I'm working on a deep learning model which takes image inputs, encodes them into a latent representation and reconstructs them.
I'm using visdom to visualise the inputs, outputs, latent variables, and monitor the loss function. I create a vis = visdom.Visdom() object and pass it into the network. As the network uses the various latent variables, the Visdom object visualises them with vis.image(...).
The problem is this design means the images get drawn in an unsynchronised way which makes it hard to track which images in the visualisation correspond to each other. I would like to make it so visdom only updates every n iterations but its not clear to me how to do this.
Of course, I could make the network return all of its latent variables and call vis.image only in the training script, but is there a way to circumvent this?
