Use with Tensorflow

[ryanmaxwell96]

So I know this is a really strange question, but I've been trying to plot a model that has been built in the subclassing way with tensorboard 2+ (or really any way I can find) but nothing works because apparently subclassing makes things really difficult. I'm trying to do this on someone else's code and it has been very challenging to figure out how to change the model to be made in the sequential way or with "Input" which is why it seems that I will have to go with subclassing.

Then I found out that Pytorch also uses model subclassing that is why I started trying to investigate if tensorboardx could solve these issues due to model subclassing. However, I've been running into the error:

AttributeError: type object 'Model' has no attribute 'param_init_net'

which I think is probably due to the fact that tensorboardx is written specifically for pytorch and not tensorflow but I'm not sure.

So do you know if it can be used with Tensorflow? If not, any ideas as to how to move forward?

[lanpa]

Hi, If you are using TensorFlow with eager mode, you should able to get the value in numpy array. Once you have the numpy value, you can use tensorboardX to log the data. But if you want to visualize the TensorFlow model, tensorboardX cannot help with that. AFAIK, you can see the TF graph in tensorboard once you checkpoint your TF model.

[dynamicwebpaige]

Thank you for flagging this as an issue, @ryanmaxwell96! Have you filed a similar bug to github.com/tensorflow/tensorboard?