def test_efficientdet_metrics(
fridge_efficientdet_model,
fridge_efficientdet_records,
metric,
expected_output,
request,
):
expected_output = request.getfixturevalue(expected_output)
fridge_efficientdet_model.eval()
batch, records = efficientdet.build_valid_batch(
fridge_efficientdet_records)
raw_preds = fridge_efficientdet_model(*batch)
preds = efficientdet.convert_raw_predictions(
batch=batch,
raw_preds=raw_preds["detections"],
records=fridge_efficientdet_records,
detection_threshold=0.0,
)
metric.accumulate(preds)
with CaptureStdout() as output:
metric.finalize()
assert output == expected_output
def after_pred(self):
self.learn.yb = [self.learn.xb[1]]
self.learn.xb = [self.learn.xb[0]]
if not self.training:
preds = efficientdet.convert_raw_predictions(
self.pred["detections"], 0)
self.learn.converted_preds = preds
def after_pred(self):
self.learn.yb = [self.learn.xb[1]]
self.learn.xb = [self.learn.xb[0]]
if not self.training:
preds = efficientdet.convert_raw_predictions(
batch=(*self.xb, *self.yb),
raw_preds=self.pred["detections"],
records=self.learn.records,
detection_threshold=0.0,
)
self.learn.converted_preds = preds
def validation_step(self, batch, batch_idx):
(xb, yb), records = batch
with torch.no_grad():
raw_preds = self(xb, yb)
preds = efficientdet.convert_raw_predictions(
raw_preds["detections"], 0)
loss = efficientdet.loss_fn(raw_preds, yb)
self.accumulate_metrics(records, preds)
for k, v in raw_preds.items():
if "loss" in k:
self.log(f"valid/{k}", v)
def test_plot_confusion_matrix(fridge_efficientdet_model,
fridge_efficientdet_records):
fridge_efficientdet_model.eval()
batch, records = efficientdet.build_valid_batch(
fridge_efficientdet_records)
raw_preds = fridge_efficientdet_model(*batch)
preds = efficientdet.convert_raw_predictions(
batch=batch,
raw_preds=raw_preds["detections"],
records=fridge_efficientdet_records,
detection_threshold=0.0,
)
cm = SimpleConfusionMatrix()
cm.accumulate(preds)
cm.finalize()
cm.plot()