flodl 0.5.0

floDl — a flow-graph deep learning framework built on libtorch
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230

//! MNIST dataset parser.
//!
//! Parses the IDX binary format (gzip-compressed) into tensors.
//! Images are normalized to [0, 1] as Float32, labels are Int64.
//!
//! # Example
//!
//! ```ignore
//! let images_gz = std::fs::read("train-images-idx3-ubyte.gz")?;
//! let labels_gz = std::fs::read("train-labels-idx1-ubyte.gz")?;
//! let mnist = Mnist::parse(&images_gz, &labels_gz)?;
//! // mnist.images: [60000, 1, 28, 28] Float32
//! // mnist.labels: [60000] Int64
//! ```

use std::io::Read;

use crate::data::BatchDataSet;
use crate::tensor::{Device, Result, Tensor, TensorError};

/// Parsed MNIST dataset (train or test split).
pub struct Mnist {
    /// Images as `[N, 1, 28, 28]` Float32, normalized to [0, 1].
    pub images: Tensor,
    /// Labels as `[N]` Int64 (0-9).
    pub labels: Tensor,
}

impl Mnist {
    /// Parse gzip-compressed IDX files into tensors.
    ///
    /// Accepts the raw bytes of `*-images-idx3-ubyte.gz` and
    /// `*-labels-idx1-ubyte.gz` files.
    pub fn parse(images_gz: &[u8], labels_gz: &[u8]) -> Result<Self> {
        let images_raw = gunzip(images_gz)?;
        let labels_raw = gunzip(labels_gz)?;

        // Parse images: magic(4) + count(4) + rows(4) + cols(4) + pixels
        if images_raw.len() < 16 {
            return Err(TensorError::new("MNIST images: file too short for header"));
        }
        let magic = read_u32_be(&images_raw[0..4]);
        if magic != 2051 {
            return Err(TensorError::new(&format!(
                "MNIST images: bad magic {magic}, expected 2051"
            )));
        }
        let n = read_u32_be(&images_raw[4..8]) as usize;
        let rows = read_u32_be(&images_raw[8..12]) as usize;
        let cols = read_u32_be(&images_raw[12..16]) as usize;
        let pixel_count = n * rows * cols;
        if images_raw.len() < 16 + pixel_count {
            return Err(TensorError::new(&format!(
                "MNIST images: expected {} pixels, got {}",
                pixel_count,
                images_raw.len() - 16
            )));
        }

        // Convert u8 pixels to f32 normalized [0, 1]
        let pixels: Vec<f32> = images_raw[16..16 + pixel_count]
            .iter()
            .map(|&b| b as f32 / 255.0)
            .collect();
        let images = Tensor::from_f32(&pixels, &[n as i64, 1, rows as i64, cols as i64], Device::CPU)?;

        // Parse labels: magic(4) + count(4) + labels
        if labels_raw.len() < 8 {
            return Err(TensorError::new("MNIST labels: file too short for header"));
        }
        let magic = read_u32_be(&labels_raw[0..4]);
        if magic != 2049 {
            return Err(TensorError::new(&format!(
                "MNIST labels: bad magic {magic}, expected 2049"
            )));
        }
        let n_labels = read_u32_be(&labels_raw[4..8]) as usize;
        if n_labels != n {
            return Err(TensorError::new(&format!(
                "MNIST: {n} images but {n_labels} labels"
            )));
        }
        if labels_raw.len() < 8 + n {
            return Err(TensorError::new("MNIST labels: truncated"));
        }

        let label_vals: Vec<i64> = labels_raw[8..8 + n]
            .iter()
            .map(|&b| b as i64)
            .collect();
        let labels = Tensor::from_i64(&label_vals, &[n as i64], Device::CPU)?;

        Ok(Mnist { images, labels })
    }

    /// Number of samples.
    pub fn len(&self) -> usize {
        self.images.shape()[0] as usize
    }

    /// True if the dataset is empty.
    pub fn is_empty(&self) -> bool {
        self.len() == 0
    }
}

impl BatchDataSet for Mnist {
    fn len(&self) -> usize {
        self.images.shape()[0] as usize
    }

    fn get_batch(&self, indices: &[usize]) -> Result<Vec<Tensor>> {
        let idx: Vec<i64> = indices.iter().map(|&i| (i % self.len()) as i64).collect();
        let idx_tensor = Tensor::from_i64(&idx, &[idx.len() as i64], Device::CPU)?;
        let images = self.images.index_select(0, &idx_tensor)?;
        let labels = self.labels.index_select(0, &idx_tensor)?;
        Ok(vec![images, labels])
    }
}

