MNIST
So we now have downloaded the MNIST parquet files, let's put them in a simple struct.
let test_samples = 10_000;
let mut test_buffer_images: Vec<u8> = Vec::with_capacity(test_samples * 784);
let mut test_buffer_labels: Vec<u8> = Vec::with_capacity(test_samples);
for row in test_parquet{
for (_name, field) in row?.get_column_iter() {
if let parquet::record::Field::Group(subrow) = field {
for (_name, field) in subrow.get_column_iter() {
if let parquet::record::Field::Bytes(value) = field {
let image = image::load_from_memory(value.data()).unwrap();
test_buffer_images.extend(image.to_luma8().as_raw());
}
}
}else if let parquet::record::Field::Long(label) = field {
test_buffer_labels.push(*label as u8);
}
}
}
let test_images = (Tensor::from_vec(test_buffer_images, (test_samples, 784), &Device::Cpu)?.to_dtype(DType::F32)? / 255.)?;
let test_labels = Tensor::from_vec(test_buffer_labels, (test_samples, ), &Device::Cpu)?;
let train_samples = 60_000;
let mut train_buffer_images: Vec<u8> = Vec::with_capacity(train_samples * 784);
let mut train_buffer_labels: Vec<u8> = Vec::with_capacity(train_samples);
for row in train_parquet{
for (_name, field) in row?.get_column_iter() {
if let parquet::record::Field::Group(subrow) = field {
for (_name, field) in subrow.get_column_iter() {
if let parquet::record::Field::Bytes(value) = field {
let image = image::load_from_memory(value.data()).unwrap();
train_buffer_images.extend(image.to_luma8().as_raw());
}
}
}else if let parquet::record::Field::Long(label) = field {
train_buffer_labels.push(*label as u8);
}
}
}
let train_images = (Tensor::from_vec(train_buffer_images, (train_samples, 784), &Device::Cpu)?.to_dtype(DType::F32)? / 255.)?;
let train_labels = Tensor::from_vec(train_buffer_labels, (train_samples, ), &Device::Cpu)?;
let mnist = candle_datasets::vision::Dataset {
train_images,
train_labels,
test_images,
test_labels,
labels: 10,
};
The parsing of the file and putting it into single tensors requires the dataset to fit the entire memory. It is quite rudimentary, but simple enough for a small dataset like MNIST.