Video support about ffcv HOT 2 CLOSED

To support additional data types one needs to create a new Field and its corresponding decoder, all the available ones are in the ffcv/fields folder. The only reason FFCV doesn't support video yet is because our lab doesn't regularly work with them. We are happy to integrate pull requests that would add that feature.

Parallelizing with FFCV is generally not an issue. As one can see in https://github.com/libffcv/ffcv/blob/main/ffcv/fields/rgb_image.py#L122 it's just a matter of using the result of Compiler.get_iterator instead of range and marking the code as is_parallel.

A sequence of jpeg would work and I could definitely see it as an option but I'm not sure it's the most space efficient method. Most video codecs leverage the information about the previous image to encode the next to save space.

If you are interested in adding Video support you can join our slack and we can guide you/help you troubleshoot.

from ffcv.

Closing due to inactivity. Feel free to submit a PR if you happen to have implemented it. We unfortunately don't have the bandwidth to add more Fields and will focus on improving FFCV's core

from ffcv.