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

use std::marker::PhantomData;

use crate::{
    checkpoint::{
        base::Checkpointer, retro_forward::RetroForward, state::BackwardStates,
        strategy::CheckpointStrategy,
    },
    grads::Gradients,
    graph::NodeID,
    ops::{unary, Backward, Ops, OpsKind},
    retro_unary, Autodiff,
};
use burn_tensor::{
    backend::Backend,
    ops::{ActivationOps, FloatTensor},
};

impl<B: Backend, C: CheckpointStrategy> ActivationOps<Autodiff<B, C>> for Autodiff<B, C> {
    fn gelu<const D: usize>(tensor: FloatTensor<Self, D>) -> FloatTensor<Self, D> {
        #[derive(Debug)]
        struct Gelu<const D: usize>;

        retro_unary!(RetroGelu, B::gelu);

        impl<const D: usize, B: Backend> Backward<B, D, 1> for Gelu<D> {
            type State = NodeID;

            fn backward(
                self,
                ops: Ops<Self::State, 1>,
                grads: &mut Gradients,
                checkpointer: &mut Checkpointer,
            ) {
                let input = checkpointer.retrieve_node_output(ops.state);

                unary::<B, D, D, _>(ops.parents, ops.node, grads, |grad| {
                    B::gelu_backward(input, grad)
                });
            }
        }

        match Gelu::<D>
            .prepare::<C>([tensor.node.clone()])
            .memory_bound()
            .retro_forward(RetroGelu::<B, D>::new(tensor.node.id))
            .parents([&tensor])
            .stateful()
        {
            OpsKind::Tracked(mut prep) => {
                let state = prep.checkpoint(&tensor);
                prep.finish(state, B::gelu(tensor.primitive.clone()))
            }
            OpsKind::UnTracked(prep) => prep.finish(B::gelu(tensor.primitive)),
        }
    }

    fn relu<const D: usize>(tensor: FloatTensor<Self, D>) -> FloatTensor<Self, D> {
        #[derive(Debug)]
        struct Relu;

        retro_unary!(RetroRelu, B::relu);

        impl<B: Backend, const D: usize> Backward<B, D, 1> for Relu {
            type State = NodeID;

            fn backward(
                self,
                ops: Ops<Self::State, 1>,
                grads: &mut Gradients,
                checkpointer: &mut Checkpointer,
            ) {
                let state = checkpointer.retrieve_node_output(ops.state);
                unary::<B, D, D, _>(ops.parents, ops.node, grads, |grad| {
                    B::relu_backward(state, grad)
                });
            }
        }

        match Relu
            .prepare::<C>([tensor.node.clone()])
            .memory_bound()
            .retro_forward(RetroRelu::<B, D>::new(tensor.node.id))
            .parents([&tensor])
            .stateful()
        {
            OpsKind::Tracked(mut prep) => {
                let state = prep.checkpoint(&tensor);
                prep.finish(state, B::relu(tensor.primitive))
            }
            OpsKind::UnTracked(prep) => prep.finish(B::relu(tensor.primitive)),
        }
    }

    fn sigmoid<const D: usize>(tensor: FloatTensor<Self, D>) -> FloatTensor<Self, D> {
        #[derive(Debug)]
        struct Sigmoid;

        retro_unary!(RetroSigmoid, B::sigmoid);

        impl<B: Backend, const D: usize> Backward<B, D, 1> for Sigmoid {
            type State = NodeID;

            fn backward(
                self,
                ops: Ops<Self::State, 1>,
                grads: &mut Gradients,
                checkpointer: &mut Checkpointer,
            ) {
                let input = checkpointer.retrieve_node_output(ops.state);
                let output = B::sigmoid(input);
                unary::<B, D, D, _>(ops.parents, ops.node, grads, |grad| {
                    B::sigmoid_backward(output, grad)
                });
            }
        }

        match Sigmoid
            .prepare::<C>([tensor.node.clone()])
            .memory_bound()
            .retro_forward(RetroSigmoid::<B, D>::new(tensor.node.id))
            .parents([&tensor])
            .stateful()
        {
            OpsKind::Tracked(mut prep) => {
                let state = prep.checkpoint(&tensor);
                prep.finish(state, B::sigmoid(tensor.primitive))
            }
            OpsKind::UnTracked(prep) => prep.finish(B::sigmoid(tensor.primitive)),
        }
    }

    fn log_sigmoid<const D: usize>(tensor: FloatTensor<Self, D>) -> FloatTensor<Self, D> {
        #[derive(Debug)]
        struct LogSigmoid<const D: usize>;

        retro_unary!(RetroLogSigmoid, B::log_sigmoid);

        impl<const D: usize, B: Backend> Backward<B, D, 1> for LogSigmoid<D> {
            type State = NodeID;

            fn backward(
                self,
                ops: Ops<Self::State, 1>,
                grads: &mut Gradients,
                checkpointer: &mut Checkpointer,
            ) {
                let input = checkpointer.retrieve_node_output(ops.state);

                unary::<B, D, D, _>(ops.parents, ops.node, grads, |grad| {
                    B::log_sigmoid_backward(input, grad)
                });
            }
        }

        match LogSigmoid::<D>
            .prepare::<C>([tensor.node.clone()])
            .memory_bound()
            .retro_forward(RetroLogSigmoid::<B, D>::new(tensor.node.id))
            .parents([&tensor])
            .stateful()
        {
            OpsKind::Tracked(mut prep) => {
                let state = prep.checkpoint(&tensor);
                prep.finish(state, B::log_sigmoid(tensor.primitive.clone()))
            }
            OpsKind::UnTracked(prep) => prep.finish(B::log_sigmoid(tensor.primitive)),
        }
    }
}