How to train own dataset?
pentaris opened this issue · comments
Hello, I'm a student studying machine running.
I found ESPNet v2 after looking for a network of real time semantic segmentation processing on TX2.
I don't have the cityscapes dataset, so I want to do training with own dataset.
I'm a beginner about machine running.
Could you provide a tutorial that trains to own dataset?
Thank you in advance.
Training has two parts:
i) Data preparation: Here, you check if the data is in write format or not. Also, you compute class imbalance weights if your dataset is imbalanced. Please see loadData.py file for more details. It is pretty simple and accepts two text files: one for training and one for validation. You can create these files easily as these contains mapping between RGB image and Segmentation mask (ground truth) in comma separated form. Note that one line per image pair.
ii) Training: One you are done with this part i), you can start training your model on your dataset. You need to adjust the number of classes in the main.py file. Please see main.py file for more details. It should be straight forward.
Let me know if you are still encountering any issues.
I tried to train in the way you showed me.
However, I encountered such an error during training.
There are 12 dataset classes, and the size is 480 x 320.
I modified the code a little, but is there something missing?
I encountered the similar issue before. After I prepare the data image/annotation with 1024x512 resolution, the problem disappeared. I can't remember very clearly. You can have a try.