axelbrando/mixture-density-networks-for-distribution-and-uncertainty-estimation

A generic Mixture Density Networks (MDN) implementation for distribution and uncertainty estimation by using Keras (TensorFlow)

mixture-density-networks-for-distribution-and-uncertainty-estimation's Issues

Version of Keras and Tensorflow

Please specify which version of Keras and Tensorflow this code is tested on.

Error in link

Hi.
In readme you have error in link on 'Introduction to MDN models and generic implementation of MDN'

#### [Introduction to MDN models and generic implementation of MDN](https://github.com/axelbrando/Mixture-Density-Networks-for-distribution-and-uncertainty-estimation/blob/master/MDN-Introduction.ipynb.ipynb)

-->

#### [Introduction to MDN models and generic implementation of MDN](https://github.com/axelbrando/Mixture-Density-Networks-for-distribution-and-uncertainty-estimation/blob/master/MDN-Introduction.ipynb)

Thesis

Hi Axel,

Is the thesis available yet ?

Regards,

Christopher

problems of implementing the cost function with covariance diagonal matrix

It's really an excellent work which helps me a lot. Recently, i'd like to implement a cost function with covariance diagonal matrix, which has been told not necessary in the original paper. However, i just want to try that. During the process, i've met a problem that the |∑| always leads to inf when the dimension of output features increase, which leads the model can't converge. Any ideas about that? @axelbrando

AttributeError: 'module' object has no attribute 'python'

when i try MDN-Introduction.ipynb
import tensorflow as tf
tf.python.control_flow_ops = tf

an error:
AttributeError Traceback (most recent call last)
in ()
1 #Import of the TensorFlow and definition of the control_flow_ops variable
2 import tensorflow as tf
----> 3 tf.python.control_flow_ops = tf
4
5 #GPU Memory allocation on demand (Remove comments if necessary)

AttributeError: 'module' object has no attribute 'python'

the tensorflow version: 1.1.0-rc1;
keras version: 2.3.0

How to sample for both Normal distribution and Laplace distribition

Hi, in your validation code, you just use the network output mu to draw the graph, I'm not quite understand this part. I think the "mu" & "sigma" is used to show the range(mu and deviation) of the distribution, is that right?

And, as we need to do sampling from the network output parameters, how could I do sampling both for Normal distribution and Laplace distribution?

Thanks very much.

Error in exponent ?

"- .5 * float(c) * K.log(sigma)" instead of "- float(c) * K.log(sigma)" ?

dataset

I want to know that what type of dataset you are using to implement this model.?
Please reply soon.
@axelbrando
@algorismes

What is the role of Log-sum-exp?

In your code for defining Loss function, you firstly take the log of distributions then add them up, whereas the loss mentioned in Bishop is log of sum of distribution.I am confused, Maybe there is something in trick you used for log-sum-exp.
Can you please formulate this?

LSTM MDN data

Hi Axel, really nice work!

I am wondering if the data for the LSTM MDN and also the weight.h5 could be provided? Thanks a lot!

Best,
Danqing

Feature request related to canceling effects of uneven x distribution

Hi there I have read your work, very good stuff.

I am wondering if you could give a hint on how to adjust he MDN layer to avoid weiging along the x-axis. Basically my dataset has many points on some regions for x-axis and I am only interested in fitting the data with weights along the y-axis.

Nealry all only examples show an "evenly" generated dataset for x and y, here the model will perform well.

Would you have any idea on how to skip weighing distribution along the x-axis and match the shape of distrubution for y? I was thinking of binning the data in zones of x and running sperate models only to combine them later.

Thanks !