Hi, I refactored the whole code

[ghost]

Every application is the same as your purpose. This code is runnable on tensorflow 1.2

I removed the unused variables, fixed some minor errors, and simplified the code, and added many comments to understand easily.

The only thing I omitted is the usage of tensorboard. Actually I don't like it, but just kept the pictures on README.md

This kind of code is just my style. I hope you like it.

[phreeza]

Hey this looks great, thanks. I am a bit busy at the moment but I will have a look over the PR as soon as I get a chance. Just out of curiosity, what are you using the vrnn for?

[ghost]

During searching of MANN(memory augmented neural network), I read some papers about GTMM(generative temporal model with memory). This model applied VAE+RNN+Memory at the same time. To understand this topic fully, I studied about VRNN(VAE+RNN). Ur code was really helpful for me to understand. I'm looking forward to finding some models to catch detect images with defect, when some kinds of images come in series.

[diegoroman17]

Hello, Thank you for share your code. I am understanding VRNN using it. Please help me with 2 specific questions:

1. What is the chunk_samples parameters and why the 3 dimension of the input and output has only 0's.
2. The code is testing for 1-D time series, but could it work for 10-D time-series? Thank you in advance

[diegoroman17]

Another question is about the use of inputs and targets. On the original paper they reconstruct xt but in your code the target is xt+1