Struct PTQEngine

Source

pub struct PTQEngine { /* private fields */ }

Expand description

Post-training quantization engine.

Implementations§

Source §

impl PTQEngine

Source

pub fn new(num_bits: u32, signed: bool) -> Self

Create a new PTQ engine.

Examples found in repository ?

examples/full_inference_pipeline.rs (line 122)

113fn quantize_model() {
114    println!("3. Model Quantization");
115
116    let calibration_data = vec![
117        Tensor::random(&[1, 224, 224, 3]).data,
118        Tensor::random(&[1, 224, 224, 3]).data,
119        Tensor::random(&[1, 224, 224, 3]).data,
120    ];
121
122    let ptq = PTQEngine::new(8, false);
123    match ptq.calibrate(&calibration_data) {
124        Ok(converter) => {
125            println!("   ✓ Calibration complete");
126            
127            let sample = &calibration_data[0];
128            let stats = QuantStats::from_tensor(sample);
129            println!("   Calibration Stats:");
130            println!("   - Min: {:.6}", stats.min_val);
131            println!("   - Max: {:.6}", stats.max_val);
132            println!("   - Mean: {:.6}", stats.mean_val);
133            println!("   - Std: {:.6}", stats.std_val);
134            
135            match converter.quantize_tensor(sample) {
136                Ok(quantized) => {
137                    println!("   ✓ Quantization complete: {} values", quantized.len());
138                    println!("   Compression: {:.2}x\n", 
139                        (sample.len() * 4) as f64 / quantized.len() as f64
140                    );
141                }
142                Err(e) => println!("   ✗ Quantization failed: {}\n", e),
143            }
144        }
145        Err(e) => println!("   ✗ Calibration failed: {}\n", e),
146    }
147}

Source

pub fn calibrate(&self, sample_data: &[ArrayD<f32>]) -> Result<QuantConverter>

Calibrate on sample data.

Examples found in repository ?

examples/full_inference_pipeline.rs (line 123)

113fn quantize_model() {
114    println!("3. Model Quantization");
115
116    let calibration_data = vec![
117        Tensor::random(&[1, 224, 224, 3]).data,
118        Tensor::random(&[1, 224, 224, 3]).data,
119        Tensor::random(&[1, 224, 224, 3]).data,
120    ];
121
122    let ptq = PTQEngine::new(8, false);
123    match ptq.calibrate(&calibration_data) {
124        Ok(converter) => {
125            println!("   ✓ Calibration complete");
126            
127            let sample = &calibration_data[0];
128            let stats = QuantStats::from_tensor(sample);
129            println!("   Calibration Stats:");
130            println!("   - Min: {:.6}", stats.min_val);
131            println!("   - Max: {:.6}", stats.max_val);
132            println!("   - Mean: {:.6}", stats.mean_val);
133            println!("   - Std: {:.6}", stats.std_val);
134            
135            match converter.quantize_tensor(sample) {
136                Ok(quantized) => {
137                    println!("   ✓ Quantization complete: {} values", quantized.len());
138                    println!("   Compression: {:.2}x\n", 
139                        (sample.len() * 4) as f64 / quantized.len() as f64
140                    );
141                }
142                Err(e) => println!("   ✗ Quantization failed: {}\n", e),
143            }
144        }
145        Err(e) => println!("   ✗ Calibration failed: {}\n", e),
146    }
147}