benchmark ssdlite detection pipeline #3
Conversation
results/20230403073901.log
Outdated
| Summaries | ||
|
|
||
| v2 / v1 | ||
| Tensor 1.66 | ||
| PIL 1.54 | ||
|
|
||
| x / PIL, v1 | ||
| Tensor, v1 0.92 | ||
| Tensor, v2 1.53 | ||
| PIL, v1 1.00 | ||
| PIL, v2 1.54 | ||
| Datapoint, v2 1.52 |
transforms.py
Outdated
| wrapper_factory = WRAPPER_FACTORIES[datasets.CocoDetection] | ||
| mock_dataset = SimpleNamespace(ids=list(range(num_samples))) | ||
| wrapper = wrapper_factory(mock_dataset) | ||
| self.wrapper = functools.partial(wrapper, num_samples // 2) |
There was a problem hiding this comment.
Why num_samples // 2 and why is it hardcoded to 117_266 above?
There was a problem hiding this comment.
This whole thing is mocking the dataset so we can re-use the dataset wrappers as transform. num_samples = 117_266 is the number of samples in the filtered COCO dataset so each sample has one annotation.
In the wrapper, we get passed the idx of the sample
and use it like this:
Since we don't have access to the idx here, I chose a proxy. Since ids is a list, I used the half of the maximum, to have on average the same performance as doing the correct lookup.
| @@ -23,6 +23,10 @@ def write(self, message): | |||
| self.stdout.write(message) | |||
| self.file.write(message) | |||
|
|
|||
| def flush(self): | |||
transforms.py
Outdated
| return image, target | ||
|
|
||
|
|
||
| class DetectionReferenceRandomHorizontalFlipV1(transforms_v1.RandomHorizontalFlip): |
There was a problem hiding this comment.
In order to ease per-transform comparison visually, perhaps keep the shorter name RandomHorizontalFlipV1 (same for other transforms).
Ideally we would even drop the V1/V2 suffix
|
We identified the slowdown from detection v2 stems from two things:
Results here are run on torchvision nightly April 4. When pytorch/vision#7488 is merged and pytorch/vision#7489 addressed, we need to re-run. |
No description provided.