Dataloading Revamp

[AntonioMacaronio]

Problems and Background

• With a sufficiently large enough dataset, the current parallel_datamanager.py will try to cache the entire dataset into RAM, which will lead to an OOM error
• parallel_datamanager.py only uses one worker to generate ray bundles. Since various subprocesses such as unprojecting during ray generation, or pixel sampling within a custom mask can be a CPU-intensive task, it may be better suited to parallelize this. While parallel_datamanager.py does support multiple workers, each worker caches the entire dataset to RAM and it does not support massive datasets, leading to duplicate copies of the dataset in computer memory. It also implements parallelism from scratch and is not friendly to build off.
• Additionally, both VanillaDataManager and ParallelDataManager rely on CacheDataloader, which subclasses torch.utils.data.DataLoader, which is a strange coding practice
• Similarly for full_images_datamanager.py: As we can not fit the entire dataset in RAM, the current implementation loads in entire dataset into the FullImageDataloader's cached_train attribute. To do this efficiently, we need multiprocess parallelization to load in a batch of images (support for batched image dataloading since gsplat now supports batched rastuerization)

Overview of Changes

• Replacing CacheDataloader with RayBatchStream, which subclasses torch.utils.data.IterableDataset. The goal of this class is to generate ray bundles directly without caching all images to RAM. This is done by collating a sampled batch of images to sample from. A new ParallelDatamanager class is written which is available side-by-side but can completely replace the original VanillaDatamanager
• Adding an ImageBatchStream to create a parallel, OOM-resistant version FullImageDataloader called ParallelFullImageDataloader which is defined inside parallel_full_images_datamanager.py
• A new pil_to_numpy() function is added. This function reads a PIL.Image's data buffer and fills an empty numpy array while reading, hastening the conversion process and removing an extra memory allocation. It is the fastest way to get from a PIL Image to a Pytorch tensor averaging ~2.5ms for a 1080x1920 image (~40% faster)

Impact

• Checkout these comparisons! The left was trained on 200 images of a 4k video, while the right was trained on 2000 images of the same 4k video.

[pwais]

nice progress! sorry its not fast but i think i know why:

i think the main reason this is slower than expected is because _get_collated_batch() gets called per raybundle and sadly _get_collated_batch() is AFAIK needlessly slow.

• take note about how the current CachedDataloader avoids doing _get_collated_batch() per raybundle. it would have been nice for the author to have left some notes about how slow _get_collated_batch() is, but evidently that author found it's necessary to not collate images per raybundle .
• in my impl, I just _get_collated_batch() once on a small set of images an keep that batch cached. the main problem I saw is that _get_collated_batch() on thousands of images seemed to use 2x or 3x as much RAM as actually needed and thus cause many minutes of swapping and stuff

Even if you only call _get_collated_batch() once tho, you might need a bigger prefetch factor and/or more workers depending on the model.

IMO it's worth trying to find a way to get the result of nerfstudio_collate on cameras (I think the cameras do need to be collated because they can be ragged? i could be wrong and they don't need collation) but on images just have the worker read image files / buffers and never call collate on those tensors.

Just to be clear, this is the line where collate on images can go nuts and start taking forever to allocate 200GB or more of RAM for many images in code in main:

nerfstudio/nerfstudio/data/utils/nerfstudio_collate.py

Line 101 in 3d27dd4

storage = elem.storage()._new_shared(numel, device=elem.device)

So! If a worker is just emitting raybundles then the images never need to be in shared tensor memory then eh? Thus should be able to save some RAM and CPU by skipping that line for images. Still need to think about the cost of reading the images themselves, but collate is definitely a troublemaker.

[pwais]

FWIW for a 3090 i was using 16 workers and prefetch factor of 16, and a train ray batch size of 24000. And I was getting the same "rays per sec" in the console output or better as with the in-repo impl (ParallelDataManger). Steady-state I was using less than 16 CPU I believe. If batch size and prefetch is small, then definitely need more workers

[pwais]

just took a quick look (can't do a full review right now), so cool to see this coming along!!

Sounds like this change will target the case that uncompressed image tensors can't fit in RAM, but the raw image files (typically jpeg) do fit in RAM. In that case I guess we do want each worker to literally load the file bytes into Python RAM (as implemented) versus let the OS disk cache work, because the idea is that the uncompressed image tensors will otherwise blow out the disk cache.

I think it would be important to test in the end like a case where the user only has limited RAM (say 16GB) and e.g. a 8GB laptop graphics card, in that case I think there are moderate or larger image datasets where the whole thing would OOM when using the current cache impl. In that case, it would be helpful to have some way to disable the cache, or just communicate to the user that they simply have too weak of a machine for the dataset (e.g. just a CONSOLE.print("[bold yellow]Warning ...") in the line where the workers start reading image files into RAM.

[pwais]

nice work keep going!