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

//! Activation functions.
use arrayfire::*;
use std::convert::TryFrom;

use crate::tensor::*;
use crate::tensor::PrimitiveType;

/// Enumeration of the activation functions.
///
#[derive(hdf5::H5Type, Clone, Copy, Debug)]
#[repr(u8)]
pub enum Activation {
    LeakyReLU = 0,
    Linear = 1,
    ReLU = 2,
    Sigmoid = 3,
    Softmax = 4,
    Tanh = 5,
}

impl TryFrom<u8> for Activation {
    type Error = ();

    fn try_from(v: u8) -> Result<Self, Self::Error> {
        match v {
            x if x == Activation::LeakyReLU as u8 => Ok(Activation::LeakyReLU),
            x if x == Activation::Linear as u8 => Ok(Activation::Linear),
            x if x == Activation::ReLU as u8 => Ok(Activation::ReLU),
            x if x == Activation::Sigmoid as u8 => Ok(Activation::Sigmoid),
            x if x == Activation::Softmax as u8 => Ok(Activation::Softmax),
            x if x == Activation::Tanh as u8 => Ok(Activation::Tanh),
            _ => Err(()),
        }
    }
}

impl Activation {
    pub fn eval(self, z: &Tensor) -> Tensor {
        match self {
            Activation::Sigmoid => sigmoid(z),
            Activation::Softmax => {
                // Input value is shifted for numerical stability
                let z_shifted = sub(z, &max(z, 0), true);
                div(&exp(&z_shifted), &sum(&exp(&z_shifted), 0), true)
            },
            Activation::Tanh => tanh(z),
            Activation::ReLU => {
                maxof(&Tensor::zeros(z.dims()), z, true)
            },
            Activation::LeakyReLU => {
                maxof(&Tensor::zeros(z.dims()), &mul(&constant(0.01 as PrimitiveType, z.dims()), z, true), true)
            },
            Activation::Linear => { z.copy() }
        }
    }

    pub(crate) fn grad(self, z: &Tensor) -> Tensor {
        match self {
            Activation::Sigmoid => sigmoid(z) * (Tensor::ones(z.dims()) - sigmoid(z)),
            Activation::Softmax => Tensor::ones(z.dims()),
            Activation::Tanh => Tensor::ones(z.dims()) - mul(&tanh(z), &tanh(z), true),
            Activation::ReLU => {
                let cond = ge(z, &(0 as PrimitiveType), true);
                cond.cast()
            },
            Activation::LeakyReLU => {
                let cond = ge(z, &(0 as PrimitiveType), true);
                //cond.cast()
                selectr(&Tensor::ones(z.dims()), &cond, 0.01)
            },
            Activation::Linear => Tensor::ones(z.dims()),
        }
    }
}


#[cfg(test)]
mod tests {
    use crate::activations::Activation;
    use crate::assert_approx_eq;
    use arrayfire::*;
    use crate::tensor::*;

    #[test]
    fn sigmoid_eval() {
        let activation = Activation::Sigmoid;
        let values: [PrimitiveType; 9] = [-10.3, -1.2, -0.8, -0.1, 0., 0.15, 1.1, 2.1, 9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.eval(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [0.000034, 0.231475, 0.310026, 0.475021, 0.5, 0.537430, 0.750260, 0.890903, 0.999945];
        assert_approx_eq!(output, expected_output);
    }


    #[test]
    fn sigmoid_grad() {
        let activation = Activation::Sigmoid;
        let values: [PrimitiveType; 9] = [-10.3, -1.2, -0.8, -0.1, 0., 0.15, 1.1, 2.1, 9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.grad(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [0.000034, 0.177894, 0.213910, 0.249376, 0.25, 0.248599, 0.187370, 0.097195, 0.000055];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn tanh_eval() {
        let activation = Activation::Tanh;
        let values: [PrimitiveType; 9] = [-10.3, -1.2, -0.8, -0.1, 0., 0.15, 1.1, 2.1, 9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.eval(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [-1.0, -0.833655, -0.664037, -0.099668, 0.0, 0.148885, 0.800499, 0.970452, 1.0];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn tanh_grad() {
        let activation = Activation::Tanh;
        let values: [PrimitiveType; 9] = [-10.3, -1.2, -0.8, -0.1, 0., 0.15, 1.1, 2.1, 9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.grad(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [0.000000, 0.305020, 0.559055, 0.990066, 1.0, 0.977833, 0.359201, 0.058223, 0.000000];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn linear_eval() {
        let activation = Activation::Linear;
        let values: [PrimitiveType; 9] = [-10.3, -1.2, -0.8, -0.1, 0., 0.15, 1.1, 2.1, 9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.eval(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        assert_approx_eq!(output, values);
    }

    #[test]
    fn linear_grad() {
        let activation = Activation::Linear;
        let values: [PrimitiveType; 9] = [-10.3, -1.2, -0.8, -0.1, 0., 0.15, 1.1, 2.1, 9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.grad(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [1.0; 9];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn relu_eval() {
        let activation = Activation::ReLU;
        let values: [PrimitiveType; 9] = [10.3, -1.2, 0.8, 0.1, 0., -0.15, 1.1, -2.1, -9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.eval(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [10.3, 0.0, 0.8, 0.1, 0.0, 0.0, 1.1, 0.0, 0.0];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn relu_grad() {
        let activation = Activation::ReLU;
        let values: [PrimitiveType; 9] = [10.3, -1.2, 0.8, 0.1, 0., -0.15, 1.1, -2.1, -9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.grad(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [1.0, 0.0, 1.0, 1.0, 1.0, 0.0, 1.0, 0.0, 0.0];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn leaky_relu_eval() {
        let activation = Activation::LeakyReLU;
        let values: [PrimitiveType; 9] = [10.3, -1.2, 0.8, 0.1, 0., -0.15, 1.1, -2.1, -9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.eval(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [0.103, 0.0, 0.008, 0.001, 0.0, 0.0, 0.011, 0.0, 0.0];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn leaky_relu_grad() {
        let activation = Activation::LeakyReLU;
        let values: [PrimitiveType; 9] = [10.3, -1.2, 0.8, 0.1, 0., -0.15, 1.1, -2.1, -9.8];
        let z = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.grad(&z);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [1.0, 0.01, 1.0, 1.0, 1.0, 0.01, 1.0, 0.01, 0.01];
        assert_approx_eq!(output, expected_output);
    }

    #[test]
    fn softmax_eval() {
        let activation = Activation::Softmax;
        let values: [PrimitiveType; 9] = [10.3, -1.2, 0.8, 0.1, 0., -0.15, 1.1, -2.1, -9.8];
        let input = Array::new(&values, Dim4::new(&[3, 3, 1, 1]));
        let eval = activation.eval(&input);
        let mut output: [PrimitiveType; 9] = [0.; 9];
        eval.host(&mut output);
        let expected_output: [PrimitiveType; 9] = [0.999915025297827, 0.000010129232797, 0.000074845469376, 0.372628471150606, 0.337168183722601, 0.290203345126792, 0.960817236817529, 0.039165028193086, 0.000017734989384];
        assert_approx_eq!(output, expected_output);
    }
}