Added a decode+resize benchmark and cuda decoder #378
Conversation
facebook-github-bot
added
the
CLA Signed
| import torchvision # noqa: F401 | ||
| from torchvision.transforms import v2 as transforms_v2 | ||
|
|
||
| self.torchvision = torchvision |
There was a problem hiding this comment.
I don't think we actually use self.torchvision? We should be able to remove it.
There was a problem hiding this comment.
Good point. done.
| @@ -535,6 +595,8 @@ def run_benchmarks( | |||
| results = [] | |||
| df_data = [] | |||
| verbose = False | |||
| # TODO: change this back before landing. | |||
| min_runtime_seconds = 0.1 | |||
There was a problem hiding this comment.
I have an unmerged change in PR #362 to make this sort of benchmark testing easier: https://github.com/pytorch/torchcodec/pull/362/files#diff-c378bd5d03e7daa116eaeeeb86921e8134f0feefb0fcdf9f020f872676c5c00dR31-R34
| frame = next(reader) | ||
| frames.append(frame["data"].permute(1, 2, 0)) | ||
| frames = [frame.to(device) for frame in frames] | ||
| frames = self.transforms_v2.functional.resize(frames, (height, width)) |
There was a problem hiding this comment.
Naive question that applies to all implementations that use the transformation: how do we ensure it's done on the GPU?
Realized after walking away from my laptop that it's controlled by where the data lives, not by some parameter on the transform. So to answer my question: in the frame.to(device) call.
|
Ideally we would also update the README description, as once this PR is merged, the chart is live. But we can also do that on a follow-up. That is, a 22-core Linux system with whatever kind of GPU. |
|
I tweaked the README.md file as well. Later on we can put the machine info in the chart itself so this doesn't have to be kept manually up-to-date. |
Benchmark results show cuda decoder is faster than CPU decoder at dataloader style benchmark.