/// Decompress gzip data.
fn gunzip(data: &[u8]) -> Result<Vec<u8>> {
    let mut decoder = flate2::read::GzDecoder::new(data);
    let mut out = Vec::new();
    decoder
        .read_to_end(&mut out)
        .map_err(|e| TensorError::new(&format!("gzip decompression failed: {e}")))?;
    Ok(out)
}

/// Read a big-endian u32 from a 4-byte slice.
fn read_u32_be(bytes: &[u8]) -> u32 {
    u32::from_be_bytes([bytes[0], bytes[1], bytes[2], bytes[3]])
}

#[cfg(test)]
mod tests {
    use super::*;

    /// Build a minimal valid IDX3 (images) file.
    fn make_idx3(n: u32, rows: u32, cols: u32, pixels: &[u8]) -> Vec<u8> {
        let mut buf = Vec::new();
        buf.extend_from_slice(&2051u32.to_be_bytes());
        buf.extend_from_slice(&n.to_be_bytes());
        buf.extend_from_slice(&rows.to_be_bytes());
        buf.extend_from_slice(&cols.to_be_bytes());
        buf.extend_from_slice(pixels);
        buf
    }

    /// Build a minimal valid IDX1 (labels) file.
    fn make_idx1(n: u32, labels: &[u8]) -> Vec<u8> {
        let mut buf = Vec::new();
        buf.extend_from_slice(&2049u32.to_be_bytes());
        buf.extend_from_slice(&n.to_be_bytes());
        buf.extend_from_slice(labels);
        buf
    }

    /// Gzip-compress bytes.
    fn gzip(data: &[u8]) -> Vec<u8> {
        use flate2::write::GzEncoder;
        use flate2::Compression;
        use std::io::Write;
        let mut enc = GzEncoder::new(Vec::new(), Compression::fast());
        enc.write_all(data).unwrap();
        enc.finish().unwrap()
    }

    #[test]
    fn parse_small_mnist() {
        // 2 images of 3x3 pixels
        let pixels = vec![0u8, 128, 255, 0, 0, 0, 255, 255, 255, 10, 20, 30, 40, 50, 60, 70, 80, 90];
        let images_raw = make_idx3(2, 3, 3, &pixels);
        let labels_raw = make_idx1(2, &[3, 7]);

        let images_gz = gzip(&images_raw);
        let labels_gz = gzip(&labels_raw);

        let mnist = Mnist::parse(&images_gz, &labels_gz).unwrap();

        assert_eq!(mnist.images.shape(), &[2, 1, 3, 3]);
        assert_eq!(mnist.labels.shape(), &[2]);

        // Check first pixel normalized: 0/255 = 0.0
        let first = mnist.images.select(0, 0).unwrap();
        let val: f64 = first.select(0, 0).unwrap()
            .select(0, 0).unwrap()
            .select(0, 0).unwrap()
            .item().unwrap();
        assert!((val - 0.0).abs() < 1e-6);

        // Check second pixel: 128/255
        let val: f64 = first.select(0, 0).unwrap()
            .select(0, 0).unwrap()
            .select(0, 1).unwrap()
            .item().unwrap();
        assert!((val - 128.0 / 255.0).abs() < 1e-4);

        // Check labels
        let l0 = mnist.labels.select(0, 0).unwrap().to_i64_vec().unwrap()[0];
        let l1 = mnist.labels.select(0, 1).unwrap().to_i64_vec().unwrap()[0];
        assert_eq!(l0, 3);
        assert_eq!(l1, 7);
    }

    #[test]
    fn get_batch_wraps_indices() {
        let pixels = vec![0u8; 2 * 3 * 3];
        let images_raw = make_idx3(2, 3, 3, &pixels);
        let labels_raw = make_idx1(2, &[0, 1]);

        let mnist = Mnist::parse(&gzip(&images_raw), &gzip(&labels_raw)).unwrap();

        // Index 5 wraps to 5 % 2 = 1
        let batch = mnist.get_batch(&[0, 5]).unwrap();
        assert_eq!(batch[0].shape(), &[2, 1, 3, 3]);
        assert_eq!(batch[1].shape(), &[2]);
    }

    #[test]
    fn bad_magic_rejected() {
        let mut images_raw = make_idx3(1, 2, 2, &[0; 4]);
        images_raw[3] = 99; // corrupt magic
        let labels_raw = make_idx1(1, &[0]);

        let result = Mnist::parse(&gzip(&images_raw), &gzip(&labels_raw));
        assert!(result.is_err());
    }
}