p2o 0.1.1

A PaddlePaddle New IR (PIR) to ONNX model converter.
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
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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227

use crate::helper::{self, dt};
use crate::proto::onnx;
use serde_json::Value;

impl super::super::Converter {
    fn add_empty_tensor_initializer(&mut self, name: String, data_type: i32) -> String {
        self.onnx_graph.initializer.push(onnx::TensorProto {
            name: name.clone(),
            dims: vec![0],
            data_type,
            ..Default::default()
        });
        name
    }

    fn push_resize_optional_input(
        &mut self,
        node: &mut onnx::NodeProto,
        out_id: i64,
        slot: &str,
        data_type: i32,
    ) {
        // ONNX Resize uses an empty-string sentinel for omitted optional inputs from opset 13
        // onward. Older opsets require a typed empty initializer in the same input slot.
        if self.target_opset >= 13 {
            node.input.push(String::new());
        } else {
            node.input.push(self.add_empty_tensor_initializer(
                format!("resize_empty_{}_{}", slot, out_id),
                data_type,
            ));
        }
    }

    fn build_resize_sizes_from_spatial_tensor(
        &mut self,
        out_id: i64,
        data_input_id: i64,
        spatial_name: String,
    ) -> anyhow::Result<String> {
        let input_name = self.get_tensor_name(data_input_id)?;
        let shape_name = format!("resize_shape_{}", out_id);
        self.onnx_graph.node.push(onnx::NodeProto {
            op_type: "Shape".to_string(),
            input: vec![input_name],
            output: vec![shape_name.clone()],
            ..Default::default()
        });

        let prefix_indices_name = format!("resize_prefix_indices_{}", out_id);
        self.push_i64_initializer(prefix_indices_name.clone(), vec![2], &[0, 1]);

        let prefix_name = format!("resize_prefix_{}", out_id);
        self.onnx_graph.node.push(onnx::NodeProto {
            op_type: "Gather".to_string(),
            input: vec![shape_name, prefix_indices_name],
            output: vec![prefix_name.clone()],
            ..Default::default()
        });

        let sizes_name = format!("resize_sizes_{}", out_id);
        let mut concat = onnx::NodeProto {
            op_type: "Concat".to_string(),
            input: vec![prefix_name, spatial_name],
            output: vec![sizes_name.clone()],
            ..Default::default()
        };
        concat.attribute.push(helper::attr_int("axis", 0));
        self.onnx_graph.node.push(concat);
        Ok(sizes_name)
    }

    fn build_resize_sizes_from_spatial_list(
        &mut self,
        out_id: i64,
        data_input_id: i64,
        size_input_id: i64,
    ) -> anyhow::Result<Option<String>> {
        let spatial_ids = match self.state.combines.get(&size_input_id) {
            Some(ids) if !ids.is_empty() => ids.clone(),
            _ => return Ok(None),
        };

        let mut spatial_inputs = Vec::with_capacity(spatial_ids.len());
        for (index, spatial_id) in spatial_ids.iter().enumerate() {
            let spatial_tensor_name = self.get_tensor_name(*spatial_id)?;
            let spatial_shape = self
                .state
                .tensor_shapes
                .get(spatial_id)
                .cloned()
                .unwrap_or_default();
            if spatial_shape.is_empty() {
                let expanded_name = format!("resize_spatial_{}_{}", out_id, index);
                self.add_unsqueeze_node(
                    spatial_tensor_name,
                    expanded_name.clone(),
                    &[0],
                    format!("resize_unsqueeze_axes_{}", out_id),
                );
                spatial_inputs.push(expanded_name);
            } else {
                spatial_inputs.push(spatial_tensor_name);
            }
        }

        let spatial_name = if spatial_inputs.len() == 1 {
            spatial_inputs.remove(0)
        } else {
            let spatial_name = format!("resize_hw_{}", out_id);
            let mut concat = onnx::NodeProto {
                op_type: "Concat".to_string(),
                input: spatial_inputs,
                output: vec![spatial_name.clone()],
                ..Default::default()
            };
            concat.attribute.push(helper::attr_int("axis", 0));
            self.onnx_graph.node.push(concat);
            spatial_name
        };

