Thank you for releasing your source code. I have a small question about the implem

Question about the implementation of pseudo label update about st3d HOT 4 CLOSED

cvmi-lab commented on August 12, 2024

Question about the implementation of pseudo label update

from st3d.

Comments (4)

jihanyang commented on August 12, 2024

Yes. I just use them to try some ideas in my previous attempts, and they are not used in current version.

from st3d.

s-ryosky commented on August 12, 2024

Thank you for your reply. I understood.

I have another question.
It seems that it uses all pseudo labels stored in memory bank for training, including those whose state is ignored.
Is this a bug?

ST3D/pcdet/datasets/dataset.py

Lines 145 to 152 in 73dda8c

 def fill_pseudo_labels(self, input_dict): 

 gt_boxes = self_training_utils.load_ps_label(input_dict['frame_id']) 

 gt_scores = gt_boxes[:, 8] 

 gt_classes = gt_boxes[:, 7] 

 gt_boxes = gt_boxes[:, :7] 

 # only suitable for only one classes, generating gt_names for prepare data 

 gt_names = np.array([self.class_names[0] for n in gt_boxes])

ST3D/pcdet/models/dense_heads/target_assigner/axis_aligned_target_assigner.py

Lines 38 to 51 in 73dda8c

 gt_classes = gt_boxes_with_classes[:, :, 7] 

 gt_boxes = gt_boxes_with_classes[:, :, :7] 

 for k in range(batch_size): 

 cur_gt = gt_boxes[k] 

 cnt = cur_gt.__len__() - 1 

 while cnt > 0 and cur_gt[cnt].sum() == 0: 

 cnt -= 1 

 cur_gt = cur_gt[:cnt + 1] 

 cur_gt_classes = gt_classes[k][:cnt + 1].int() 

 target_list = [] 

 for anchor_class_name, anchors in zip(self.anchor_class_names, all_anchors): 

 if cur_gt_classes.shape[0] > 1: 

 mask = torch.from_numpy(self.class_names[cur_gt_classes.cpu().abs() - 1] == anchor_class_name)

from st3d.

jihanyang commented on August 12, 2024

No. If I don't store ignored pseudo labels, anchors don't have corresponding GTs will activate negative loss. I set the labels of ignored pseudo labels with negative values, and ignore their losses when calculating loss.

from st3d.

s-ryosky commented on August 12, 2024

I understood it.
I didn't notice that the anchors corresponding to the labels with negative values were trained as negative samples.

Thank you.

from st3d.

Recommend Projects

Question about the implementation of pseudo label update about st3d HOT 4 CLOSED

Comments (4)

Related Issues (20)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	def fill_pseudo_labels(self, input_dict):
	gt_boxes = self_training_utils.load_ps_label(input_dict['frame_id'])
	gt_scores = gt_boxes[:, 8]
	gt_classes = gt_boxes[:, 7]
	gt_boxes = gt_boxes[:, :7]

	# only suitable for only one classes, generating gt_names for prepare data
	gt_names = np.array([self.class_names[0] for n in gt_boxes])

	gt_classes = gt_boxes_with_classes[:, :, 7]
	gt_boxes = gt_boxes_with_classes[:, :, :7]
	for k in range(batch_size):
	cur_gt = gt_boxes[k]
	cnt = cur_gt.__len__() - 1
	while cnt > 0 and cur_gt[cnt].sum() == 0:
	cnt -= 1
	cur_gt = cur_gt[:cnt + 1]
	cur_gt_classes = gt_classes[k][:cnt + 1].int()

	target_list = []
	for anchor_class_name, anchors in zip(self.anchor_class_names, all_anchors):
	if cur_gt_classes.shape[0] > 1:
	mask = torch.from_numpy(self.class_names[cur_gt_classes.cpu().abs() - 1] == anchor_class_name)