tract_tensorflow/ops/nn/
dw_conv2d.rs

1use crate::model::ParsingContext;
2use crate::tfpb::tensorflow::NodeDef;
3use tract_hir::internal::*;
4use tract_hir::ops::cnn::*;
5use tract_hir::ops::nn::*;
6
7pub fn depthwise_conv2d(_ctx: &ParsingContext, pb: &NodeDef) -> TractResult<Box<dyn InferenceOp>> {
8    let data_format = super::data_format(pb)?;
9    let padding = super::padding(pb)?;
10    let strides = super::strides(pb)?.into();
11    let dilations: TVec<usize> = pb.get_attr_list_int("dilations")?.into();
12    if dilations.len() != 4 || dilations[0] != 1 && dilations[3] != 1 {
13        bail!("dilations must be of the form [1, h, v, 1], found {:?}", dilations)
14    };
15    Ok(expand(DepthwiseConv2d::new(data_format, padding, strides, dilations)))
16}
17
18#[derive(Debug, Clone, new, Hash)]
19pub struct DepthwiseConv2d {
20    data_format: DataFormat,
21    padding: PaddingSpec,
22    strides: TVec<usize>,
23    dilations: TVec<usize>,
24}
25
26
27
28impl Expansion for DepthwiseConv2d {
29    fn name(&self) -> StaticName {
30        "DepthwiseConv2dNative".into()
31    }
32
33
34    fn rules<'r, 'p: 'r, 's: 'r>(
35        &'s self,
36        s: &mut Solver<'r>,
37        inputs: &'p [TensorProxy],
38        outputs: &'p [TensorProxy],
39    ) -> InferenceResult {
40        check_input_arity(inputs, 2)?;
41        check_output_arity(outputs, 1)?;
42        s.equals(&inputs[0].rank, 4)?;
43        s.equals(&inputs[1].rank, 4)?;
44        s.equals(&inputs[0].datum_type, &inputs[1].datum_type)?;
45        s.equals(&inputs[0].datum_type, &outputs[0].datum_type)?;
46        s.equals(&outputs[0].rank, 4)?;
47        s.given_2(&inputs[0].shape, &inputs[1].shape, move |s, img, ker| {
48            let img = self.data_format.shape(img)?;
49            s.equals(&inputs[1].shape[2], &inputs[0].shape[img.c_axis()])?;
50            s.equals(&outputs[0].shape[img.n_axis().unwrap()], img.n_dim().unwrap())?;
51            if let Ok(ker) = ker.iter().map(|d| d.to_usize()).collect::<TractResult<TVec<_>>>() {
52                let output_shape = self.padding.compute(
53                    img.hw_dims(),
54                    &ker[0..2],
55                    &self.dilations[img.hw_axes()],
56                    &self.strides[img.hw_axes()],
57                );
58                let in_channels = ker[2].to_usize()?;
59                let multiplier = ker[3].to_usize()?;
60                s.equals(&outputs[0].shape[img.h_axis()], &output_shape[0].convoluted)?;
61                s.equals(&outputs[0].shape[img.h_axis() + 1], &output_shape[1].convoluted)?;
62                s.equals(&outputs[0].shape[img.c_axis()], (in_channels * multiplier).to_dim())?;
63            }
64            Ok(())
65        })?;
66        Ok(())
67    }
68
69    fn wire(
70        &self,
71        prefix: &str,
72        model: &mut TypedModel,
73        inputs: &[OutletId],
74    ) -> TractResult<TVec<OutletId>> {
75        let input = model.outlet_fact(inputs[0])?;
76        let kernel = model.outlet_fact(inputs[1])?;
77        let input_shape = input.shape.to_tvec();
78        let kernel_shape = if let Some(s) = kernel.shape.as_concrete() {
79            s
80        } else {
81            bail!("Do not expect streaming on kernel dims");
82        };
83        let shape = self.data_format.shape(&input_shape)?;
84        let mut conv = Conv::default()
85            .hwio()
86            .group(kernel_shape[2])
87            .dilations(self.dilations[shape.hw_axes()].into())
88            .strides(self.strides[shape.hw_axes()].into())
89            .padding(self.padding.clone());
90        if self.data_format == DataFormat::NHWC {
91            conv = conv.nhwc()
92        }
93        conv.wire(prefix, model, inputs)
94    }
95}
tract_tensorflow/ops/nn/dw_conv2d.rs

tract_tensorflow/ops/nn/
dw_conv2d.rs
