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

use burn_store::TensorSnapshot;

use super::prelude::*;

impl NodeCodegen for onnx_ir::node::layer_norm::LayerNormalizationNode {
    fn inputs(&self) -> &[Argument] {
        &self.inputs
    }

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

    fn field(&self) -> Option<Field> {
        let name = Ident::new(&self.name, Span::call_site());
        let scale_shape = self.inputs[1]
            .ty
            .static_shape_known()
            .expect("LayerNorm: scale tensor shape must be known at codegen time");
        let num_features = scale_shape[0].to_tokens();
        let epsilon = self.config.epsilon;
        let has_bias = self.inputs.len() > 2;

        Some(Field::new(
            self.name.clone(),
            quote! {
                LayerNorm<B>
            },
            quote! {
                let #name = LayerNormConfig::new(#num_features)
                    .with_epsilon(#epsilon)
                    .with_bias(#has_bias)
                    .init(device);
            },
        ))
    }

    fn collect_snapshots(&self, field_name: &str) -> Vec<TensorSnapshot> {
        use crate::burn::node_traits::create_lazy_snapshot;

        let mut snapshots = vec![];

        // Gamma (scale) tensor at input index 1
        if let Some(gamma_input) = self.inputs.get(1) {
            let gamma_path = format!("{}.gamma", field_name);
            if let Some(snapshot) = create_lazy_snapshot(gamma_input, &gamma_path, "LayerNorm") {
                snapshots.push(snapshot);
            }
        }

        // Beta (bias) tensor at input index 2 - only if present
        if self.inputs.len() > 2
            && let Some(beta_input) = self.inputs.get(2)
        {
            let beta_path = format!("{}.beta", field_name);
            if let Some(snapshot) = create_lazy_snapshot(beta_input, &beta_path, "LayerNorm") {
                snapshots.push(snapshot);
            }
        }
        // When bias is absent, Burn's LayerNorm is configured with .with_bias(false),
        // so no beta parameter exists and no snapshot is needed

        snapshots
    }

    fn forward(&self, scope: &mut ScopeAtPosition<'_>) -> TokenStream {
        let input = scope.arg(self.inputs.first().unwrap());
        let output = arg_to_ident(self.outputs.first().unwrap());
        let field = Ident::new(&self.name, Span::call_site());

        if self.config.full_precision {
            quote! {
                let #output = {
                    let dtype = #input.dtype();
                    self.#field.forward(#input.cast(burn::tensor::DType::F32)).cast(dtype)
                };
            }
        } else {
            quote! {
                let #output = self.#field.forward(#input);
            }
        }
    }

    fn register_imports(&self, imports: &mut BurnImports) {
        imports.register("burn::nn::LayerNorm");
        imports.register("burn::nn::LayerNormConfig");
    }
}

#[cfg(test)]
mod tests {
    use super::super::test_helpers::*;
    use burn::tensor::DType;
    use insta::assert_snapshot;
    use onnx_ir::node::layer_norm::{
        LayerNormConfig, LayerNormalizationNode, LayerNormalizationNodeBuilder,
    };

    fn create_layer_norm_node(name: &str) -> LayerNormalizationNode {
        let config = LayerNormConfig::new(1e-5, true);

        LayerNormalizationNodeBuilder::new(name)
            .input_tensor("input", 3, DType::F32)
            .input_static_tensor_shape("scale", vec![512], DType::F32)
            .input_static_tensor_shape("bias", vec![512], DType::F32)
            .output_tensor("output", 3, DType::F32)
            .config(config)
            .build()
    }

    #[test]
    fn test_layer_norm_forward() {
        let node = create_layer_norm_node("layer_norm1");
        let code = codegen_forward_default(&node);
        assert_snapshot!(code, @r"
        pub fn forward(&self, input: Tensor<B, 3>) -> Tensor<B, 3> {
            let output = {
                let dtype = input.dtype();
                self.layer_norm1.forward(input.cast(burn::tensor::DType::F32)).cast(dtype)
            };
            output
        }
        ");
    }

    #[test]
    fn test_layer_norm_forward_with_clone() {
        let node = create_layer_norm_node("layer_norm1");
        let code = codegen_forward_with_clone(&node);
        assert_snapshot!(code, @r"
        pub fn forward(&self, input: Tensor<B, 3>) -> Tensor<B, 3> {
            let output = {
                let dtype = input.clone().dtype();
                self.layer_norm1
                    .forward(input.clone().cast(burn::tensor::DType::F32))
                    .cast(dtype)
            };
            output
        }
        ");
    }
}