Custom data set

[IgnacioSan22]

I'm trying to train the model with a completely different dataset. I'm struggling to set the values in the config file, is there some documentation about setting them correctly or their meaning/influence?

[thangvubk]

Please check here for the config explanation.

[IgnacioSan22]

I only have two classes, with unbalance proportion of points, and independently of how I set the config parameters after 1 or 2 epochs of training the semantic part predicts zero points for the second class. Do you know why this could be?

[theshapguy]

I'm also facing a similar issue with unbalanced points, I have 4 classes and only 1 of them is being classified.

I even tried using the WeightedRandomSampler rather than the DistrubutedSampler but I'm still having the same issue

[SijanNeupane49]

I am also having same kind of issue. I also have unbalanced proportion of points in my datasets. I have 3 classes Leaf, stem and node. For instance, in one of my 3D model(16th model) the no .of points are as following:

1. Leaf = 3,77,131
2. Stem = 17,811
3. Node = 4667

I also tried changing many parameters in config file, updated weights according to class distribution ratio

def point_wise_loss(self, semantic_scores, pt_offsets, semantic_labels, instance_labels,
                        pt_offset_labels):
        losses = {}
        semantic_loss = F.cross_entropy(
            semantic_scores, semantic_labels, weights= torch.tensor([1,17,75]), ignore_index=self.ignore_label)
        losses['semantic_loss'] = semantic_loss

to model/softgroup.py (on line 144) to balance the class imbalance but still the semantic part predicts to zero for the second and third class. I have also posted similar kind of issues in #76 with screenshots of my results.

[SijanNeupane49]

@IgnacioSan22 @theshapguy did you guys manage to solve these issues?If yes, please help me as well. Thanks in advance.

[SijanNeupane49]

@thangvubk can you please maybe suggest us something regarding this issue? It would be immense help. Thank you in advance.

[thangvubk]

On dataset with imbalanced point per class i have two suggestions.

1. Using weighted semantic loss. Follow this.
2. Train longer. Ussually, class with more points will converge first, then class with less points.

[SijanNeupane49]

@thangvubk Thank you for your quick reply. How did you calculated those semantic weights for stpls3d? How can I do that for my custom dataset?

[zhongxiaj]

Please check here for the config explanation.

Great work! I have some questions; How can I organize my own my dataset ? is these any guideline for more detail?(such as how to organize my own .tsv(scannetv2-labels.combined.tsv), .aggregation.json,.segs.json) is there any codes for detail ?

[xiaotiancai899]

Please check here for the config explanation.

Great work! I have some questions; How can I organize my own my dataset ? is these any guideline for more detail?(such as how to organize my own .tsv(scannetv2-labels.combined.tsv), .aggregation.json,.segs.json) is there any codes for detail ?

Hi! Have you done this? And I also have the same question as you do.