an image only 'road', 'sidewalk' label

[Shaiken]

my enviorment GPU V100 docker : nvcr.io/nvidia/pytorch:20.12-py3 NVIDIA-SMI 470.63.01 Driver Version: 470.63.01 CUDA Version: 11.4

sorry ask this question, i try edit cityscapes.py and nullloader.py, adjust num_classes = 19 to num_classes = 2 and try edit cityscapes_labels.py label id,

• 'ego vehicle' 1 --> 8
• 'rectification border' --> 7
• 'road' 7 --> 2
• 'sidewalk' --> 1 because my map color image only 3 type value, unlabeled: [0], 'road': [2], 'sidewalk': [1]

loss rate: [epoch 0], [iter 11 / 986], [train main loss 1.809400], [lr 0.005000] [batchtime 1.47] ... [epoch 5], [iter 528 / 986], [train main loss 0.749431], [lr 0.002222] [batchtime 1.41] ...

i look my best image prediction just one color ...

how solve it

[Shaiken]

i found problem, if i wnat ignore a lot area, need inset a class not ignore val [255]