VAE in Pytorch: The VAE is realized in Pytorch 0.4.The Varaiable container is replaced by Tensor in this version.
Variational autoencoder (MLP encoder/decoder) and related models A good introduction of VAE can be found in Jann's blog: https://jaan.io/what-is-variational-autoencoder-vae-tutorial/.
The VAE in pytorch is inspired by the Dmitry Efimov's VAE Tutorials in Theano:http://efimov-ml.com/vae.html.
Compared with Theano framework, the realiziation in Pytorch has several advantages: The declaration of class is seperated from the other memeber functions like in C++. The forward function in Pytorch is straight forward and the back-propagation is left to the backward() function with optimizer. The update function in Theano is not necessary. The archiecture of Pytorch model is much clearer than model in Theano.
LSTM in Pytorch: The LSTMcell is realized with nn.Linear layer. The structure is simple and the result is tested with a online time series code. A good understanding of LSTM can be found in colah's blog: http://colah.github.io/posts/2015-08-Understanding-LSTMs/ For this moment, some people do not like RNN + LSTM. However, the electrical circul like stucture and the backpropagation is very well combined. This kind of mechanism (backpropagation) is indeed a good option for 19th century machine-like automation.
The test code is from https://github.com/L1aoXingyu/code-of-learn-deep-learning-with-pytorch with necessary modification.