Setting panoptic_label_divisor=None gives error

[mi2celis]

I am trying to run training with the cityscapes dataset with only semantic segmentation annotations.
I have set panoptic_label_divisor=None in dataset.py. I'm using ${DEEPLAB2}/configs/cityscapes/panoptic_deeplab/ resnet50_os32_semseg.textproto, where the instance branch is not initiated.

When running:
python trainer/train.py
--config_file= deeplab2/configs/cityscapes/panoptic_deeplab/resnet50_os32_semseg.textproto
--mode=train
--model_dir=deeplab2/model
--num_gpus=1

I get the following error

Traceback (most recent call last):
File "trainer/train.py", line 81, in
app.run(main)
File "/home/mimi/Envs/deeplab2/lib/python3.6/site-packages/absl/app.py", line 312, in run
_run_main(main, args)
File "/home/mimi/Envs/deeplab2/lib/python3.6/site-packages/absl/app.py", line 258, in _run_main
sys.exit(main(argv))
File "trainer/train.py", line 77, in main
FLAGS.num_gpus)
File "/home/mimi/code/deeplab2_proj/deeplab2/trainer/train_lib.py", line 137, in run_experiment
trainer = trainer_lib.Trainer(config, deeplab_model, losses, global_step)
File "/home/mimi/code/deeplab2_proj/deeplab2/trainer/trainer.py", line 137, in init
only_semantic_annotations=not support_panoptic)
File "/home/mimi/code/deeplab2_proj/deeplab2/trainer/runner_utils.py", line 122, in create_dataset
focus_small_instances=focus_small_instances)
File "/home/mimi/code/deeplab2_proj/deeplab2/data/sample_generator.py", line 127, in init
*self._dataset_info['panoptic_label_divisor'],
TypeError: unsupported operand type(s) for : 'int' and 'NoneType'

Any suggestions of what may be the problem.

[aquariusjay]

Hi mi2celis,

Thanks for the question.

To train the model on Cityscaspes semantic segmentation, you do not need to change anything except use the resnet50_os32_semseg.textproto.

The dataset name in the textproto should still be "cityscapes_panoptic" (as shown here) and the dataset is converted in the same way as the tutorial.

The code currently only reads panoptic dataset format, as shown in this line.

Hope that helps.

Cheers,

[mi2celis]

Got it! Thank you for the prompt reply.

So, if my dataset only contains semantic segmentation annotations, can I still use DeepLab2 to train as explained in the section to convert your own dataset ?

[aquariusjay]

Hi mi2celis,

Yes, that is correct. Basically, you consider your dataset as the one that contains only stuff classes.
I think we have a typo in the doc about setting panoptic_label_divisor=None where you should actually still set a panoptic_label_divisor (e.g., 256 or even smaller) since the code only reads panoptic dataset format at the moment (e.g., here).
We will update the doc soon. Thanks for bringing up the issue.

Additionally, you may first want to make sure if you could successfully train Cityscapes semantic segmentation.

Cheers,

[aquariusjay]

Hi mi2celis,

Let me add a bit more details here which may be helpful when you will be converting your own dataset.

Since the code only reads panoptic data at the moment, you need to set panoptic_label_divisor = k, where k is any integer, instance_id = 0, and class_has_instances_list = [] (i.e., we treat the dataset as the one that contains only stuff classes), when you are (1) converting the dataset to TFRecord, and (2) adding dataset information in dataset.py.

We will update the doc to include this information soon.

Please feel free to report some more issues afterwards.

Cheers,

[aquariusjay]

Closing the issue as there is no active discussion for a while.
Please feel free to reopen or create a new one if you still have any questions.