Vision dataset (#179)

* Add some readers for the mnist dataset.

* Import the cifar and mnist dataset.

candle-nn/src/lib.rs

@@ -8,6 +8,7 @@ pub mod layer_norm;
 pub mod linear;
 pub mod optim;
 pub mod var_builder;
+pub mod vision;
 
 pub use activation::Activation;
 pub use conv::{Conv1d, Conv1dConfig};

candle-nn/src/vision/cifar.rs

@@ -0,0 +1,62 @@
+//! The CIFAR-10 dataset.
+//!
+//! The files can be downloaded from the following page:
+//! <https://www.cs.toronto.edu/~kriz/cifar.html>
+//! The binary version of the dataset is used.
+use crate::vision::Dataset;
+use candle::{DType, Device, Result, Tensor};
+use std::fs::File;
+use std::io::{BufReader, Read};
+
+const W: usize = 32;
+const H: usize = 32;
+const C: usize = 3;
+const BYTES_PER_IMAGE: usize = W * H * C + 1;
+const SAMPLES_PER_FILE: usize = 10000;
+
+fn read_file(filename: &std::path::Path) -> Result<(Tensor, Tensor)> {
+    let mut buf_reader = BufReader::new(File::open(filename)?);
+    let mut data = vec![0u8; SAMPLES_PER_FILE * BYTES_PER_IMAGE];
+    buf_reader.read_exact(&mut data)?;
+    let mut images = vec![];
+    let mut labels = vec![];
+    for index in 0..SAMPLES_PER_FILE {
+        let content_offset = BYTES_PER_IMAGE * index;
+        labels.push(data[content_offset]);
+        images.push(&data[1 + content_offset..content_offset + BYTES_PER_IMAGE]);
+    }
+    let images: Vec<u8> = images
+        .iter()
+        .copied()
+        .flatten()
+        .copied()
+        .collect::<Vec<_>>();
+    let labels = Tensor::from_vec(labels, SAMPLES_PER_FILE, &Device::Cpu)?;
+    let images = Tensor::from_vec(images, (SAMPLES_PER_FILE, C, H, W), &Device::Cpu)?;
+    let images = (images.to_dtype(DType::F32)? / 255.)?;
+    Ok((images, labels))
+}
+
+pub fn load_dir<T: AsRef<std::path::Path>>(dir: T) -> Result<Dataset> {
+    let dir = dir.as_ref();
+    let (test_images, test_labels) = read_file(&dir.join("test_batch.bin"))?;
+    let train_images_and_labels = [
+        "data_batch_1.bin",
+        "data_batch_2.bin",
+        "data_batch_3.bin",
+        "data_batch_4.bin",
+        "data_batch_5.bin",
+    ]
+    .iter()
+    .map(|x| read_file(&dir.join(x)))
+    .collect::<Result<Vec<_>>>()?;
+    let (train_images, train_labels): (Vec<_>, Vec<_>) =
+        train_images_and_labels.into_iter().unzip();
+    Ok(Dataset {
+        train_images: Tensor::cat(&train_images, 0)?,
+        train_labels: Tensor::cat(&train_labels, 0)?,
+        test_images,
+        test_labels,
+        labels: 10,
+    })
+}

candle-nn/src/vision/mnist.rs

@@ -0,0 +1,65 @@
+//! The MNIST hand-written digit dataset.
+//!
+//! The files can be obtained from the following link:
+//! <http://yann.lecun.com/exdb/mnist/>
+use candle::{DType, Device, Result, Tensor};
+use std::fs::File;
+use std::io::{self, BufReader, Read};
+
+fn read_u32<T: Read>(reader: &mut T) -> Result<u32> {
+    let mut b = vec![0u8; 4];
+    reader.read_exact(&mut b)?;
+    let (result, _) = b.iter().rev().fold((0u64, 1u64), |(s, basis), &x| {
+        (s + basis * u64::from(x), basis * 256)
+    });
+    Ok(result as u32)
+}
+
+fn check_magic_number<T: Read>(reader: &mut T, expected: u32) -> Result<()> {
+    let magic_number = read_u32(reader)?;
+    if magic_number != expected {
+        Err(io::Error::new(
+            io::ErrorKind::Other,
+            format!("incorrect magic number {magic_number} != {expected}"),
+        ))?;
+    }
+    Ok(())
+}
+
+fn read_labels(filename: &std::path::Path) -> Result<Tensor> {
+    let mut buf_reader = BufReader::new(File::open(filename)?);
+    check_magic_number(&mut buf_reader, 2049)?;
+    let samples = read_u32(&mut buf_reader)?;
+    let mut data = vec![0u8; samples as usize];
+    buf_reader.read_exact(&mut data)?;
+    let samples = data.len();
+    Tensor::from_vec(data, samples, &Device::Cpu)
+}
+
+fn read_images(filename: &std::path::Path) -> Result<Tensor> {
+    let mut buf_reader = BufReader::new(File::open(filename)?);
+    check_magic_number(&mut buf_reader, 2051)?;
+    let samples = read_u32(&mut buf_reader)? as usize;
+    let rows = read_u32(&mut buf_reader)? as usize;
+    let cols = read_u32(&mut buf_reader)? as usize;
+    let data_len = samples * rows * cols;
+    let mut data = vec![0u8; data_len];
+    buf_reader.read_exact(&mut data)?;
+    let tensor = Tensor::from_vec(data, (samples, rows * cols), &Device::Cpu)?;
+    tensor.to_dtype(DType::F32)? / 255.
+}
+
+pub fn load_dir<T: AsRef<std::path::Path>>(dir: T) -> Result<crate::vision::Dataset> {
+    let dir = dir.as_ref();
+    let train_images = read_images(&dir.join("train-images-idx3-ubyte"))?;
+    let train_labels = read_labels(&dir.join("train-labels-idx1-ubyte"))?;
+    let test_images = read_images(&dir.join("t10k-images-idx3-ubyte"))?;
+    let test_labels = read_labels(&dir.join("t10k-labels-idx1-ubyte"))?;
+    Ok(crate::vision::Dataset {
+        train_images,
+        train_labels,
+        test_images,
+        test_labels,
+        labels: 10,
+    })
+}

candle-nn/src/vision/mod.rs

@@ -0,0 +1,12 @@
+use candle::Tensor;
+
+pub struct Dataset {
+    pub train_images: Tensor,
+    pub train_labels: Tensor,
+    pub test_images: Tensor,
+    pub test_labels: Tensor,
+    pub labels: usize,
+}
+
+pub mod cifar;
+pub mod mnist;