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

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

//pub struct Conv1d {
//    alpha: f32,
//}
//impl Conv1d {
//    pub fn new(alpha: f32) -> Conv1d {
//        Conv1d {
//            alpha: alpha,
//        }
//    }
//}
//impl OpTrait for Conv1d {
//    fn get_name(&self) -> String {
//        "Conv1d".to_string()
//    }
//    fn get_input_size(&self) -> usize {
//        2
//    }
//    fn get_output_size(&self) -> usize {
//        1
//    }
//    /// Forward pass
//    fn apply(&mut self, input: &[&Tensor], output: &[&Tensor]) {
//        unimplemented!();
//    }
//    
//    /// 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]) {
//        unimplemented!();
//    }
//
//    /// 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()
//    }
//}

// Conv2d
#[derive(Clone, Copy, PartialEq)]
pub enum PaddingMode{
    Zeros,
    Reflect,
    Replicate,
    Circular,
}

pub struct Conv2d {
    in_channels: usize,
    out_channels: usize,
    kernel_size: (usize, usize),
    stride: (usize, usize),
    padding: (usize, usize),
    dilation: (usize, usize),
    groups: usize,
    bias_option: bool,
    padding_mode: PaddingMode,
    
    weight: Tensor,
    bias: Tensor,
    weight_grad: Tensor,
    bias_grad: Tensor,
}
impl Conv2d {
    pub fn new(in_channels: usize, out_channels: usize,
               kernel_size: (usize, usize),
               stride: (usize, usize),
               padding: (usize, usize),
               dilation: (usize, usize),
               bias: bool,
               padding_mode: PaddingMode
    ) -> Conv2d {
        Conv2d {
            in_channels: in_channels,
            out_channels: out_channels,
            kernel_size: kernel_size,
            stride: stride,
            padding: padding,
            dilation: dilation,
            groups: 1,
            bias_option: bias,
            padding_mode: padding_mode,
            
            weight: Tensor::empty(&vec![out_channels, in_channels, kernel_size.0, kernel_size.1]),
            bias: Tensor::empty(&vec![out_channels, ]),
            weight_grad: Tensor::empty(&vec![out_channels, in_channels, kernel_size.0, kernel_size.1]),
            bias_grad: Tensor::empty(&vec![out_channels, ]),
        }
    }
}
impl OpTrait for Conv2d {
    fn get_name(&self) -> String {
        "Conv2d".to_string()
    }
    fn get_input_size(&self) -> usize {
        1
    }
    fn get_output_size(&self) -> usize {
        1
    }
    /// Forward pass
    fn apply(&mut self, input: &[&Tensor], output: &[&Tensor]) {
        if self.groups > 1 {
            unimplemented!();
        }
        if self.weight.size()[2] != self.kernel_size.0 || self.weight.size()[3] != self.kernel_size.1 {
            panic!("this is conv2d");
        }
        let input_size = input[0].size();
        if input_size[1] != self.in_channels {
            panic!("conv2d expect the same input channel: input: {:?}, config: {:?}", input_size[1], self.in_channels);
        }
        let conv_output = input[0].conv2d(&self.weight, self.stride, self.padding, self.dilation, self.padding_mode);
        //println!("conv_output: {:?}, {:?}, {:?}, {:?}, {:?}, {:?}", self.weight.size(), self.stride, self.padding, self.dilation, conv_output.size(), input[0].size());
        if conv_output.size()[1] != self.out_channels {
            panic!("conv2d expect the same input channel {:?}, {:?}", input_size[1], self.in_channels);
        }

        if self.bias_option {
            //println!("{:?}, {:?}", self.weight.size(), self.bias.size());
            let expanded_bias = self.bias
                .unsqueeze(1)
                .unsqueeze(2)
                .repeat(&[1, conv_output.size()[2], conv_output.size()[3]]);
            //println!("conv_output: {:?}, expanded_bias.size() {:?}", conv_output.size(), expanded_bias.size());
            let ret = conv_output.add(&expanded_bias);
            output[0].swap(ret);
        } else {
            output[0].swap(conv_output);            
        }
    }
    
    /// 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]) {
        let (w_grad, d_grad) = input[0].conv2d_grad(&self.weight, self.stride, self.padding, self.dilation, self.padding_mode, output_grad[0]);
        self.weight_grad.swap(w_grad);
        input_grad[0].swap(d_grad);

        if self.bias_option {
            self.bias_grad.swap(output_grad[0].mean(Some(&[0, 2, 3]), false));
        }
    }

    /// access weight values
    fn get_values(&self) -> Vec<&Tensor> {
        vec![&self.weight, &self.bias]
    }
    fn set_values(&self, v: &[Tensor]) {
        self.weight.swap(v[0].clone());
        self.bias.swap(v[1].clone());
    }
    /// access gradient values
    fn get_grads(&self) -> Vec<&Tensor> {
        vec![&self.weight_grad, &self.bias_grad]
    }
}
// Conv3d
// ConvTranspose1d
// ConvTranspose2d
// ConvTranspose3d
// Unfold
// Fold