Variable reuse issue with wide resnet model training script in autoaugment/train_cifar.py

[preethiseshadri518]

I am trying to train a wide resnet model for CIFAR10 using the models/research/autoaugment/train_cifar.py directory. I just copied over all the necessary files to train a CIFAR10 model and created a local repository. Since this requires Python 2 and Tensorflow 1, I am running the code with tensorflow 1.15 and python 2.7.6. I am using MacOS and CPU.

To train a wrn model, I run the following (I haven't modified the original code):

python train_cifar.py --model_name=wrn \
                      --checkpoint_dir=training \
                      --data_path=data \
                      --dataset='cifar10' \
                      --use_cpu=1

However, when I run this, I get the following error:
ValueError: Variable model/init/init_conv/weights already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?

I've included screenshots so you can see the entire traceback. The issue appears to be with custom_ops.py and reusing variables. I am not super familiar with tensorflow and gravitate towards use keras, so I'm not sure how to fix this issue and where any changes should be made. Thank you!

Also, this is my first time submitting an issue so apologies if I did not follow the guidelines completely and would welcome any feedback. I did not see any other issues filed about variable reuse for the cifar10/wide resnet training script.

[preethiseshadri518]

@ravikyram thank you for assigning this issue. Is there anything I can do to gain more visibility?