OCF's DataPipes for training and inference in Pytorch.
ocf-data-sampler
These datapipes are designed to be composable and modular, and follow the same setup as for the in-built Pytorch Datapipes. There are some great docs on how they can be composed and used here.
End to end examples are given in ocf_datapipes.training and
ocf_datapipes.production.
This repo is organized as follows. The general flow of data loading and
processing goes from the
ocf_datapipes.load -> .select -> .transform.xarray -> .convert and then
optionally .transform.numpy.
training and production contain datapipes that go through all the steps of
loading the config file, data, selecting and transforming data, and returning
the numpy data to the PyTorch dataloader.
Modules have their own README's as well to go into further detail. This is part of a tree showing the general repo structure.
.
└── ocf_datapipes/
├── batch/
│ ├── fake/
├── config/
├── convert/
│ ├── numpy/
│ ├── numpy_batch/
├── experimental/
├── load/
│ ├── gsp/
│ ├── nwp/
│ ├── pv/
│ ├── satellite.py
│ ├── sensor/
│ ├── topographic.py
│ └── wind/
├── production/
├── select/
├── training/
│ ├── common.py
│ ├── example/
│ ├── metnet/
│ ├── pseudo_irradience.py
│ ├── pvnet.py
│ ├── pvnet_site.py
│ └── windnet.py
├── transform/
│ ├── numpy_batch/
│ └── xarray/
├── utils/
│ ├── gsp_shape/
│ ├── split/
├── validation/
└── visualization/
A general outline for a new DataPipe should go something like this:
from torch.utils.data.datapipes.datapipe import IterDataPipe
from torch.utils.data.datapipes._decorator import functional_datapipe
@functional_datapipe("<pipelet_name>")
class <PipeletName>IterDataPipe(IterDataPipe):
def __init__(self):
pass
def __iter__(self):
pass## The below code snippets have been picked from ocf_datapipes\training\pv_satellite_nwp.py file
# 1. read the configuration model for the dataset, detailing what kind of data is the dataset holding, e.g., pv, pv+satellite, pv+satellite+nwp, etc
config_datapipe = OpenConfiguration(configuration)
# 2. create respective data pipes for pv, nwp and satellite
pv_datapipe, pv_location_datapipe = (OpenPVFromNetCDF(pv=configuration.input_data.pv).pv_fill_night_nans().fork(2))
nwp_datapipe = OpenNWP(configuration.input_data.nwp.nwp_zarr_path)
satellite_datapipe = OpenSatellite(zarr_path=configuration.input_data.satellite.satellite_zarr_path)
# 3. pick all or random location data based on pv data pipeline
location_datapipes = pv_location_datapipe.pick_locations().fork(4, buffer_size=BUFFER_SIZE)
# 4. for the above picked locations get their respective spatial space slices from all the data pipes
pv_datapipe, pv_time_periods_datapipe, pv_t0_datapipe = pv_datapipe.select_spatial_slice_meters(...)
nwp_datapipe, nwp_time_periods_datapipe = nwp_datapipe.select_spatial_slice_pixels(...)
satellite_datapipe, satellite_time_periods_datapipe = satellite_datapipe.select_spatial_slice_pixels(...)
# 5. get contiguous time period data for the above picked locations
pv_time_periods_datapipe = pv_time_periods_datapipe.find_contiguous_t0_time_periods(...)
nwp_time_periods_datapipe = nwp_time_periods_datapipe.find_contiguous_t0_time_periods(...)
satellite_time_periods_datapipe = satellite_time_periods_datapipe.find_contiguous_t0_time_periods(...)
# 6. since all the datapipes have different sampling period for their data, lets find the time that is common between all the data pipes
overlapping_datapipe = pv_time_periods_datapipe.filter_to_overlapping_time_periods(secondary_datapipes=[nwp_time_periods_datapipe, satellite_time_periods_datapipe])
# 7. take time slices for the above overlapping time from all the data pipes
pv_datapipe = pv_datapipe.select_time_slice(...)
nwp_datapipe = nwp_datapipe.convert_to_nwp_target_time(...)
satellite_datapipe = satellite_datapipe.select_time_slice(...)
# 8. Finally join all the data pipes together
combined_datapipe = MergeNumpyModalities([nwp_datapipe, pv_datapipe, satellite_datapipe])Ensure you have the dev requirements installed:
pip install requirements-dev.txt
Run pytest to run th tests
For new datapipes being developed for new models or input modalities, to
somewhat separate the more experimental and in development datapipes from the
ones better tested for production purposes, there is an
ocf_datapipes.experimental namespace for developing these more research-y
datapipes. These datapipes might not, and probably are not, tested. Once the
model(s) using them are in production, they should be upgraded to one of the
other namespaces and have tests added.
If you find this code useful, please cite the following:
@misc{ocf_datapipes,
author = {Bieker, Jacob, and Dudfield, Peter, and Kelly, Jack},
title = {OCF Datapipes},
year = {2022},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/openclimatefix/ocf_datapipes}},
}
Thanks goes to these wonderful people (emoji key):
 Jacob Bieker 💻 |
 Raj 💻 |
 James Fulton 💻 |
 Ritesh Mehta 💻 |
 Chris Briggs 💻 |
 Markus 💻 |
 Code/OS 💻 📖 |
 Sukh-P 📖 |
This project follows the all-contributors specification. Contributions of any kind welcome!