        self.build_resize_sizes_from_spatial_tensor(out_id, data_input_id, spatial_name)
            .map(Some)
    }

    fn op_resize_interp(&mut self, op: &Value, mode: &str) -> anyhow::Result<()> {
        self.require_opset(11, mode)?;
        let data_format = helper::attr(op, "data_format")
            .and_then(|d| d.as_str())
            .unwrap_or("NCHW");
        if data_format != "NCHW" {
            anyhow::bail!(
                "{} currently only supports data_format=NCHW (got {})",
                mode,
                data_format
            );
        }
        let align_corners = helper::attr(op, "align_corners")
            .and_then(|d| d.as_bool())
            .unwrap_or(false);
        let align_mode = helper::attr(op, "align_mode")
            .and_then(|d| d.as_i64())
            .unwrap_or(1);
        // Paddle nearest_interp matches ORT Resize with asymmetric coordinates and floor
        // rounding for the models we convert here, including SLANet's odd-sized CSPPAN
        // upsampling path (e.g. 16 -> 31). The half_pixel/round_prefer_ceil mapping
        // shifts samples and causes large feature drift.
        let coordinate_mode = if align_corners {
            "align_corners"
        } else if mode == "nearest" {
            "asymmetric"
        } else if align_mode == 0 {
            "half_pixel"
        } else {
            "asymmetric"
        };
        let nearest_mode = "floor";

        let mut onnx_node = onnx::NodeProto {
            op_type: "Resize".to_string(),
            ..Default::default()
        };

        onnx_node.attribute.push(helper::attr_str("mode", mode));
        onnx_node.attribute.push(helper::attr_str(
            "coordinate_transformation_mode",
            coordinate_mode,
        ));
        if mode == "nearest" {
            onnx_node
                .attribute
                .push(helper::attr_str("nearest_mode", nearest_mode));
        }

        let out_id = helper::op_out_id(op)?;
        let inputs = helper::op_input_ids(op);
        if !inputs.is_empty() {
            onnx_node.input.push(self.get_tensor_name(inputs[0])?);
        }
        self.push_resize_optional_input(&mut onnx_node, out_id, "roi", dt::FLOAT);
        if inputs.len() > 2
            && inputs[2] != 0
            && let Some(sizes_name) =
                self.build_resize_sizes_from_spatial_list(out_id, inputs[0], inputs[2])?
        {
            self.push_resize_optional_input(&mut onnx_node, out_id, "scales", dt::FLOAT);
            onnx_node.input.push(sizes_name);
        } else if let Some(output_shape) = self.state.tensor_shapes.get(&out_id).cloned()
            && output_shape.len() == 4
            && output_shape[2] > 0
            && output_shape[3] > 0
        {
            self.push_resize_optional_input(&mut onnx_node, out_id, "scales", dt::FLOAT);

            let hw_name = format!("resize_static_hw_{}", out_id);
            self.push_i64_initializer(hw_name.clone(), vec![2], &output_shape[2..4]);

            let sizes_name =
                self.build_resize_sizes_from_spatial_tensor(out_id, inputs[0], hw_name)?;
            onnx_node.input.push(sizes_name);
        } else {
            let scale_name = format!("scales_{}", out_id);
            onnx_node.input.push(scale_name.clone());
            let mut scale_vals = vec![1.0f32, 1.0, 2.0, 2.0];
            if let Some(d_arr) = helper::attr(op, "scale").and_then(|d| d.as_array())
                && d_arr.len() >= 2
            {
                scale_vals[2] = d_arr[0].get("D").and_then(|v| v.as_f64()).unwrap_or(2.0) as f32;
                scale_vals[3] = d_arr[1].get("D").and_then(|v| v.as_f64()).unwrap_or(2.0) as f32;
            }

            self.push_f32_initializer(scale_name, vec![4], &scale_vals);
        }

        onnx_node.output.push(self.get_tensor_name(out_id)?);
        self.onnx_graph.node.push(onnx_node);
        Ok(())
    }

    pub fn op_nearest_interp(&mut self, op: &Value) -> anyhow::Result<()> {
        self.op_resize_interp(op, "nearest")
    }

    pub fn op_bilinear_interp(&mut self, op: &Value) -> anyhow::Result<()> {
        self.op_resize_interp(op, "linear")
    }
}