Improve object detection task guideline #29967
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| @@ -362,7 +363,7 @@ Face to upload your model). | |||
| >>> training_args = TrainingArguments( | |||
| ... output_dir="detr-resnet-50_finetuned_cppe5", | |||
| ... per_device_train_batch_size=8, | |||
| ... num_train_epochs=10, | |||
| ... num_train_epochs=100, | |||
There was a problem hiding this comment.
Did you manage to investigate this? 100 epochs is very large. If we train for this many steps do we still see a change in metrics over the first 10 epochs?
There was a problem hiding this comment.
cc @qubvel perhaps this would be great to investigate as part of making it easier to train object detection models
|
@amyeroberts could we already merge this PR which addresses the first 3 points of #29964? |
|
@NielsRogge Did you run and check re 100 epochs? #29967 (comment) If not, then we should revert back until confirmed |
|
With 100 epochs it converges better but still has pretty low metrics. Also, predictions are bad because we use it for inference with batch size = 1. |
|
@amyeroberts people have reported to replicate it with 100 epochs here: #30557 (comment) Feel free to approve the PR in order to clarify this for now in the docs. Later on we could improve metrics more |
* Add improvements * Address comment
What does this PR do?
This PR fixes some issues mentioned in #29964