Open icey-zhang opened 4 years ago
In the train_cityscapes.yml recipe, we use the ocrnet.HRNet_Mscale
architecture. The attention head is created here: https://github.com/NVIDIA/semantic-segmentation/blob/main/network/ocrnet.py#L175, and the classification head here: https://github.com/NVIDIA/semantic-segmentation/blob/main/network/ocrnet.py#L66
Thank you for your answer. I think I understand. Your work is excellent!
Hello, sorry, I have another question. How should I choose arch when testing. I directly modified train.py to test.py, and used the same arch as in train. I don't know whether it is correct?
Hello, sorry, I have another question. How should I choose arch when testing. I directly modified train.py to test.py, and used the same arch as in train. I don't know whether it is correct?
If you're asking how to run evaluation to generate a validation score for a single model, please refer to eval_cityscapes.yml. You specify the arch in the same way as for training, i.e. https://github.com/NVIDIA/semantic-segmentation/blob/main/scripts/eval_cityscapes.yml#L16. There is an example in the README.
Thank you very, very much for your prompt reply. I was a little surprised.I didn't have any success with runx, so I went straight to train.py.
Hi were you able to train the model succesfully?
how did you train it ? what was your labels?
Hello, I am using HRnet-OCR as my own data set for arch training, and I do not find the seg and attn modules mentioned in your paper, are they in the deeper.DeeperS8 module?In addition, if you want the network to be able to achieve multi-scale training, should you modify the parameter scales?