search

xmindflow / DermoSegDiff

[MICCAI 2023] DermoSegDiff: A Boundary-aware Segmentation Diffusion Model for Skin Lesion Delineation
https://arxiv.org/abs/2308.02959
MIT License
57 stars 9 forks source link

Inference speed #2

Closed saskra closed 9 months ago

saskra commented 10 months ago

The inference, i.e. the application of a trained model to test images, is very slow for me. Are there any ways to speed it up?

bravo-hq commented 10 months ago

Dear @saskra,

Thank you for reaching out regarding the performance concerns with inference times in our DDPM-based model.

Just to let you know, the inherent nature of Denoising Diffusion Probabilistic Models (DDPMs) requires the completion of each diffusion step sequentially, which can be time-intensive. The duration of the inference process is influenced by two primary factors: the total number of diffusion steps implemented and the dimensions of the images within your dataset.

We are aware of the importance of efficiency in your work, so we are actively exploring ways to expedite this process. Our team is developing an update to incorporate Denoising Diffusion Implicit Models (DDIMs), which we anticipate will offer improved inference speeds.

In the interim, I highly recommend consulting the following comprehensive survey, which provides an in-depth examination of various diffusion model architectures, specifically addressing the challenge of extended inference times:

Diffusion models in medical imaging: A comprehensive survey

This resource may offer valuable insights and alternative approaches to mitigate the latency you are currently experiencing.

We appreciate your patience and understanding as we endeavor to enhance our model's performance. If you have further questions or need additional assistance, please do not hesitate to contact us.

Sincerely, Yousef Sadegheih