Linear Variational Autoencoder - Not Learning? Help!

Hello, PyTorch community.

For the past few weeks I have been struggling real hard to make a piece of code work, something I supposedly thought to be a trivial task. Here it is:

I. The task.

  1. Download some Bitcoin data and pre-process it. (Transform closing prices into returns).
  2. Slice the BTC data set in windows of length 100.
  3. Define a VAE architecture, so I could use the decoder to simulate time series.

II. The problem.

First I started with LSTM and GRU layers. Somehow, I wasn´t sure if my model was learning something at all since it seemed the results were flattening, no matter how much I tweaked the hyperparameters. Then before going for other more sophisticated layers (Conv1d), I decided just to go for Linear layers (the code I present below), just to test if it is able to work.

Somehow, the output of the decoder part does indeed look somehow a time series. However, the scale is different from the input (similar to: The network output matches the target curve but with a different scale).

I tried: Increasing the number of epochs, adding more layers (linear transformations), changing learning rate, etc, etc. The results are the same. It looks like a time series, but in a very different scale.

Definitely, I am new to PyTorch, I was wondering if someone could please shed a light in this matter because cause literally I don´t know what else to try. I am uploading my Notebook. I would really appreciate if someone could take a look!


Output_BTC
Norm BTC

Colab Link: https://colab.research.google.com/drive/1ieJPbLrHnscJCP-BrPvtYd6F4PkHDyxu?usp=sharing

I thank you for your time!