a try when I have time. Therr is some slightly complicated logic at infrrencr time thst might make it difficult to convert.

[YJonmo]

Thank you. I knew it was not possible to convert the LoFTR into the .trt formant. I hope this is not the same.

[Parskatt]

It might be possible with a simplified inference format (which may perform slightly worse).

…

________________________________ From: YJonmo ***@***.***> Sent: Wednesday, October 11, 2023 8:06:55 AM To: Parskatt/RoMa ***@***.***> Cc: Johan Edstedt ***@***.***>; Comment ***@***.***> Subject: Re: [Parskatt/RoMa] Sharing the checkpoint (Issue #1) Thank you. I knew it was not possible to convert the LoFTR into the .trt formant. I hope this is not the same. — Reply to this email directly, view it on GitHub<#1 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AFIIB7WPRG6JZW6TR6Q6YPDX6YZP7AVCNFSM6AAAAAAYW5GJCWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONJWHA3DGNBXGA>. You are receiving this because you commented.Message ID: ***@***.***>