To allow support for large datasets (10gb+).

See http://nbviewer.jupyter.org/github/zer0n/MLx/blob/master/demo/pymlx.ipynb#Parameter-Sweep for example.

Code: https://github.com/zer0n/MLx/blob/master/pymlx/parameter_sweep.py

Resource: https://arxiv.org/pdf/1703.04691.pdf

Make the number of levels parameter match the severity levels in business rather than a sensitive parameter with no physical meaning in business.

Once we have that, then users wouldn't even need to input the number of anomalies parameter and still obtain a good anomaly detection model.

We want input and output dimensions to be user specified when initialization the model and forecaster class. This causes problems, because we have currently delayed initializations. For example, no real custom models or loss functions are currently accepted by the forecaster.

This makes it easier for the user to understand what the user can provide.

https://arxiv.org/pdf/1705.07795.pdf

github_total_push_events_2011-2018.csv file that is to be read in is wrongly named and should be github_dau_2011-2018.csv to match with the data file in raw

Let's build this after the AI class

Build multi-step loss function into model. Capture aleatoric uncertainty in loss function.

This can be built using straight up pytorch, should also include an example tutorial.

Maybe combine with initialization issue.

We should add a stacked GRU with residual connections to have this in our model repertoire.

1. Horizon = 50, Lag = 500 wrt original timestamp
2. Downsample rate: 1, 2, 3, 10, 25, 50
3. Sub-lag: Lag/Downsample rate

Extensions:

1. Dealing with prime numbers or number with fewer divisors
2. Elastic network capacity corresponding to different sub-lag(input-size)

Users should be able to input either a WaveNet model, a string 'wavenet', 'gru', 'lstm', etc. No matter how cool WaveNet is, many users would expect to be able to use LSTM.

This allows a more diverse set of models to be built using a small and flexible codebase

GefCom has a bivariate time series dataset. It's nicer to show bivariate case.

Ideally we have a very clear and clean set of script to run models on benchmark datasets.

Could be distributional losses, or CDF-based losses, or combinations.

This function takes the following data specs <input dimension, output dimension, history length, forecast horizon> as input and outputs the network topology.

Do this for CNN first. This would eliminate the need to parameter sweeping for CNNs, which include lots of nested parameters.

Hi,

In the tutorial on Github traffic data, how the input_channel parameter's value was set to 21? An explanation of that would be helpful.

Resource for Keras code:
https://github.com/yaringal/ConcreteDropout

and integrate it into this library if it works well.

Also compare this with the Tensor-Train RNNs paper.

distributeddataparallel claims better performance than dataparallel

Original Paper

Follow-up paper for RNN

For quick testing and productionization

Hyperoptimizer issues error tracing back to CNTK. Same code runs fine in tensorflow. Error can be observed in tutorials/hyperoptimization.ipynb

Take the notebook template and refactor into Python functions and small notebook.

We will use this as the very first baseline.

Pre-compute

instead of just

.

Winning Kaggle entry for web traffic forecasting:
https://github.com/shellshock1911/Sky-Cast-Capstone/blob/master/final_report.pdf
https://github.com/Arturus/kaggle-web-traffic
https://github.com/Arturus/kaggle-web-traffic/blob/master/how_it_works.md

• Study object detection metric
• Design TS anomaly detection metric
• Explain to Mobius team about the metric and the label protocol

Add static and time-dependent covariates to fit and predict methods in forecaster module. I suspect it is best to have keras model that exepcts covariates all conform to the same interface in terms of how they expect the input data to propagate through the network.

1. Consolidate dictionary list of topology parameters
2. Accept Keras models

Create readthedocs.org documentation that documents how the frameworks works,

We need to have a custom loss function similar to the ones from this paper:
https://arxiv.org/abs/1703.04977

The dimensionality of the data need to be specified as types and comments in the forecaster.