burn-onnx 0.21.0-pre.3

Library for importing ONNX models into the Burn framework
Documentation 
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

use super::broadcast_helpers::align_rhs_for_lhs_rank;
use super::prelude::*;

impl NodeCodegen for onnx_ir::quantize_linear::QuantizeLinearNode {
    fn inputs(&self) -> &[Argument] {
        &self.inputs
    }

    fn outputs(&self) -> &[Argument] {
        &self.outputs
    }

    fn forward(&self, scope: &mut ScopeAtPosition<'_>) -> TokenStream {
        let x_arg = self.inputs.first().unwrap();
        let scale_arg = self.inputs.get(1).unwrap();
        let output = arg_to_ident(self.outputs.first().unwrap());

        let x = scope.arg(x_arg);
        let scale = scope.arg(scale_arg);

        let target_dtype = self.outputs.first().unwrap().ty.elem_type();
        let target_dtype_tokens = target_dtype.to_tokens();

        let x_expr = if x_arg.ty.elem_type().is_float() {
            quote! { (#x).cast(burn::tensor::DType::F32) }
        } else if x_arg.ty.elem_type().is_uint() {
            quote! { (#x).cast(burn::tensor::DType::I32).float().cast(burn::tensor::DType::F32) }
        } else {
            quote! { (#x).float().cast(burn::tensor::DType::F32) }
        };

        let scale_expr = if scale_arg.ty.elem_type().is_float() {
            quote! { (#scale).cast(burn::tensor::DType::F32) }
        } else {
            quote! { (#scale).float().cast(burn::tensor::DType::F32) }
        };

        let x_rank = x_arg.ty.rank();
        let scale_rank = scale_arg.ty.rank();
        let scale_expr = align_rhs_for_lhs_rank(scale_expr, x_rank, scale_rank, self.config.axis);

        let with_zero_point = if let Some(zp_arg) = self.inputs.get(2) {
            let zp = scope.arg(zp_arg);
            let zp_expr = if zp_arg.ty.elem_type().is_float() {
                quote! { (#zp).cast(burn::tensor::DType::F32) }
            } else if zp_arg.ty.elem_type().is_uint() {
                quote! { (#zp).cast(burn::tensor::DType::I32).float().cast(burn::tensor::DType::F32) }
            } else {
                quote! { (#zp).float().cast(burn::tensor::DType::F32) }
            };
            let zp_expr =
                align_rhs_for_lhs_rank(zp_expr, x_rank, zp_arg.ty.rank(), self.config.axis);
            quote! { ((#x_expr).div(#scale_expr)).round().add(#zp_expr) }
        } else {
            quote! { ((#x_expr).div(#scale_expr)).round() }
        };

        let clamped_expr = match target_dtype {
            DType::U8 => quote! { (#with_zero_point).clamp(0f32, 255f32) },
            DType::I8 => quote! { (#with_zero_point).clamp(-128f32, 127f32) },
            DType::U16 => quote! { (#with_zero_point).clamp(0f32, 65535f32) },
            DType::I16 => quote! { (#with_zero_point).clamp(-32768f32, 32767f32) },
            _ => panic!(
                "QuantizeLinear output dtype {:?} not supported",
                target_dtype
            ),
        };

        quote! {
            let #output = (#clamped_expr).int().cast(#target_dtype_tokens);
        }
    }
}

#[cfg(test)]
mod tests {
    use super::super::test_helpers::*;
    use burn::tensor::DType;
    use insta::assert_snapshot;
    use onnx_ir::quantize_linear::{
        QuantizeLinearConfig, QuantizeLinearNode, QuantizeLinearNodeBuilder,
    };

    fn create_node(name: &str, with_zero_point: bool) -> QuantizeLinearNode {
        let mut builder = QuantizeLinearNodeBuilder::new(name)
            .input_tensor("x", 2, DType::F32)
            .input_tensor("y_scale", 0, DType::F32)
            .output_tensor("y", 2, DType::U8)
            .config(QuantizeLinearConfig::default());

        if with_zero_point {
            builder = builder.input_tensor("y_zero_point", 0, DType::U8);
        }

        builder.build()
    }

    #[test]
    fn test_quantize_linear_forward_without_zero_point() {
        let node = create_node("q", false);
        let code = codegen_forward_default(&node);
        assert_snapshot!(code, @r"
        pub fn forward(&self, x: Tensor<B, 2>, y_scale: Tensor<B, 0>) -> Tensor<B, 2, Int> {
            let y = (((((x).cast(burn::tensor::DType::F32))
                .div(
                    ((y_scale).cast(burn::tensor::DType::F32)).unsqueeze_dims(&[0isize, 1isize]),
                ))
                .round())
                .clamp(0f32, 255f32))
                .int()
                .cast(burn::tensor::DType::U8);
            y
        }
        ");
    }

    #[test]
    fn test_quantize_linear_forward_with_zero_point() {
        let node = create_node("q", true);
        let code = codegen_forward_default(&node);
        assert_snapshot!(code, @r"
        pub fn forward(
            &self,
            x: Tensor<B, 2>,
            y_scale: Tensor<B, 0>,
            y_zero_point: Tensor<B, 0, Int>,
        ) -> Tensor<B, 2, Int> {
            let y = (((((x).cast(burn::tensor::DType::F32))
                .div(
                    ((y_scale).cast(burn::tensor::DType::F32)).unsqueeze_dims(&[0isize, 1isize]),
                ))
                .round()
                .add(
                    ((y_zero_point)
                        .cast(burn::tensor::DType::I32)
                        .float()
                        .cast(burn::tensor::DType::F32))
                        .unsqueeze_dims(&[0isize, 1isize]),
                ))
                .clamp(0f32, 255f32))
                .int()
                .cast(burn::tensor::DType::U8);
            y
        }
        ");
    }
}