Detailed technical report in https://github.com/Darthholi/Sap500/blob/master/VolumePredictionWithWavenet.pdf
Includes pretrained models.
First look at the data was done using colab notebook: https://colab.research.google.com/drive/1yuOsEa0zRQkE4p-qIeYYWsm7Ov8yBmH9
The experiments did run on Intel® Core™ i7-6700K CPU @ 4.00GHz × 8 with 15,6 GiB memory and
GeForce GTX 1080 Ti/PCIe/SSE2 graphic card (and environment from
https://lambdalabs.com/lambda-stack-deep-learning-software.
Since the results were dependent on the initial random training starting point,
there is present a saved modelwavenet-startpoint-suggest.h5, which can be used to continue the training
(the model is different only in some layers, so most of the weights can be reused).
Running the code in google colab is shown in this section of the notebook: https://colab.research.google.com/drive/1yuOsEa0zRQkE4p-qIeYYWsm7Ov8yBmH9#scrollTo=ea2QjosCbNc6
Setting up the environment:
curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
sudo apt-get install git-lfs
pip install --upgrade -r requirements.txt
Download the data not present here: ./download.sh
Train from scratch:
python cmd-train --model_name="wavenet/wavenet.h5"
Use a pretrained optimum:
cmd-train --model_name="wavenet/wavenet.h5" --start_weights_from="wavenet-startpoint-suggest.h5"
Evaluate a model and plot graphs:
cmd-eval --model_name="wavenet1452/wavenet.h5" --cache_name="wavenet1452/cached/"