flashlight 0.0.12

neural network library
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

use flashlight::{layers::*, prelude::*};
#[allow(unused)]
use flashlight_tensor::prelude::*;
use async_trait::async_trait;

use rand::prelude::*;
use std::time::Instant;

pub struct NewModel{
    linear1: Linear<Gpu>,
    linear2: Linear<Gpu>,
    linear3: Linear<Gpu>,
    activation: Relu<Cpu>,
    output_activation: Sigmoid,
}



impl NewModel{
    fn new() -> Self{
        Self{
            linear1: Linear::new(2, 16, 0.01),
            linear2: Linear::new(16, 16, 0.01),
            linear3: Linear::new(16, 1, 0.01),
            activation: Relu::new(),
            output_activation: Sigmoid::new(),
        }
    }
    fn grad_output(&self, target: &Tensor<f32>) -> Tensor<f32>{
        self.output_activation.grad_output(target)
    }
}

#[async_trait]
impl ModelGpu for NewModel{
    async fn forward(&mut self, input: Tensor<f32>) -> Tensor<f32> {
        let x = self.linear1.forward(&input).await;
        let x = self.activation.forward(&x);
        let x = self.linear2.forward(&x).await;
        let x = self.activation.forward(&x);
        let x = self.linear3.forward(&x).await;
        
        self.output_activation.forward(&x)
    }
    async fn backward(&mut self, grad_output: Tensor<f32>) {
        let x = self.output_activation.backward(&grad_output);

        let x = self.linear3.backward(&x).await;
        let x = self.activation.backward(&x);
        let x = self.linear2.backward(&x).await;
        let x = self.activation.backward(&x);
        self.linear1.backward(&x).await;
    }
    fn clear_buffers(&mut self){
        self.linear1.clear();
        self.linear2.clear();
        self.linear3.clear();
    }
}

#[tokio::main]
async fn main() {
    let number_of_epochs = 100;
    let number_of_samples = 100;
    let bach_size = 100;

    let mut input_data: DataPreparaton = DataPreparaton::new();

    let mut rng = rand::rng();

    for _i in 0..number_of_samples{
        let num1 = rng.random_range(-100.0..100.0);
        let num2 = rng.random_range(-100.0..100.0);
        let input_vec = vec!{num1, num2};

        let output_vec;
        if num1>num2 {
            output_vec = vec!{1.0};
        }
        else{
            output_vec = vec!{0.0};
        }

        //println!("Input: {}, {}", input_vec[0], input_vec[1]);

        //print!("Output: ");
        for _i in 0..output_vec.len(){
            //print!("{}", output_vec[i]);
        }
        //println!("");

        input_data.append(&Tensor::from_data(&input_vec, &[1, input_vec.len() as u32]).unwrap(), &Tensor::from_data(&output_vec, &[1, output_vec.len() as u32]).unwrap());
    }

    input_data.set_bach_size(bach_size);

    let handler = input_data.to_handler();

    let mut model: NewModel = NewModel::new();

    let mut correct_counter: u32 = 0;

    for i in 0..input_data.data.len(){
        let output_data = model.forward(input_data.data[i].input_data.matrix_transpose().unwrap()).await;

        //println!("Sample {}", i);
        //println!("Data: {}", input_data.input_data[i].matrix_to_string().unwrap());
        //println!("Expected: {}", input_data.output_data[i].matrix_to_string().unwrap());
        //println!("Output: {}", outpud_data[outpud_data.len()-1].matrix_to_string().unwrap());
        if input_data.data[i].output_data.get_data()[0] == 1.0 && output_data.get_data()[0] > 0.5 {
                //println!("correct");
                correct_counter += 1;
        }
        else if input_data.data[i].output_data.get_data()[0] == 0.0 && output_data.get_data()[0] < 0.5 {
                //println!("correct");
                correct_counter += 1;
        }
        else {
            //println!("WRONG");
        }
        //println!("\n");
    }
    println!("Ratio: {}/{}", correct_counter, number_of_samples);

    model.clear_buffers();

    //learning time
    let start = Instant::now();

    let mut forward_duration: std::time::Duration = std::time::Duration::new(0, 0);
    let mut backward_duration: std::time::Duration = std::time::Duration::new(0, 0);
    let mut clear_duration: std::time::Duration = std::time::Duration::new(0, 0);
    let mut cpu_duration: std::time::Duration = std::time::Duration::new(0, 0);

    for epoch in 0..=number_of_epochs{
    
        let mut cost_sum: f32 = 0.0;

        for i in 0..handler.len(){
            let start = Instant::now();
            let input_bach = handler.input_bach(i);
            let output_bach = &handler.output_bach(i);
            cpu_duration += start.elapsed();

            let start = Instant::now();
            let output = model.forward(input_bach).await;
            forward_duration += start.elapsed();

            let start = Instant::now();
            let grad_output = model.grad_output(output_bach);
            cpu_duration += start.elapsed();

            let start = Instant::now();
            model.backward(grad_output).await;
            backward_duration += start.elapsed();
            
            let start = Instant::now();
            //println!("{:?}, {:?}", output.get_data(), handler.output_bach(i).get_data());
            cost_sum += cross_entropy_cost(&output, &handler.output_bach(i)).unwrap();
            //println!("{}", cross_entropy_cost(&output, &handler.output_bach(i)).unwrap());
            cpu_duration += start.elapsed();

            let start = Instant::now();
            model.clear_buffers();
            clear_duration += start.elapsed();
        }
        if epoch % 10 == 0{
            println!("Epoch {}, Cost: {}", epoch, cost_sum/handler.len() as f32);
        }
    }
    let duration: std::time::Duration = start.elapsed();

    let mut correct_counter: u32 = 0;

    for _i in 0..number_of_samples{
        let num1 = rng.random_range(-100.0..100.0);
        let num2 = rng.random_range(-100.0..100.0);
        let input_vec = vec!{num1, num2};

        let output_vec;
        if num1>num2 {
            output_vec = vec!{1.0};
        }
        else{
            output_vec = vec!{0.0};
        }
        let input_data = Tensor::from_data(&input_vec, &[2, 1]).unwrap();
        let output_data = model.forward(input_data.clone()).await;

        //println!("Sample {}", _i);
        //println!("Data: {}", input_data.matrix_transpose().unwrap().matrix_to_string().unwrap());
        //println!("Expected: {}", output_vec[0]);
        //println!("Output: {}", output_data.matrix_to_string().unwrap());
        if output_vec[0] == 1.0 && output_data.get_data()[0] >= 0.5 {
                //println!("correct");
                correct_counter += 1;
        }
        else if output_vec[0] == 0.0 && output_data.get_data()[0] < 0.5 {
                //println!("correct");
                correct_counter += 1;
        }
        else {
            //println!("WRONG");
        }
        //println!("\n");
    }
    println!("Ratio: {}/{}", correct_counter, number_of_samples);
    println!("Learning time: {:?}", duration);
    println!("Forward time: {:?}\nBackward time: {:?}\nClear time: {:?}\nCpu time: {:?}", forward_duration, backward_duration, clear_duration, cpu_duration);
}