about gpu about slater

[yilmazkorkmaz1]

Hi, sorry for late response. Did not receive any notification.

Of course. A single 4090 would be enough for both training and inference.

[greenyjwk]

hi, Thanks for your contribution! I see you used 5x2080ti gpu.2x4090 gpu with 24 GB each can run the project?

How would you be able to run TensorFlow version1 on RTX 4090 with the prerequisites mentioned in the README file? Because I thought the RTX 4090 requires CUDA 11.8 at least.

[yilmazkorkmaz1]

Hi Ji,

I have not used 4090 at that time but it might be doable using either these solutions (https://forums.developer.nvidia.com/t/can-nvidia-tensorflow-1-x-be-used-with-rtx-4090/241211) or migrating to TF2 (https://www.tensorflow.org/guide/migrate).

You can start with using tf2:
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
whenever you import tensorflow to deactivate v2 specialities (other than this you probably need to change a few lines in other dnnlib files).
You might check these repos too.
https://github.com/icon-lab/FedGIMP
https://github.com/christophmeyer/stylegan2-tf2

[greenyjwk]

Hello Yilmaz,

Thank you for the response.
Yes, I used the below approach as you suggested, but I have been facing with the following error.

[tf2 -> tf1]
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

[Error]
File "/home/jie/SLATER-main_tfv2/dnnlib/tflib/network.py", line 149, in _init_graph
out_expr = self._build_func(*self.input_templates, **build_kwargs)
File "", line 1211, in G_synthesis
File "", line 250, in get_positional_embeddings
AttributeError: module 'tensorflow' has no attribute 'variable_scope'

By the way, thank you for suggesting other repos.

[yilmazkorkmaz1]

Yes, it is expected to see these kind of migration issues. These might help.
https://www.tensorflow.org/api_docs/python/tf/compat/v1/variable_scope
https://www.tensorflow.org/guide/migrate/model_mapping

Or easier solution might be to check the repos I shared in the previous response.

[greenyjwk]

Thank you so much for the suggestion.

As you suggested, I have tried three different ways, but the error still occurs.

1. For all the python scripts that import tensorflow, I changed to tensorflow.compat.v1. And, the following error still occured,
   AttributeError: module 'tensorflow' has no attribute 'variable_scope' still occurs from the network.py.

2. As you specified there is also network.py file in FedGIMP, so I compared and changed some lines from network.py from SLATER. The error still occurred.

3. I replaced the (networks.py from SLATER) With FedGIMP/dnnlin/tflib/network.py, where the error comes out. The errror still occurred.

For your reference, the error occured "SLATER-main_tfv2/dnnlib/tflib/network.py", line 143
out_expr = self._build_func(*self.input_templates, **build_kwargs)