use tract_hir::internal::*;
use crate::model::ParsingContext;
use crate::tfpb::tensorflow::NodeDef;
pub fn build(_ctx: &ParsingContext, _pb: &NodeDef) -> TractResult<Box<dyn InferenceOp>> {
Ok(expand(ExpandDims))
}
#[derive(Debug, Clone, Hash)]
pub struct ExpandDims;
impl Expansion for ExpandDims {
fn name(&self) -> StaticName {
"ExpandDims".into()
}
fn rules<'r, 'p: 'r, 's: 'r>(
&'s self,
s: &mut Solver<'r>,
inputs: &'p [TensorProxy],
outputs: &'p [TensorProxy],
) -> InferenceResult {
let data = &inputs[0];
let dims = &inputs[1];
let output = &outputs[0];
check_input_arity(inputs, 2)?;
check_output_arity(outputs, 1)?;
s.equals(&dims.datum_type, DatumType::I32)?;
s.equals(&data.datum_type, &output.datum_type)?;
s.equals(data.rank.bex() + 1, &output.rank)?;
s.given_2(&dims.value, &data.rank, move |s, index, rank| {
let mut index = index.cast_to_scalar::<i64>()?;
if index < 0 {
index += rank + 1
}
let index = index as usize;
for i in 0..index {
s.equals(&output.shape[i], &data.shape[i])?;
}
s.equals(output.shape[index].bex(), 1i64.to_dim().bex())?;
s.given(&data.rank, move |s, rank| {
for i in index..(rank as usize) {
s.equals(&output.shape[i + 1], &data.shape[i])?;
}
Ok(())
})
})
}
fn wire(
&self,
prefix: &str,
target: &mut TypedModel,
inputs: &[OutletId],
) -> TractResult<TVec<OutletId>> {
if let Some(ref axes) = target.outlet_fact(inputs[1])?.konst {
let mut axes = axes
.cast_to::<i32>()?
.as_slice::<i32>()?
.iter()
.map(|&axis| {
Ok(if axis < 0 {
axis + target.outlet_fact(inputs[0])?.shape.rank() as i32
} else {
axis
})
})
.collect::<TractResult<Vec<_>>>()?;
axes.sort();
let mut wire = inputs[0];
for axis in axes.iter().rev() {
wire = target.wire_node(
format!("{prefix}.axis-{axis}"),
AxisOp::Add(*axis as _),
&[wire],
)?[0];
}
Ok(tvec!(wire))
} else {
bail!("Need axes to be const")
}
}
}