tract_core/ops/downsample/
mod.rs1use crate::internal::*;
2use crate::ops;
3use ndarray::prelude::*;
4
5use super::identity::Identity;
6
7mod array;
8mod conv;
9mod scan;
10
11#[derive(Debug, Clone, new, Default, PartialEq, Eq, Hash)]
12pub struct Downsample {
13 pub axis: usize,
14 pub stride: isize,
15 pub modulo: usize,
16}
17
18impl Downsample {
19 pub(crate) fn transform_dim(&self, input_dim: &TDim) -> TDim {
20 (input_dim.clone() - self.modulo).div_ceil(self.stride.unsigned_abs() as u64)
21 }
22
23 pub(crate) fn transform_fact(&self, input_fact: &TypedFact) -> TractResult<TypedFact> {
24 let mut downed = input_fact.clone();
25 let down_len = self.transform_dim(&input_fact.shape[self.axis]);
26 downed.shape.set(self.axis, down_len);
27 if let Some(k) = downed.konst {
28 let mut outputs = self.eval(tvec!(k.into_tvalue()))?;
29 downed.konst = Some(outputs.remove(0).into_arc_tensor())
30 }
31 if cfg!(debug_assertions) {
32 downed.consistent()?;
33 }
34 Ok(downed)
35 }
36}
37
38impl Op for Downsample {
39 fn name(&self) -> StaticName {
40 "Downsample".into()
41 }
42
43 fn info(&self) -> TractResult<Vec<String>> {
44 Ok(vec![format!("axis:{} stride:{} modulo:{}", self.axis, self.stride, self.modulo)])
45 }
46
47 op_as_typed_op!();
48}
49
50impl EvalOp for Downsample {
51 fn is_stateless(&self) -> bool {
52 true
53 }
54
55 fn eval(&self, inputs: TVec<TValue>) -> TractResult<TVec<TValue>> {
56 let input = args_1!(inputs);
57 unsafe {
58 let t = if self.modulo > input.shape()[self.axis] {
59 let mut shape: TVec<usize> = input.shape().into();
60 shape[self.axis] = 0;
61 Tensor::uninitialized_dt(input.datum_type(), &shape)?
62 } else {
63 let slice = ndarray::Slice::new(self.modulo as isize, None, self.stride);
64 unsafe fn do_slice<T: Datum>(
65 t: &Tensor,
66 axis: usize,
67 slice: ndarray::Slice,
68 ) -> Tensor {
69 unsafe {
70 let dt = t.datum_type();
71 let mut t2 = t
72 .to_array_view_unchecked::<T>()
73 .slice_axis(Axis(axis), slice)
74 .into_owned()
75 .into_tensor();
76 t2.set_datum_type(dt);
77 t2
78 }
79 }
80 dispatch_datum_by_size!(do_slice(input.datum_type())(&*input, self.axis, slice))
81 };
82 Ok(tvec!(t.into_tvalue()))
83 }
84 }
85}
86
87impl TypedOp for Downsample {
88 fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
89 ensure!(self.axis < inputs[0].rank());
90 ensure!(
91 self.modulo == 0 || self.stride > 0,
92 "non-zero modulo is only defined with forward strides"
93 );
94 let mut downed = inputs[0].without_value();
95 let down_len = self.transform_dim(&downed.shape[self.axis]);
96 downed.shape.set(self.axis, down_len);
97 Ok(tvec!(downed))
98 }
99
100 fn declutter(
101 &self,
102 model: &TypedModel,
103 node: &TypedNode,
104 ) -> TractResult<Option<TypedModelPatch>> {
105 if self.stride == 1 {
106 return Ok(Some(TypedModelPatch::replace_single_op(
107 model,
108 node,
109 &node.inputs,
110 Identity,
111 )?));
112 }
113 pull_downsample_up(model, node)
114 .with_context(|| format!("Pulling {} over {}", node, model.node(node.inputs[0].node)))
115 }
116
117 as_op!();
118}
119
120fn pull_downsample_up(
121 model: &TypedModel,
122 down_node: &TypedNode,
123) -> TractResult<Option<TypedModelPatch>> {
124 model.check_consistency()?;
125 let down_op = down_node.op_as::<Downsample>().unwrap();
126 if let Some(prec) = model.linear_prec(down_node.id)? {
127 let (input_facts, output_facts) = model.node_facts(prec.id)?;
128 let axes_mapping = prec.op.axes_mapping(&input_facts, &output_facts)?;
129 debug!("Consider pull {down_op:?} over {prec:?} (invariants: {axes_mapping:?})");
130 if let Some(slice_op) = prec.op_as::<ops::array::Slice>() {
131 if let Some(p) =
132 array::pull_downsample_over_slice(model, prec, slice_op, down_node, down_op)?
133 {
134 return Ok(Some(p));
135 }
136 } else if let Some(other_op) = prec.op_as::<AxisOp>() {
137 return array::pull_downsample_over_axis_op(model, prec, other_op, down_node, down_op);
138 } else if let Some(conv_op) = prec.op_as::<ops::cnn::conv::Conv>() {
139 return conv::fuse_downsample_into_conv(model, prec, conv_op, down_node, down_op);
140 } else if let Some(other_op) = prec.op_as::<ops::scan::Scan>() {
141 return scan::pull_downsample_over_scan(model, prec, other_op, down_node, down_op);
142 }
143 rule_if!(prec.outputs.len() <= 1 && prec.inputs.len() > 0);
144 let axis_info = axes_mapping.axis((InOut::Out(0), down_op.axis))?;
145 let mut patch = TypedModelPatch::default();
146 let mut inputs = vec![];
147 for (ix, (outlet, axis_info)) in prec.inputs.iter().zip(&axis_info.inputs).enumerate() {
148 let mut wire = patch.tap_model(model, *outlet)?;
149 if let &[axis] = &**axis_info {
150 if !patch.outlet_fact(wire)?.shape[axis].is_one() {
151 let mut op = down_op.clone();
152 op.axis = axis;
153 wire = patch.wire_node(
154 format!("{}.{}-{}", down_node.name, prec.name, ix),
155 op,
156 &[wire],
157 )?[0];
158 }
159 } else {
160 return Ok(None);
161 }
162 inputs.push(wire);
163 }
164 let other = patch.wire_node(&prec.name, prec.op.clone(), &inputs)?;
165 patch.shunt_outside(model, OutletId::new(down_node.id, 0), other[0])?;
166 return Ok(Some(patch));
167 }
168 Ok(None)
169}