Skip to content

Latest commit

 

History

History
257 lines (170 loc) · 11.7 KB

readme.md

File metadata and controls

257 lines (170 loc) · 11.7 KB
Raw

Neural Processes for Sequential Data

This repository houses an implementation of the "Recurrent Attentive Neural Process for Sequential Data" (ANP-RNN), tested on a toy regression problem and real smart meter data.

ANP-RNN Diagram

The repository provides options for running as an ANP-RNN, ANP, or NP.

Numerous tweaks have been made for flexibility and stability, including a replication of the DeepMind ANP results in PyTorch. This replication appears to be a better qualitative match than other PyTorch versions of ANP (as of 2019-11-01). See also section for other code repositories.

The code is not extensively documented due to limited usage. If you are using it and find it confusing, please raise a GitHub issue and we can enhance the documentation together.

Experiment: Comparing models on real world data

Here I compare the models on smartmeter power demand data.

The black dots are input data, the dotted line is the true data. The blue line is the prediction, and the blue shadow is the uncertainty to one standard deviation.

I chose a difficult example below, it's a window in the test set that deviates from the previous pattern. Given 3 days inputs, it must predict the next day, and the next day has higher power usage than previously. The trained model manages to predict it based on the inputs.

Results

Results on Smartmeter prediction (lower is better)

Model val_np_loss val_mse_loss
ANP-RNN(impr)(MCDropout) -1.48
ANP-RNN_imp -1.38 .00423
ANP-RNN -1.27 0.0047
ANP -1.3 0.0072
NP -1.3 0.0040
LSTM -0.78 0.0074

Example LSTM baseline

Here is an LSTM with a similar setup: it has access to the y value in the context (first half). It's output is inferier and it's uncertainty estimation if poor. It starts of high since it hasn't had much data yet, but it should increase, or at least stay high in the second half as it moves away from it's data.

Example NP

Here we see underfitting, since the curve doesn't match the data

Example ANP outputs (sequential)

Here we see overfitting, but the uncertainty seems to small, and the fit could be improved

Example ANP-RNN outputs

This has a better calibrated uncertainty and a better fit

Example of ANP-RNN with MCDropout

Experiment: Comparing models on toy 1d regression

I put some work into replicating the behaviour shown in the original deepmind tensorflow notebook. At the same time I compared multiple models.

Results

Results on toy 1d regression (lower is better)

model val_loss
ANP-RNN(impr) -1.3217
ANP-RNN -0.62
ANP -0.4228
ANP(impr) -0.3182
NP -1.2687

Example outputs

Compare deepmind:

And this repo with an ANP (anp_1d_regression.ipynb)

And a ANP-RNN

It's just a qualitative comparison but we see the same kind of overfitting with uncertainty being tight where lots of data points exist, and wide where they do not. However this repo seems to miss points occasionally.

Experiment: Using ANP-RNN + Monte Carlo Dropout

One more experiment is included:

The model tries to estimate the how unsure it is, but what about when it is out of sample? What about what it doesn't know that it doesn't know?

Name val_loss (n=100) [lower is better]
MCDropout -1.31
Normal -1.04

We can estimate additional uncertainty by using Monte Carlo Dropout to see how uncertain the model acts in the presence of dropout. This doesn't capture all uncertainty, but I found that is does improve (decrease) the validation loss. The loss is calculated by the negative overlap of the output distribution and the target value so this improvement in the loss shows that MCDropout improved the estimation of the uncertainty.

Why didn't the model just learn to be more uncertain? Well I choose a challenging train, val/test split where the val data was in the future and showed quite differen't behaviour. That means that the validation data had behaviour the model has never seen before.

With MCDropout:

Without

For more details see the notebook ./smartmeters-ANP-RNN-mcdropout.ipynb

Usage

  1. Clone this repository.
git clone
git-lfs pull
  1. Refer to requirements.txt for software requirements and versions.
  2. Run the notebook smartmeters.ipynb.
  3. For a toy 1D regression problem, refer to anp-rnn_1d_regression.ipynb.

Smartmeter Data

Code

The code is based on the code listed in the next section, with modifications for stability and to ensure it can handle future predictions. Notable changes include:

Changes for a sequential/predictive use case:

  • target points are always in the future, context is in the past
  • context and targets are still sampled randomly during training

Changes for stability:

  • in eval mode, take mean of latent space, and mean of output isntead of sampling
  • use log_variance where possible (there is a flag to try without this, and it seems to help)
    • and add a minimum bound to std (in log domain) to avoid mode collapse (one path using log_var one not)
  • use log_prob loss (not mseloss or BCELoss)
  • use pytorch attention (which has dropout and is faster) instead of custom attention
  • use_deterministic option, although it seems to do better with this off
  • use batchnorm and dropout on channel dimensions
  • check and skip nonfinite values because for extreme inputs we can still get nan's. Also gradient clipping
  • use pytorch lightning for early stopping, hyperparam opt, and reduce learning rate on plateau

ANP-RNN diagram

Tips

  • Make you normalise all data, ideally the output two, this seems to be very important
  • Batchnorm, lvar, dropout: these seem ok but it's unclear to me how to make these help reliably. Attention dropout or lstm dropout can be especially unreliable.
  • sometimes you need quite a large hidden space to model a process. Making the network deep seems to stop it learning effectivly. It would be helpfull to try differen't activations, initializations and make sure the gradient flows effectivly to deeper networks,
  • The deterministic path had unclear value, I found it best to leave it out
  • The absolute size and comparitive size of the context and target is important for performance.
    • If the context is too long and complex the model cannot summarize it
    • If the target is too long and complex the model cannot fit it well
    • If the context is in the target, the model may collapse to just fitting this. To fix
      • make it small
      • or make the loss on this part downweighted, this seems like the best approach since x_context->y_context may still be a usefull secondary task
      • or do not include context in target
    • however including the target in the context may sometimes be helpfull
  • This repo compares models, but the biggest difference in this situation would be from additional data sources, but that is outside the scope of these experiments

See also:

A list of projects I used as reference or modified to make this one:

I'm very grateful for all these authors for sharing their work. It was a pleasure to dive deep into these models compare the different implementations.

Neural process papers:

Blogposts:

Citing

If you like our work and end up using this code for your reseach give us a shout-out by citing or acknowledging