Paweł Potrykus

There is an option to test a trained model. Assuming its weights are already loaded you can use for fastai: ```python learner.validate(test_dl) ``` for pytorch-lightning ```python result = trainer.test(lightning_model, test_dl)...

I'd say this could be a FAQ so we should think of showcasing this in a notebook linked on a website.

The current adapter for fastai still returns a scalar. What's worse, in case of COCOMetric it only keeps track of the first one: ```python class FastaiMetricAdapter(fastai.Metric): def __init__(self, metric: Metric):...

We have `model.param_groups` return List of List, which cannot be passed directly to a regular torch Optimizer init cause it expects a list of dicts and fails upon conversion of...

this needs to be fixed in all examples that use pytorch lightning, keeping this one in the TODO

Any idea why black formatting check might failing here?

I had such a use case, but cannot this be done with: ```python inference_metric = COCOMetric(print_summary=True) inference_metric.accumulate(preds) inference_metric_result = inference_metric.finalize() ```

But looking at it again, it's the same number of lines of code and the `metric.accumulate` method is exactly for this purpose. Take a look at the tests here: https://github.com/airctic/icevision/blob/9c17564b82b2c17367ee2eb02fd24dbf8bf6376c/tests/models/efficient_det/test_metrics.py#L63...

some early results for bbox unaware mosaic 

bbox aware mosaic with enforced scaling