Error with checkpoint on evaluate.py and visualize.py about mobile-segmentation

[sercant]

Hello @thepuremind,

The model that you are trying is the DPC encoder head variant. I assume that this is due to a missing parameter that needs to be passed to the scripts.

To fix this error pass this parameter to the script:

--dense_prediction_cell_json=./core/dense_prediction_cell_branch5_top1_cityscapes.json

Let me know if it fixes the issue. I am going to be updating the documentation in the following months. Until the documentation is updated I will keep the issue open.

[thepuremind]

@sercant , thanx!
It fixed this problem. Also I have a questions. How I can start to evaluate on test set of Cityscapes data?
I think that need to add test set in tfrecord. How to do it?
What I need doing to run to training on other data?

[sercant]

@thepuremind, test set doesn't have the ground truth files provided. The evaluation should be done at www.cityscapes-dataset.com. To send to the evaluation one can use the visualize.py to generate the output, zip and send to the server.

To generate the test tfrecord files:

1. Modify the line

   mobile-segmentation/dataset/build_cityscapes_data.py

   Line 169 in 1ca269e

   for dataset_split in ['train', 'val']:

   and add 'test' to the list.
2. Run the build script again.
3. To visualize the test set, pass parameter --vis_split=test to the visualize.py.

To run on your own dataset, please examine the scripts under the dataset folder and generate your own tfrecord files. Then add the information of the dataset to segmentation_dataset.py file. As there are already several examples within the file, it should be straight forward.