search

autonomousvision / differentiable_volumetric_rendering

This repository contains the code for the CVPR 2020 paper "Differentiable Volumetric Rendering: Learning Implicit 3D Representations without 3D Supervision"
http://www.cvlibs.net/publications/Niemeyer2020CVPR.pdf
MIT License
794 stars 90 forks source link

How to train a new model #48

Closed szk1509 closed 3 years ago

szk1509 commented 3 years ago

Hi, first of all I want to thank you for sharing your work :smiley: on the other hand, after several attempts, I was wondering what I was missing/misinterpreting on the default.yaml file to train a new model.

The structure of my dataset looked like this:

The config.yaml: data: dataset_name: DataSetX path: data/DataSetX ... input_type: image ... n_views: 0 n_views_input: 0 pointcloud_chamfer_file: pointcloud.npz ... img_folder: image img_folder_input: img_choy2016 ... img_extension: png img_extension_input: jpg ... train_split: train val_split: val test_split: test ...

Questions:

  1. What are the differences btw. img_folder and img_folder_input? After modifying both values independently I still got the following error: "Error occurred when loading field img of model rgb (05)"

  2. Would you recommend to change the number of views value? or it is not strictly necessary?

  3. Since my training set, validation set and cad models are separated in folders, do you consider it easier to change the structure to the structure presented by shapenet?

Thank you for your help, Lizeth

m-niemeyer commented 3 years ago

Hi,

I would suggest you download one of our datasets (e.g. like you mentioned the ShapeNet one), and you write a script which converts the data from your format to our format. Then it should be much easier to adapt the config to your needs - you rather change small things like the image folder name or the image extension in the config, but not structural things like what you mention.

Good luck!

daviszhu88 commented 7 months ago

I got the same problem when trying the train.py, with DTU data. I really think there should be something work that people can reproduce the results.