Support num_outputs attribute in Split operator

Author: robertknight

rten-convert/rten_convert/converter.py

@@ -653,7 +653,7 @@ def op_node_from_onnx_operator(
         case "Split":
             attrs = sg.SplitAttrsT()
             attrs.axis = attr_reader.get_attr("axis", "int", 0)
-            attr_reader.check_attr("num_outputs", "int", 0)
+            attrs.numOutputs = attr_reader.get_attr("num_outputs", "int", None)
             attr_reader.generate_input_from_attr(1, "split", "ints")
 
         case "Squeeze":

rten-convert/rten_convert/schema_generated.py

@@ -4911,12 +4911,22 @@ def Axis(self):
             return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
         return 0
 
+    # SplitAttrs
+    def NumOutputs(self):
+        o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+        if o != 0:
+            return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+        return None
+
 def SplitAttrsStart(builder):
-    builder.StartObject(1)
+    builder.StartObject(2)
 
 def SplitAttrsAddAxis(builder, axis):
     builder.PrependInt32Slot(0, axis, 0)
 
+def SplitAttrsAddNumOutputs(builder, numOutputs):
+    builder.PrependInt32Slot(1, numOutputs, None)
+
 def SplitAttrsEnd(builder):
     return builder.EndObject()
 
@@ -4927,6 +4937,7 @@ class SplitAttrsT(object):
     # SplitAttrsT
     def __init__(self):
         self.axis = 0  # type: int
+        self.numOutputs = None  # type: Optional[int]
 
     @classmethod
     def InitFromBuf(cls, buf, pos):
@@ -4950,11 +4961,13 @@ def _UnPack(self, splitAttrs):
         if splitAttrs is None:
             return
         self.axis = splitAttrs.Axis()
+        self.numOutputs = splitAttrs.NumOutputs()
 
     # SplitAttrsT
     def Pack(self, builder):
         SplitAttrsStart(builder)
         SplitAttrsAddAxis(builder, self.axis)
+        SplitAttrsAddNumOutputs(builder, self.numOutputs)
         splitAttrs = SplitAttrsEnd(builder)
         return splitAttrs
 

src/model.rs

@@ -1646,7 +1646,10 @@ mod tests {
         let split_out_2 = graph_builder.add_value("Split_out_2", None, None);
         graph_builder.add_operator(
             "Split",
-            OpType::Split(ops::Split { axis: 1 }),
+            OpType::Split(ops::Split {
+                axis: 1,
+                num_outputs: None,
+            }),
             &[input_2d, split_splits].map(Some),
             &[split_out_1, split_out_2],
         );

src/model_builder.rs

@@ -858,6 +858,7 @@ impl<'mb, 'a> GraphBuilder<'mb, 'a> {
             OpType::Split(args) => op_with_attrs!(Split, SplitAttrs, {
                 sg::SplitAttrsArgs {
                     axis: args.axis as i32,
+                    num_outputs: args.num_outputs.map(|n| n as i32),
                 }
             }),
             OpType::Sqrt => op!(Sqrt),

src/op_registry.rs

@@ -880,7 +880,11 @@ impl_read_op!(Size);
 impl_read_op!(Slice);
 impl_read_op!(Softmax, attrs_as_softmax_attrs, axis);
 impl_read_op!(Softplus);
-impl_read_op!(Split, attrs_as_split_attrs, axis);
+impl_read_op!(Split, attrs_as_split_attrs, |attrs: sg::SplitAttrs| {
+    let axis = attrs.axis() as isize;
+    let num_outputs = attrs.num_outputs().map(|n| n as u32);
+    Ok(ops::Split { axis, num_outputs })
+});
 impl_read_op!(Sqrt);
 impl_read_op!(Squeeze);
 impl_read_op!(Sub);

src/ops/rnn.rs

@@ -807,7 +807,7 @@ mod tests {
         let splits = &[size as i32; 4];
 
