export model on `train.py`

[zengjingming]

Hi, Thank you for your great work!

I just want to ask how to export .h5 file when I run python train/train.py. It seems that the export part in the code is at the front, before training.

[zengjingming]

I'm sorry for the ignorance. When I run the train.py first time. I don't have any weights to load.So how can I export the model?

[felixrosberg]

Hello and thank you!

Yes, this is probably a odd solution. But if the export flag is set to True it will just export the model and exit. If you see a couple of rows above, you have G = load_model_internal(..) and opt.load. So if you set the --load flag to an int of a checkpoint you can load a model and train from there. If you also set --export to True you will load the model, export and exit. If you set --export to True but keep --load as None it will skip the export step. :)

[zengjingming]

Sorry to bother you again. I'm trying to tun this project on linux server with GPUs. Could you please tell me what version of tensorflow should I install ? I have tried tensorflow-cpu and tensorflow-directml-plugin you mentioned. But it seems it only use cpu.

[felixrosberg]

No worries! That way of intallation is suggested by good buddy and contributer netrunner-exe, so I know not much of it and would not install TensorFlow personally that way. I think any of the later 2.0 version should work, but try installing tensorflow-gpu instead. Also make sure you have correct version of CUDA. There is a second installation suggestion in the READ.ME, perhaps that would work for you. Otherwise I suggest you install tensorflow-gpu and look up what more you need on tensorflows webpage.

[netrunner-exe]

Sorry to bother you again. I'm trying to tun this project on linux server with GPUs. Could you please tell me what version of tensorflow should I install ? I have tried tensorflow-cpu and tensorflow-directml-plugin you mentioned. But it seems it only use cpu.

The directml installation method is more suitable for Windows or WSL. For Linux, I would also suggest trying the installation method described on the official Tensorflow website.

conda install -c conda-forge cudatoolkit=11.2.2 cudnn=8.1.0
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/
python3 -m pip install tensorflow

Verify install:
python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

You may need to fix the command a bit for more appropriate versions of cudatoolkit and cudnn for a particular GPU driver, or maybe not - in your case this can only be found out by experience. If everything was successful, the last command should show the number of GPUs in the system.

[zengjingming]

Thank you ! I have built my system successfully!

[zengjingming]

Can I ask what's the version of python,tensorflow this project uses?