Provide dataset

[YoannSo]

Hi ! Its an amazing work ! Can you provide some compatible dataset to test your program first ? thanks a lot

[wanmeihuali]

@YoannSo Thanks! For the test, I'm using the tank and temple dataset(truck). See the link here: https://www.tanksandtemples.org/download/ I'm not sure if I can provide an extra download link due to the license. then you can first use colmap to generate a sparse reconstruction, then save it as txt and use this file to generate the two json files.

I can provide your with the generated two JSON files, but you may need to download the images separately.

[YoannSo]

Yes, if you can I'd love to have the json files you've generated thank you!

[wanmeihuali]

@YoannSo Here it is: https://github.com/wanmeihuali/taichi_3d_gaussian_splatting#train-on-tank-and-temple-truck-scene

[YoannSo]

thanks !

[YoannSo]

Hi, i have one error when trainning, maybe something is deprecated on my project ? Im running this in unbuntu and cuda 11.4 with torch 1.12.1 Thanks for your help

[wanmeihuali]

@YoannSo Emm interesting, my environment is CUDA12.1+torch2.0, and it doesn't have such issue... one possible problem is that point_id_in_camera_list is defined as int32, can you try converting it to long? e.g.

self.accumulated_num_in_camera[input_data.point_id_in_camera_list.long()] += 1

maybe int32 type tensor is not supported as index in earlier version?

[YoannSo]

thanks, already tried this on all variable and i got this error futher

[wanmeihuali]

Seems still a problem related to data type? You can try commenting self._plot_grad_histogram out, it's just a logging function for tensorboard that plot gradient distribution, I guess may be it's because the earlier version of torch/tensorboard does not have good support for int32....

[YoannSo]

Ok that's work correctly now, thanks !