ocr 0.1.2

A pure Rust CLI OCR tool for printed text extraction
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
231
232

//! INT8 quantization for edge deployment
//!
//! Provides symmetric per-tensor quantization:
//!   quantized = round(clamp(value / scale, -127, 127))
//!   dequantized = quantized * scale
//!
//! Benefits: 4x memory reduction, faster cache-friendly inference on CPU.

use ndarray::{Array1, Array2};

/// Quantized tensor with symmetric INT8 weights
#[derive(Debug, Clone)]
pub struct QuantizedTensor {
    pub weights: Vec<i8>,
    pub scale: f32,
    pub shape: (usize, usize), // rows, cols
}

impl QuantizedTensor {
    /// Create from a flattened INT8 buffer with known shape
    pub fn from_vec(weights: Vec<i8>, scale: f32, rows: usize, cols: usize) -> Self {
        Self {
            weights,
            scale,
            shape: (rows, cols),
        }
    }

    /// Convert back to f32 Array2
    pub fn to_array2(&self) -> Array2<f32> {
        let (rows, cols) = self.shape;
        let mut arr = Array2::zeros((rows, cols));
        for i in 0..rows {
            for j in 0..cols {
                arr[[i, j]] = self.weights[i * cols + j] as f32 * self.scale;
            }
        }
        arr
    }

    /// Number of parameters
    pub fn len(&self) -> usize {
        self.weights.len()
    }

    /// Memory size in bytes
    pub fn size_bytes(&self) -> usize {
        self.weights.len() * std::mem::size_of::<i8>()
    }
}

/// Quantize an f32 Array2 to symmetric INT8
pub fn quantize_array2(arr: &Array2<f32>) -> QuantizedTensor {
    let (rows, cols) = (arr.nrows(), arr.ncols());

    // Find max absolute value for scale
    let max_abs = arr.iter().map(|&v| v.abs()).fold(0.0f32, f32::max);
    if max_abs < 1e-8 {
        return QuantizedTensor {
            weights: vec![0i8; rows * cols],
            scale: 1.0,
            shape: (rows, cols),
        };
    }

    let scale = max_abs / 127.0;
    let mut weights = Vec::with_capacity(rows * cols);

    for i in 0..rows {
        for j in 0..cols {
            let q = (arr[[i, j]] / scale).round().clamp(-127.0, 127.0) as i8;
            weights.push(q);
        }
    }

    QuantizedTensor {
        weights,
        scale,
        shape: (rows, cols),
    }
}

/// Quantize an f32 Array1 to symmetric INT8
pub fn quantize_array1(arr: &Array1<f32>) -> (Vec<i8>, f32) {
    let max_abs = arr.iter().map(|&v| v.abs()).fold(0.0f32, f32::max);
    if max_abs < 1e-8 {
        return (vec![0i8; arr.len()], 1.0);
    }

    let scale = max_abs / 127.0;
    let weights: Vec<i8> = arr
        .iter()
        .map(|&v| (v / scale).round().clamp(-127.0, 127.0) as i8)
        .collect();

    (weights, scale)
}

/// Dequantize INT8 weights back to f32 Array2
pub fn dequantize_to_array2(weights: &[i8], scale: f32, rows: usize, cols: usize) -> Array2<f32> {
    let mut arr = Array2::zeros((rows, cols));
    for i in 0..rows {
        for j in 0..cols {
            arr[[i, j]] = weights[i * cols + j] as f32 * scale;
        }
    }
    arr
}

/// Dequantize INT8 weights back to f32 Array1
pub fn dequantize_to_array1(weights: &[i8], scale: f32) -> Array1<f32> {
    Array1::from_iter(weights.iter().map(|&w| w as f32 * scale))
}

