Bug fix: 'ArgsParser' object has no attribute 'reg_loss'

[aniruddhapal211316]

Issue :

while running the pretrained model using clean.py or run_segan+_clean.sh , it shows an error AttributeError: 'ArgsParser' object has no attribute 'reg_loss' .
This same problem was also reported earlier in this repo, here is the link for it: #36

Fix:

added "reg_loss": "l1_loss", in segan_pytorch/ckpt_segan+/train.opts

Due to this commit 9ade47b , a new command line option --reg_loss was added for training the model which initialized a variable self.reg_loss present in segan/models/model.py in class SEGAN to l1_loss or mse_loss depending on the command line argument passed by the user.

But the pretrained model was trained before that commit was made therefore it was trained without using --reg_loss option in command line and self.reg_loss was initialized to l1_loss. All the command line arguments used for training the pretrained model are stored in segan_pytorch/ckpt_segan+/train.opts and are loaded again when using the pretrained model using clean.py. So by adding "reg_loss": "l1_loss", in segan_pytorch/ckpt_segan+/train.opts solves this issue.

[woshiZS]

@aniruddhapal211316
sorry to bother you, the pre-trained model link in README seems to be empty? Where to find the pre-trained model ?
any reply will be appreciated !!!