This is the deep learning model that we proposed in Black Hole Weather Forecasting with Deep Learning: A Pilot Study (Duarte, Nemmen & Navarro, MNRAS, in press). All the details are available and described in the paper. This repository includes the following files:
We trained the model using Tensorflow 1.8.0 with multi_gpu (P6000 and GP100).
model.py: the architecture we used is based on a U-Net with modifications in the number of layers and the input/output.
The main difference is that our input accepts the temporal dimension with two-spatial dimensions, while the classical U-Net accepts only the spatial dimensions.
The input is a tensor (N, 256, 192, 5).
params.py: hyperparameters you can change before training the model, which are the following:
- epochs: the number of epochs (default = 100)
- batch_size: the size of the mini-batch (default = 64)
- filters: how many filters are in the first layer (default = 32)
- results_path: the path where you will save the training
- alpha, beta, delta, gamma: loss function parameters (default = 0.1)
train.py: training settings without generator
train_gen.py: training setting with a generator to save memory
inference.py: how to create predictions using .h5 file
You can find the data files with the trained neural network weights in our group's data repository. They are in the HDF5 binary format, following Tensor Flow standards. You can load the weights using
model.load("/path/to/.h5/dl_fluids.h5")