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

extern crate ndarray;
extern crate ndarray_linalg;

use ndarray::*;

use crate::schmear::*;
use crate::linalg_utils::*;
use crate::func_scatter_tensor::*;

///Represents a probability distribution over linear mappings
///whose covariance structure is separable between input and output.
///This can be conceptualized as a specialization of [`Schmear`]
///for a distribution over matrices where the covariance of the (vectorized)
///random variable takes the separable form `kron(out_covariance, in_covariance)`
///for `out_covariance` and `in_covariance` output coordinate and input coordinate
///covariances, respectively, and `kron` referring to [`crate::linalg_utils::kron`] 
///See also [`FuncScatterTensor`].
pub struct FuncSchmear {
    pub mean : Array2<f32>,
    pub covariance : FuncScatterTensor
}

impl FuncSchmear {
    ///Given a transformation matrix from the full, flattened dimension
    ///of this [`FuncSchmear`] to a smaller dimension, performs a fused
    ///[`FuncSchmear::flatten`] and [`Schmear::transform`] operation using
    ///the specified transformation. This fused operation is written
    ///to be much faster than manually performing the aforementioned operations.
    pub fn compress(&self, mat : ArrayView2<f32>) -> Schmear {
        let t = self.mean.shape()[0];
        let s = self.mean.shape()[1];

        let mean_flat = self.mean.clone().into_shape((t * s,)).unwrap();
        let mean_transformed = mat.dot(&mean_flat);

        let covariance_transformed  = self.covariance.compress(mat);

        Schmear {
            mean : mean_transformed,
            covariance : covariance_transformed
        }
    }
    
    ///Converts this [`FuncSchmear`] over linear maps to its corresponding [`Schmear`]
    ///over vectorized linear mappings.
    pub fn flatten(&self) -> Schmear {
        let t = self.mean.shape()[0];
        let s = self.mean.shape()[1];

        let mean = self.mean.clone().into_shape((t * s,)).unwrap();
        let covariance = self.covariance.flatten();
        Schmear {
            mean,
            covariance
        }
    }
    ///Computes the output [`Schmear`] of this [`FuncSchmear`] applied
    ///to a given argument [`Schmear`].
    ///Given an input schmear, computes the output schmear which
    ///would result from sampling `(function, input)` pairs,
    ///computing `function(input)` for each of them, and then
    ///obtaining the [`Schmear`] over those results.
    pub fn apply(&self, x : &Schmear) -> Schmear {
        let sigma_dot_u = frob_inner(self.covariance.in_scatter.view(), x.covariance.view());
        let u_inner_product = x.mean.dot(&self.covariance.in_scatter).dot(&x.mean);
        let v_scale = sigma_dot_u + u_inner_product;
        let v_contrib = v_scale * &self.covariance.out_scatter;

        if (v_scale < 0.0f32) {
            println!("v scale became negative: {}", v_scale);
            println!("components: {}, {}",  sigma_dot_u, u_inner_product);
            if (u_inner_product < 0.0f32) {
                println!("Non-psd in scatter: {}", &self.covariance.in_scatter);
                println!("x mean: {}", &x.mean);
            }
        }

        let m_sigma_m_t = self.mean.dot(&x.covariance).dot(&self.mean.t());

        let result_covariance = v_contrib + &m_sigma_m_t;
        let result_mean = self.mean.dot(&x.mean);
        let result = Schmear {
            mean : result_mean,
            covariance : result_covariance
        };
        result
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::*;
    use ::rand_distr::StandardNormal;
    use ndarray_linalg::*;
    use crate::test_utils::*;
    use ndarray_rand::*;

    #[test]
    fn schmear_application_accurate() {
        let t = 3;
        let s = 3;
        let num_samps = 1000;
        let arg_scale_mult = 0.01f32;

        let normal_inverse_wishart = random_normal_inverse_wishart(s, t);

        let func_schmear = normal_inverse_wishart.get_schmear();

        let arg_mean = random_vector(s);
        let mut arg_covariance_sqrt = random_matrix(s, s);
        arg_covariance_sqrt *= arg_scale_mult; 
        let arg_covariance = arg_covariance_sqrt.dot(&arg_covariance_sqrt.t());
        let arg_schmear = Schmear {
            mean : arg_mean.clone(),
            covariance : arg_covariance.clone()
        };

        let actual_out_schmear = func_schmear.apply(&arg_schmear);

        let mut expected_out_mean = Array::zeros((t,));
        let mut expected_out_covariance = Array::zeros((t, t));

        let mut rng = rand::thread_rng();

        let scale_fac = 1.0f32 / (num_samps as f32);

        for _ in 0..num_samps {
            let func_samp = normal_inverse_wishart.sample(&mut rng);

            let standard_normal_arg_vec = Array::random((s,), StandardNormal);
            let arg_samp = &arg_mean + &arg_covariance_sqrt.dot(&standard_normal_arg_vec);

            let out_samp = func_samp.dot(&arg_samp);

            let out_diff = &out_samp - &actual_out_schmear.mean;

            expected_out_mean += &(scale_fac * &out_samp);
            expected_out_covariance += &(scale_fac * &outer(out_diff.view(), out_diff.view()));
        }

        assert_equal_vectors_to_within(actual_out_schmear.mean.view(), expected_out_mean.view(), 1.0f32);
        assert_equal_matrices_to_within(actual_out_schmear.covariance.view(), expected_out_covariance.view(), 10.0f32);
    }
}