7 releases (4 breaking)
0.5.1 | Oct 13, 2022 |
---|---|
0.5.0 | Apr 1, 2022 |
0.4.0 | Feb 26, 2022 |
0.3.0 | Jan 3, 2022 |
0.1.0 | Sep 20, 2020 |
#699 in Machine learning
Used in tsuga
21KB
334 lines
cifar-ten
This library parses the binary files of the CIFAR-10 data set and returns them as a tuple struct
CifarResult
:(Vec<u8>, Vec<u8>, Vec<u8>, Vec<u8>)
which is organized as(train_data, train_labels, test_data, test_labels)
Convenience methods for converting these to the Rust ndarray
numeric arrays are provided using the to_ndarray_0xx
feature flag, as
well as for automatically downloading binary training data from a remote url.
// $ cargo build --features=download,to_ndarray_015
use cifar_ten::*;
fn main() {
let (train_data, train_labels, test_data, test_labels) = Cifar10::default()
.download_and_extract(true)
.encode_one_hot(true)
.build()
.unwrap()
.to_ndarray::<f32>()
.expect("Failed to build CIFAR-10 data");
}
Various ndarray
versions can be used with the following feature flags:
version | feature flag |
---|---|
0.15 |
to_ndarray_015 |
0.14 |
to_ndarray_014 |
0.13 |
to_ndarray_013 |
A tar.gz
file with the original binaries can be found here. The crate's author also
provides several ML data mirrors here which are used for running tests on this library. Please feel free to use,
but should you expect to make heavy use of these files, please consider creating your own mirror.
If you'd like to verify that the correct images and labels are being provided, the examples/preview_images.rs
file using show-image
to
preview a RGB representation of a given image with the corresponding one-hot formatted label.
Note: Early commits included the dataset, which will make the download size large. For development, it's suggested to clone using
$ git clone --depth=1 https://github.com/quietlychris/cifar-ten.git
Dependencies
~0.2–10MB
~112K SLoC