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

mod compress;
mod slice;

use tract_hir::internal::*;
use tract_hir::ops::array;

use crate::model::{OnnxOpRegister, ParsingContext};
use crate::pb::*;
use tract_num_traits::AsPrimitive;

pub fn register_all_ops(reg: &mut OnnxOpRegister) {
    reg.insert("Compress", compress::compress);
    reg.insert("Concat", concat);
    reg.insert("ConstantLike", constant_like);
    reg.insert("ConstantOfShape", constant_of_shape);
    reg.insert("Expand", |_, _| Ok((Box::new(array::MultiBroadcastTo::default()), vec![])));
    reg.insert("EyeLike", eye_like);
    reg.insert("Flatten", flatten);
    reg.insert("Gather", gather);
    reg.insert("Pad", pad);
    reg.insert("Reshape", |_, _| Ok((Box::new(array::Reshape::default()), vec![])));
    reg.insert("Shape", |_, _| Ok((Box::new(array::Shape::new(DatumType::I64)), vec![])));
    reg.insert("Size", |_, _| Ok((Box::new(array::Size::new(DatumType::I64)), vec![])));
    reg.insert("Transpose", transpose);
    reg.insert("Tile", |_, _| Ok((Box::new(array::Tile::default()), vec![])));
    reg.insert("Slice", slice::slice);
    reg.insert("Split", split);
    reg.insert("Squeeze", squeeze);
    reg.insert("Unsqueeze", unsqueeze);
}

pub fn concat(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let axis = node.get_attr("axis")?;
    Ok((Box::new(array::Concat::new(axis)), vec![]))
}

pub fn make_const<T>(shape: &[usize], v: f32) -> TractResult<Arc<Tensor>>
where
    T: Copy + Datum,
    f32: AsPrimitive<T>,
{
    Ok(tract_ndarray::Array::<T, _>::from_elem(shape, v.as_()).into_arc_tensor())
}

pub fn constant_like(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let value = node.get_attr_opt("value")?.unwrap_or(0.);
    if node.input.len() == 0 {
        let dt = node.get_attr_opt("dtype")?.unwrap_or(f32::datum_type());
        let shape: Vec<usize> = node.get_attr_vec("shape")?;
        let tensor = dispatch_numbers!(self::make_const(dt)(&shape, value))?;
        Ok((Box::new(tract_hir::ops::konst::Const::new(tensor)), vec![]))
    } else {
        Ok((Box::new(array::ConstantLike::new(value)), vec![]))
    }
}

pub fn constant_of_shape(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let value = match node.get_attr_opt::<Tensor>("value")? {
        Some(val) => val.into_arc_tensor(),
        None => make_const::<f32>(&vec![1], 0.0 as f32)?,
    };
    Ok((Box::new(array::ConstantOfShape::new(value)), vec![]))
}

pub fn eye_like(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let dt = node.get_attr_opt("dtype")?;
    let k = node.get_attr_opt("k")?.unwrap_or(0);
    Ok((Box::new(array::EyeLike::new(dt, k)), vec![]))
}

pub fn flatten(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let axis = node.get_attr_opt("axis")?.unwrap_or(1);
    Ok((Box::new(array::Flatten::new(axis)), vec![]))
}

pub fn gather(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let axis = node.get_attr_opt("axis")?.unwrap_or(0);
    Ok((Box::new(array::Gather::new(axis)), vec![]))
}

pub fn pad(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let value: f32 = node.get_attr_opt("value")?.unwrap_or(0.0);
    let mode = match node.get_attr_opt("mode")? {
        None | Some("constant") => None,
        Some(mode) => node.check_value(
            "mode",
            match mode {
                "reflect" => Ok(Some(array::PadMode::Reflect)),
                "edge" => Ok(Some(array::PadMode::Edge)),
                _ => Err(mode),
            },
        )?,
    }
    .unwrap_or_else(|| array::PadMode::Constant(Arc::new(value.into())));
    let pads = node.get_attr_tvec("pads")?;
    let rank = pads.len() / 2;
    let pads = (0..rank).map(|ax| (pads[ax], pads[ax + rank])).collect();
    Ok((Box::new(array::Pad::new(pads, mode)), vec![]))
}

pub fn split(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let axis = node.get_attr_opt("axis")?.unwrap_or(0);
    let split = node.get_attr_opt_vec("split")?;
    Ok((Box::new(array::Split::new(axis, node.output.len(), split)), vec![]))
}

pub fn squeeze(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let axes = node.get_attr_opt_vec("axes")?;
    Ok((Box::new(array::Squeeze::new(axes)), vec![]))
}

pub fn transpose(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let perm = node.get_attr_opt_vec("perm")?;
    Ok((Box::new(array::PermuteAxes::new(perm)), vec![]))
}

pub fn unsqueeze(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let axes = node.get_attr_vec("axes")?;
    Ok((Box::new(array::AddDims::new(axes)), vec![]))
}