/// Matrix multiplication with quantized weights: output = input @ quantized_weight.T
/// Returns f32 output. Input stays f32; weights are dequantized on-the-fly per element.
/// This saves memory bandwidth (1 byte reads vs 4 byte reads) at the cost of
/// per-element dequantization.
pub fn quantized_matmul(input: &Array2<f32>, qw: &QuantizedTensor) -> Array2<f32> {
    let (batch, in_features) = input.dim();
    let (out_features, _) = qw.shape;
    assert_eq!(in_features, qw.shape.1, "Input features must match weight cols");

    let mut output = Array2::zeros((batch, out_features));
    let scale = qw.scale;

    for b in 0..batch {
        for o in 0..out_features {
            let mut sum = 0.0f32;
            for i in 0..in_features {
                let w = qw.weights[o * in_features + i] as f32 * scale;
                sum += input[[b, i]] * w;
            }
            output[[b, o]] = sum;
        }
    }

    output
}

/// Compute compression ratio: original f32 size / quantized size
pub fn compression_ratio(element_count: usize) -> f32 {
    (element_count * std::mem::size_of::<f32>()) as f32
            / (element_count * std::mem::size_of::<i8>()) as f32
}

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

    #[test]
    fn test_quantize_dequantize_roundtrip() {
        let mut arr = Array2::zeros((3, 4));
        arr[[0, 0]] = 1.0;
        arr[[0, 1]] = -0.5;
        arr[[1, 2]] = 2.0;
        arr[[2, 3]] = -1.5;

        let qt = quantize_array2(&arr);
        let recovered = qt.to_array2();

        // Max error should be <= scale (quantization step)
        let max_err = arr
            .iter()
            .zip(recovered.iter())
            .map(|(a, b)| (a - b).abs())
            .fold(0.0f32, f32::max);
        assert!(max_err <= qt.scale + 1e-6, "Max error {} > scale {}", max_err, qt.scale);
    }

    #[test]
    fn test_quantize_all_zeros() {
        let arr = Array2::zeros((5, 5));
        let qt = quantize_array2(&arr);
        assert!(qt.weights.iter().all(|&w| w == 0));
        assert_eq!(qt.scale, 1.0);
    }

    #[test]
    fn test_compression_ratio() {
        let ratio = compression_ratio(1000);
        assert!((ratio - 4.0).abs() < 0.01, "Expected ~4x compression, got {}", ratio);
    }

    #[test]
    fn test_quantized_matmul() {
        let mut w = Array2::zeros((2, 3));
        w[[0, 0]] = 1.0;
        w[[0, 1]] = 2.0;
        w[[0, 2]] = 3.0;
        w[[1, 0]] = -1.0;
        w[[1, 1]] = 0.0;
        w[[1, 2]] = 1.0;

        let qw = quantize_array2(&w);
        let input = Array2::from_shape_vec((1, 3), vec![1.0, 1.0, 1.0]).unwrap();
        let output = quantized_matmul(&input, &qw);

        // Expected: [6.0, 0.0] (with some quantization error)
        assert!((output[[0, 0]] - 6.0).abs() < 0.1, "Expected ~6.0, got {}", output[[0, 0]]);
        assert!(output[[0, 1]].abs() < 0.1, "Expected ~0.0, got {}", output[[0, 1]]);
    }

    #[test]
    fn test_quantized_tensor_size() {
        let arr = Array2::zeros((100, 100));
        let qt = quantize_array2(&arr);
        assert_eq!(qt.size_bytes(), 10000);
        assert_eq!(qt.len(), 10000);
    }

    #[test]
    fn test_quantize_array1() {
        let arr = Array1::from_vec(vec![-2.0, -1.0, 0.0, 1.0, 2.0]);
        let (weights, scale) = quantize_array1(&arr);
        assert_eq!(weights.len(), 5);
        assert!(scale > 0.0);
        // Max should map to 127
        assert_eq!(weights[4], 127);
        assert_eq!(weights[0], -127);
    }

    #[test]
    fn test_dequantize_to_array1() {
        let weights = vec![-127i8, 0, 127];
        let scale = 0.1;
        let arr = dequantize_to_array1(&weights, scale);
        assert!((arr[0] - (-12.7)).abs() < 1e-5);
        assert!((arr[1] - 0.0).abs() < 1e-5);
        assert!((arr[2] - 12.7).abs() < 1e-5);
    }
}