train my own data, the results is mussy

I used this project to train human segmentation data, change the num_class and .odgt file, but the results is mussy. Are there any other changes that need to be made

Start by reproducing the human-segmentation training with the changed num_class and .odgt file, then compare the training configuration and dataset format with the project's expected inputs. The issue does not name a file, test, dataset sample, or expected output, so completion cannot be verified until those details and a concrete diagnosis are provided.

Written by the indexing model from the issue text.