buntyke/pytorch-hed

pytorch-hed's Issues

Does anyone get good results for training?

I follow the train.py to train the model with the default hyper parameters in the code. However, the results are not good. Does anyone successfully train the model yet? Appreciate if providing the hyper parameters. Thanks

Details to get it working

I had to jump through the following (minor) hoops to get this code to work. Hope this helps the next guy:

1. If you don't want to deal with the notebook: python train.py
2. In "train.py", change dataRoot to point to the location of your own "HED-BSDS" folder or "data/HED-BSDS" if you are following the readme.
3. Original "HED-BSDS" does not have a val-pair.lst file. I just took 10% of the train-pair.lst file entries and created a new "val-pair.lst" file, with ~2800 entries. Don't forget to shuffle the train-pair.lst entries first. This can be done online, search for "Shuffle Text Lines online".
4. Validation is too often and ineffecient GPU-wise. Increase "self.valInterval = 500" in trainer.py to make it
   infrequent.

buntyke, if you make these changes to your original Readme, please feel free to remove this "Issue".

Cheers,

question about the completeness of this code.

Is this code complete? I wonder if you have result images of your code.
thanks :)

Train from scratch

Hi,

Is there some way to train it from scratch? Probably use my own architecture?