Start doc revamp for detection models #5876
Conversation
There was a problem hiding this comment.
Thanks @NicolasHug, I've added a few comments below to get your thoughts. Happy to chat more.
docs/source/conf.py
Outdated
|
|
||
| generate_classification_table() | ||
| generate_weights_table(module=M, table_name="classification", metrics_names=["Acc@1", "Acc@5"]) | ||
| generate_weights_table(module=M.detection, table_name="detection", metrics_names=["box_map"]) |
There was a problem hiding this comment.
I think this is going to put the KeypointRCNN model along with the rest of the models. Unfortunately the specific model was trained with different data from COCO (keypoints), so its focus and validation numbers are the same as for other models.
Have a look on the current documentation here to see how the models are grouped and how extra accuracy metrics are being shown. We don't have to do exactly the same thing but good to get some ideas.
There was a problem hiding this comment.
Is it OK with you if I create an issue to special-case KeypointRCNN? This way I'll still be able to ask for external contributors while I'm thinking of a decent solution for this.
There was a problem hiding this comment.
Yeah I think that would be fine, let's do this.
There was a problem hiding this comment.
Thanks, I opened #5882 to keep track of this
docs/source/conf.py
Outdated
| weights = [w for weight_enum in weight_enums for w in weight_enum] | ||
|
|
||
| column_names = ("**Weight**", "**Acc@1**", "**Acc@5**", "**Params**", "**Recipe**") | ||
| def clean_name(name): |
There was a problem hiding this comment.
Nit: Might be cleaner if you pass metrics_names as Tuple[str, str] where one part of the tuple is the key and the second the doc-friendly name of it.
docs/source/conf.py
Outdated
|
|
||
| generate_classification_table() | ||
| generate_weights_table(module=M, table_name="classification", metrics_names=["Acc@1", "Acc@5"]) | ||
| generate_weights_table(module=M.detection, table_name="detection", metrics_names=["box_map"]) |
There was a problem hiding this comment.
Yeah I think that would be fine, let's do this.
|
Hey @NicolasHug! You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py |
Summary: * Start doc revamp for detection models * Minor cleanup * Use list of tuples for metrics Reviewed By: jdsgomes, NicolasHug Differential Revision: D36095699 fbshipit-source-id: 16314c2d0a13fb233fa87b34925c1502c6047c17
This PR adds table generation for the detection weights, and adds doc for RetinaNet.
Rendered docs: https://output.circle-artifacts.com/output/job/a31ba961-8635-4fbf-8dc0-af4039a5c788/artifacts/0/docs/models_new.html#object-detection-instance-segmentation-and-person-keypoint-detection