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

use crate::tensor::Tensor;
use super::OpTrait;


//
// Common Cost function
//
pub enum Reduction{
    None,
    Mean,
    Sum,
}

/// MSELoss
/// The left-most dimension is the N.
pub struct MSELoss {
    reduction: Reduction,
}
impl MSELoss {
    pub fn new() -> MSELoss {
        MSELoss {
            reduction: Reduction::None,
        }
    }
}
impl OpTrait for MSELoss {
    fn get_name(&self) -> String {
        "MSE".to_string()
    }
    fn apply(&mut self, input: &[&Tensor], output: &[&Tensor]) {
        // TODO: wait for Tensor to have lazy evaluation for elemwise operation.
        let tmp = input[0].sub(input[1]);
        let tmp2 = tmp.mul(&tmp);
        let tmp3 = tmp2.sum();
        let ret = tmp3.div(&input[0].get_N().mul(&input[0].get_C()));
        output[0].swap(ret);
    }
    fn grad(&self, input: &[&Tensor], output_grad: &[&Tensor], input_grad: &[&Tensor]) {
        
        if input.len() < 2 {
            panic!("MSELoss expect two input, get {}", input.len());
        }
        if input_grad.len() < 2 {
            panic!("MSELoss expect two input gradient tensor, get {}", input_grad.len());
        }
        if output_grad.len() < 1 {
            panic!("MSELoss expect one output gradient, get {}", output_grad.len());
        }
        if ! input[0].same_shape(input[1]) {
            panic!("MSELoss expect two input have the same shape, get {:?}, {:?}", input[0].size(), input[1].size());
        }


        let tmp1 = input[0].sub(input[1]);
        let tmp2 = tmp1.div(&input[0].numel_tensor());
        let tmp3 = tmp2.mul(output_grad[0]);
        input_grad[0].swap(tmp3);

        let tmp1 = input[1].sub(input[0]);
        let tmp2 = tmp1.div(&input[0].numel_tensor());
        let tmp3 = tmp2.mul(output_grad[0]);
        input_grad[1].swap(tmp3);
    }

    fn get_values(&self) -> Vec<&Tensor> {
        Vec::new()
    }
    fn set_values(&self, v: &[Tensor]) {
    }

    fn get_grads(&self) -> Vec<&Tensor> {
        Vec::new()
    }
}


/// This loss combines a Sigmoid layer and the BCELoss in one single class.
/// This version is more numerically stable than using a plain Sigmoid followed
/// by a BCELoss as, by combining the operations into one layer,
/// we take advantage of the log-sum-exp trick for numerical stability.
///
/// -y log (1/(1 + exp(-x))) - (1-y) log(1 - 1/(1 + exp(-x)))
/// 
/// Prediction comes first, label comes second.
pub struct BCEWithLogitsLoss {
    
}
impl BCEWithLogitsLoss {
    pub fn new() -> BCEWithLogitsLoss {
        BCEWithLogitsLoss {
        }
    }
}
impl OpTrait for BCEWithLogitsLoss {
    
    fn get_name(&self) -> String {
        "BCEWithLogitsLoss".to_string()
    }

    /// The first is the prediction, the second input is the label
    fn apply(&mut self, input: &[&Tensor], output: &[&Tensor]) {
        if input.len() < 2 {
            panic!("{} expect two input, get {}", self.get_name(), input.len());
        }
        let ret = input[1].mul(&input[0].neg().log1pexp())
            .add(&(input[1].neg().add(&input[1].ones_like())).mul(&input[0].log1pexp()));
        output[0].swap(ret);
    }
    
    /// Given the forward input value and backward output_grad,
    /// Update weight gradient.
    /// return backward input gradeint.
    fn grad(&self, input: &[&Tensor], output_grad: &[&Tensor], input_grad: &[&Tensor]) {
        // ddx y log (1 + exp(-x)) = -y  / (1 + exp(x))
        // ddx (1-y) log (1 + exp(x)) = (1-y) / (1 + exp(-x))
    }

    /// access weight values
    fn get_values(&self) -> Vec<&Tensor> {
        Vec::new()
    }
    
    fn set_values(&self, v: &[Tensor]) {
        
    }
    
    /// access gradient values
    fn get_grads(&self) -> Vec<&Tensor> {
        Vec::new()
    }
}