How do I use HardNegativeSampler ?

[DSRajesh]

Hello
How do I use HardNegativeSampler while training a classifier using DenseNet121 ?

Thanks
Rajesh

[KumoLiu]

Hi @DSRajesh, here is the training workflow.

MONAI/monai/apps/detection/utils/hard_negative_sampler.py

Lines 104 to 108 in 0a3cdd9

	The training workflow is described as the follows:
	1) forward network and get prediction scores (classification prob/logits) for all the samples;
	2) use hard negative sampler to choose negative samples with high prediction scores and some positive samples;
	3) compute classification loss for the selected samples;
	4) do back propagation.

So you can put HardNegativeSampler after out = model(input).

You can also refer to:

MONAI/monai/apps/detection/networks/retinanet_detector.py

Lines 938 to 956 in 0a3cdd9

	if isinstance(self.fg_bg_sampler, HardNegativeSampler):
	max_cls_logits_per_image = torch.max(cls_logits_per_image.to(torch.float32), dim=1)[0]
	sampled_pos_inds_list, sampled_neg_inds_list = self.fg_bg_sampler(
	[matched_idxs_per_image + 1], max_cls_logits_per_image
	)
	elif isinstance(self.fg_bg_sampler, BalancedPositiveNegativeSampler):
	sampled_pos_inds_list, sampled_neg_inds_list = self.fg_bg_sampler([matched_idxs_per_image + 1])
	else:
	raise NotImplementedError(
	"Currently support torchvision BalancedPositiveNegativeSampler and monai HardNegativeSampler matcher. "
	"Other types of sampler not supported. "
	"Please override self.get_cls_train_sample_per_image(*) for your own sampler."
	)
	sampled_pos_inds = torch.where(torch.cat(sampled_pos_inds_list, dim=0))[0]
	sampled_neg_inds = torch.where(torch.cat(sampled_neg_inds_list, dim=0))[0]
	valid_idxs_per_image = torch.cat([sampled_pos_inds, sampled_neg_inds], dim=0)
	return cls_logits_per_image[valid_idxs_per_image, :], gt_classes_target[valid_idxs_per_image, :]

Thanks!