how to test the dataset with efficient_ad and torch_inference?

[YuRongHao2233]

If I use torch_inference to load efficient_ad and it will report these.

Traceback (most recent call last):

File "E:\anomalib-main\inference.py", line 117, in
infer(weights="E:\anomalib-main\results\efficient_ad\Zn_eff_03_m\run\weights\torch\model.pt")
File "E:\anomalib-main\inference.py", line 49, in infer
predictions = inferencer.predict(image=image)
File "E:\anomalib-main\src\anomalib\deploy\inferencers\base_inferencer.py", line 84, in predict
output = self.post_process(predictions, metadata=metadata)
File "E:\anomalib-main\src\anomalib\deploy\inferencers\torch_inferencer.py", line 139, in post_process
if isinstance(predictions[1], (Tensor)):
KeyError: 1

[samet-akcay]

@YuRongHao2233 have you set the export mode to torch? By default anomalib saves the model based on Lightning. If you want to use torch inferencer you need to export the model as torch

[YuRongHao2233]

I have exported it as torch

[YuRongHao2233]

I have review the output of efficient_ad,it is a dic,so it can not be used as

if isinstance(predictions[1], (Tensor))

[samet-akcay]

Alright then, we'll have a look.

[blaz-r]

It seems like Efficient AD is returning dictionary, where other models return tensor or tuple. One thing I also noticed is that torch inferencer does have this tuple check for patchcore, while openvino inferencer doesn't, so the score returned by patchcore isn't really accounted for, or am I missing something?

@YuRongHao2233 a quick fix in this case would be to just modify Efficient AD torch model forward function. I believe you can just return combined map in this case and it should work with openvino (this is probably not a good general solution).

[blaz-r]

@alexriedel1 how exactly are the results from forward pass used in case of inference? Are non-combined values even used, because it seems they aren't.

[alexriedel1]

yes thats right, only the combined feature map is used. you can also read it in the torch model

anomalib/src/anomalib/models/efficient_ad/torch_model.py

Lines 302 to 351 in cdedf05

	if self.training:
	# Student loss
	distance_st = reduce_tensor_elems(distance_st)
	d_hard = torch.quantile(distance_st, 0.999)
	loss_hard = torch.mean(distance_st[distance_st >= d_hard])
	student_output_penalty = self.student(batch_imagenet)[:, : self.teacher_out_channels, :, :]
	loss_penalty = torch.mean(student_output_penalty**2)
	loss_st = loss_hard + loss_penalty
	# Autoencoder and Student AE Loss
	aug_img = self.choose_random_aug_image(batch)
	ae_output_aug = self.ae(aug_img)
	with torch.no_grad():
	teacher_output_aug = self.teacher(aug_img)
	if self.is_set(self.mean_std):
	teacher_output_aug = (teacher_output_aug - self.mean_std["mean"]) / self.mean_std["std"]
	student_output_ae_aug = self.student(aug_img)[:, self.teacher_out_channels :, :, :]
	distance_ae = torch.pow(teacher_output_aug - ae_output_aug, 2)
	distance_stae = torch.pow(ae_output_aug - student_output_ae_aug, 2)
	loss_ae = torch.mean(distance_ae)
	loss_stae = torch.mean(distance_stae)
	return (loss_st, loss_ae, loss_stae)
	else:
	with torch.no_grad():
	ae_output = self.ae(batch)
	map_st = torch.mean(distance_st, dim=1, keepdim=True)
	map_stae = torch.mean(
	(ae_output - student_output[:, self.teacher_out_channels :]) ** 2, dim=1, keepdim=True
	)
	if self.pad_maps:
	map_st = F.pad(map_st, (4, 4, 4, 4))
	map_stae = F.pad(map_stae, (4, 4, 4, 4))
	map_st = F.interpolate(map_st, size=(self.input_size[0], self.input_size[1]), mode="bilinear")
	map_stae = F.interpolate(map_stae, size=(self.input_size[0], self.input_size[1]), mode="bilinear")
	if self.is_set(self.quantiles):
	map_st = 0.1 * (map_st - self.quantiles["qa_st"]) / (self.quantiles["qb_st"] - self.quantiles["qa_st"])
	map_stae = (
	0.1 * (map_stae - self.quantiles["qa_ae"]) / (self.quantiles["qb_ae"] - self.quantiles["qa_ae"])
	)
	map_combined = 0.5 * map_st + 0.5 * map_stae
	return {"anomaly_map_combined": map_combined, "map_st": map_st, "map_ae": map_stae}