Parameter/Checkpoint Compatability and Discriminator Network

[AdityaAgg]

Hi @maxorange, I am a Computer Science student at USC taking a class on deep learning. For my final project, I am trying to find a technique to color a 3D model given an image or texture. I found your github repos, esp. this one, super useful resource so far towards my project and going to credit and reference you all over my final deliverables. Thank you for the super thorough implementation.

However, I am having a few issues with this repo currently. Here they are:

1. The parameters in the checkpoint file don't load for newer versions of TensorFlow OutOfRangeError (see above for traceback): Read less bytes than requested [[Node: save_3/RestoreV2_53 = RestoreV2[dtypes=[DT_FLOAT], _device="/job:localhost/replica:0/task:0/cpu:0"](_recv_save_3/Const_0, save_3/RestoreV2_53/tensor_names, save_3/RestoreV2_53/shape_and_slices)]] [[Node: save_3/RestoreV2_52/_43 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/gpu:0", send_device="/job:localhost/replica:0/task:0/cpu:0", send_device_incarnation=1, tensor_name="edge_243_save_3/RestoreV2_52", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"]()]]

I tried loading the meta file and clearing devices too before restoring the data too and that did not seem to help.

2. I decided to go about retraining the stacked gan model (because I'll have to modify it and figure out how to be trained towards solving my problem anyways). I was looking for the discriminator network and I couldn't find it! I was wondering if you had any sample code that shows how you trained the network. Let me know if you could send me an email at [email protected] if you wanted to keep communication for this in a different place (since this roadblock not so much of an issue with the code as the first one). I am going to see if I can somehow use the discriminator in your DCGAN implementation but let me know if you had better suggestions over email or through this issue.

Best, Aditya

[maxorange]

Hi Aditya,

1. A file named sgan_model.ckpt.data-00000-of-00001 may be missing. Could you check if this file exists in your params folder?

2. I added the discriminator to my code. You can see it in models/sgan.py. I also added sample code for training so that you can modify and retrain the networks. If there are any errors, please let me know.

[AdityaAgg]

Hi @maxorange thank you for the quick and invaluable response. I think I do have the file but I get a warning with the file when pulling or cloning which could be the issue. Error attached as an image:

Two additional questions unrelated to this issue but I was curious about (feel free to answer for these through [email protected]):

1. Do you have recommended parameters for the adam's optimizers? I was planning to use what was recommended in the Stack GAN paper. But let me know if you had found better ideas through experimentation.
2. I was wondering why you chose to feed the edge encoder results for style discriminator instead of the 3 outputs from the color encoder? Sorry in advance if this is an ill-informed question!

Edit Figured out the issue with the checkpoint file by looking at a closed issue. #2 If you are able to answer the other questions that'd also be great!

[maxorange]

Okay, I will reply to your email address.

[ahmedshingaly]

just for reference, 3d model coloring is well explained in this article, take a look. Pix2Vox is also referenced in it