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

use crate::internal::*;
use crate::ops;
use ndarray::prelude::*;

mod array;
mod conv;
mod scan;

#[derive(Debug, Clone, new, Default, PartialEq, Hash)]
pub struct Downsample {
    pub axis: usize,
    pub stride: usize,
    pub modulo: usize,
}

impl Downsample {
    fn eval_t<T: Datum>(&self, input: &Tensor) -> TractResult<Arc<Tensor>> {
        let input = input.to_array_view::<T>()?;
        let sampled = if self.modulo < input.shape()[self.axis] {
            input
                .slice_axis(
                    Axis(self.axis),
                    ndarray::Slice::new(self.modulo as isize, None, self.stride as isize),
                )
                .to_owned()
                .into_arc_tensor()
        } else {
            let mut shape = input.shape().to_vec();
            shape[self.axis] = 0;
            unsafe { Tensor::uninitialized::<T>(&shape)?.into_arc_tensor() }
        };
        Ok(sampled)
    }

    pub(crate) fn transform_dim(&self, input_dim: &TDim) -> TDim {
        (input_dim.clone() - self.modulo).div_ceil(self.stride as u32)
    }

    pub(crate) fn transform_fact(&self, input_fact: &TypedFact) -> TractResult<TypedFact> {
        let mut downed = input_fact.clone();
        let down_len = self.transform_dim(&input_fact.shape.dim(self.axis));
        downed.shape.set_dim(self.axis, down_len.clone())?;
        Ok(downed)
    }
}

tract_linalg::impl_dyn_hash!(Downsample);

impl Op for Downsample {
    fn name(&self) -> Cow<str> {
        "Downsample".into()
    }

    fn info(&self) -> TractResult<Vec<String>> {
        Ok(vec![format!("axis:{} stride:{} modulo:{}", self.axis, self.stride, self.modulo)])
    }

    op_core_mir!();
    impl_op_same_as!();
    op_as_typed_op!();
    op_as_pulsed_op!();
}

impl StatelessOp for Downsample {
    fn eval(&self, mut inputs: TVec<Arc<Tensor>>) -> TractResult<TVec<Arc<Tensor>>> {
        let input = args_1!(inputs);
        Ok(tvec!(dispatch_datum!(Self::eval_t(input.datum_type())(self, &*input))?))
    }
}

impl TypedOp for Downsample {
    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
        let mut downed = inputs[0].clone();
        let down_len = self.transform_dim(&downed.shape.dim(self.axis));
        downed.shape.set_dim(self.axis, down_len.clone())?;
        Ok(tvec!(downed))
    }

    fn declutter(
        &self,
        model: &TypedModel,
        node: &TypedNode,
    ) -> TractResult<Option<TypedModelPatch>> {
        if self.stride == 1 {
            return Ok(Some(TypedModelPatch::shunt_one_op(model, node)?));
        }
        pull_downsample_up(model, node)
    }

    fn pulsify(
        &self,
        _source: &NormalizedModel,
        node: &NormalizedNode,
        target: &mut PulsedModel,
        mapping: &HashMap<OutletId, OutletId>,
        _pulse: usize,
    ) -> TractResult<TVec<OutletId>> {
        let input = mapping[&node.inputs[0]];
        let pulse = target.outlet_fact(input)?.pulse();
        if pulse % self.stride != 0 {
            bail!("Pulsificaton requires pulse to be a stride multiple")
        }
        target.wire_node(&*node.name, self.clone(), &[input])
    }

    as_op!();
}

impl PulsedOp for Downsample {
    fn pulsed_output_facts(&self, inputs: &[&PulsedFact]) -> TractResult<TVec<PulsedFact>> {
        let mut fact = inputs[0].clone();
        fact.shape[self.axis] /= self.stride;
        fact.dim = fact.dim.div_ceil(self.stride as u32);
        Ok(tvec!(fact))
    }

    as_op!();
    pulsed_op_to_typed_op!();
}

fn pull_downsample_up(
    model: &TypedModel,
    down_node: &TypedNode,
) -> TractResult<Option<TypedModelPatch>> {
    let down_op = down_node.op_as::<Downsample>().unwrap();
    if let Some(prec) = model.single_prec(down_node.id)? {
        let invariants = prec.op.invariants(model, prec)?;
        debug!("Consider pull {:?} over {:?} (invariants: {:?})", down_op, prec, invariants);
        if let Some(crop_op) = prec.op_as::<ops::array::Slice<TDim>>() {
            return array::pull_downsample_over_slice(model, prec, crop_op, down_node, down_op);
        } else if let Some(crop_op) = prec.op_as::<ops::array::Slice<usize>>() {
            return array::pull_downsample_over_slice(model, prec, crop_op, down_node, down_op);
        } else if let Some(other_op) = prec.op_as::<AxisOp>() {
            return array::pull_downsample_over_axis_op(model, prec, other_op, down_node, down_op);
        } else if let Some(conv_op) = prec.op_as::<ops::cnn::conv::ConvUnary>() {
            return conv::fuse_downsample_into_conv(model, prec, conv_op, down_node, down_op);
        } else if let Some(other_op) = prec.op_as::<ops::scan::Scan>() {
            return scan::pull_downsample_over_scan(model, prec, other_op, down_node, down_op);
        } else if let Some(above_axis) = invariants.unary_track_axis_up(down_op.axis, false) {
            let mut patch = TypedModelPatch::default();
            let mut inputs = vec![];
            for (ix, &oo) in prec.inputs.iter().enumerate() {
                let source = patch.tap_model(model, oo)?;
                let mut op = down_op.clone();
                op.axis = above_axis;
                let ds = patch.wire_node(format!("{}-{}", prec.name, ix), op, [source].as_ref())?;
                inputs.push(ds[0]);
            }
            let other = patch.wire_node(&*prec.name, prec.op.clone(), &*inputs)?;
            patch.shunt_outside(model, OutletId::new(down_node.id, 0), other[0])?;
            return Ok(Some(patch));
        }
    }
    Ok(None)
}