anton-jeran / mesh2ir

This is the official implementation of our mesh-based neural network (MESH2IR) to generate acoustic impulse responses (IRs) for indoor 3D scenes represented using a mesh.

Hi there 👋

This is Anton.

• 🔭 I’m currently a 5th year PhD student at the University of Maryland
• 👯 I’m looking to collaborate on audio processing and machine learning
• 📫 How to reach me: [email protected]
• 😄 Pronouns: He/Him/His

mesh2ir's Issues

How to generate the csv data inside **Paths** folder for a new given scene?

Very nice work and thank you very much for sharing.
Read.me has a sentence
``
For 3 different indoor scenes, we have stored sample source-recevier locations in a csv format inside Paths folder.
"
The question I want to ask is how is this data obtained?
How to generate the csv data inside Paths folder for a new given scene?
Are there any tools or scripts to share?
Thank you.

Would you provide the ckpt file of the model？

Very wonderful work, would you provide the ckpt file of the model？
Thanks.

The size of the RIR generated

Very nice work and thank you very much for sharing.
The size of the RIR generated by "python3 evaluate.py" is very different from the size of the corresponding RIR of the training dataset (or GWA dataset). What's going on? Would you like to help?
Thank you!

Size of RIR generated is 15.6KB

path: "1aa91215-cba7-4c40-8b37-6b21584b5924/hybrid/L3_R0012.wav"

Size of RIR in GWA or training dataset is 281.4KB

path:"GWA_Dataset_small/1aa91215-cba7-4c40-8b37-6b21584b5924/L3_R0012.wav"

Training Backward propagation is slow

Hi, I am trying to reproduce the mesh2ir following the training parts. However, I found that the backward step is quite slow.
I tried cuda11.0 / cuda 10.2 and added torch.backends.cudnn.benchmark = True.
For batch size=256, it takes about 5 seconds for 1 batch training (includes 3 generation steps, which are the most time-consuming). (GPU is V100)
Could you please let me know if this is reasonable or if there is anything I need to modify?