         // Split input into seperate tensor for each of the gates.
-        let ifco = split(&pool, x.view(), axis, &splits.into()).expect("split failed");
+        let ifco = split(&pool, x.view(), axis, splits.as_slice().into()).expect("split failed");
 
         // Recombine in a new gate order.
         concat(
@@ -831,7 +831,7 @@ mod tests {
         let splits = &[size as i32; 3];
 
         // Split input into seperate tensor for each of the gates.
-        let ruh = split(&pool, x.view(), axis, &splits.into()).expect("split failed");
+        let ruh = split(&pool, x.view(), axis, splits.as_slice().into()).expect("split failed");
 
         // Recombine in a new gate order.
         concat(&pool, &[ruh[1].view(), ruh[0].view(), ruh[2].view()], axis).expect("concat failed")

src/ops/split.rs

@@ -1,46 +1,82 @@
 use rten_tensor::prelude::*;
 use rten_tensor::{NdTensorView, Tensor, TensorView};
 
+use crate::iter_util::range_chunks;
 use crate::ops::{
     map_input, resolve_axis, static_dims, Input, InputList, OpError, Operator, OutputList,
 };
 use crate::tensor_pool::TensorPool;
 
+#[derive(Clone, Debug)]
+pub enum SplitSizes<'a> {
+    /// Split a tensor into pieces with sizes specified by a vector. The sum of
+    /// the piece sizes must match the size of the axis.
+    Sizes(NdTensorView<'a, i32, 1>),
+    /// Split a tensor into N equal-sized pieces. If the size of the axis being
+    /// split is not evenly divisible by N, the last chunk will be smaller.
+    NumSplits(u32),
+}
+
+impl<'a> From<&'a [i32]> for SplitSizes<'a> {
+    fn from(val: &'a [i32]) -> Self {
+        Self::Sizes(val.into())
+    }
+}
+
 pub fn split<T: Copy>(
     pool: &TensorPool,
     input: TensorView<T>,
     axis: isize,
-    split: &NdTensorView<i32, 1>,
+    split: SplitSizes,
 ) -> Result<Vec<Tensor<T>>, OpError> {
     let axis = resolve_axis(input.ndim(), axis)?;
 
-    if split.iter().any(|size| *size < 0) {
-        return Err(OpError::InvalidValue("Split sizes must be >= 0"));
-    }
-    let split_sum = split.iter().sum::<i32>() as usize;
-    if split_sum != input.size(axis) {
-        return Err(OpError::InvalidValue(
-            "Split sizes do not sum to dimension size",
-        ));
-    }
-
-    let mut split_start = 0;
-    let outputs = split
-        .iter()
-        .map(|&split_size| {
-            let split_size = split_size as usize;
-            let split_range = split_start..split_start + split_size;
-            split_start += split_size;
-            input.slice_axis(axis, split_range).to_tensor_in(pool)
-        })
-        .collect();
+    let outputs = match split {
+        SplitSizes::Sizes(split) => {
+            if split.iter().any(|size| *size < 0) {
+                return Err(OpError::InvalidValue("Split sizes must be >= 0"));
+            }
+            let split_sum = split.iter().sum::<i32>() as usize;
+            if split_sum != input.size(axis) {
+                return Err(OpError::InvalidValue(
+                    "Split sizes do not sum to dimension size",
+                ));
+            }
+
+            let mut split_start = 0;
+            split
+                .iter()
+                .map(|&split_size| {
+                    let split_size = split_size as usize;
+                    let split_range = split_start..split_start + split_size;
+                    split_start += split_size;
+                    input.slice_axis(axis, split_range).to_tensor_in(pool)
+                })
+                .collect()
+        }
+        SplitSizes::NumSplits(n_splits) => {
+            let n_splits = n_splits as usize;
+            if n_splits == 0 {
+                return Err(OpError::InvalidValue("num_outputs must be > 0"));
+            }
+            let dim_size = input.size(axis);
+            if n_splits > dim_size {
+                return Err(OpError::InvalidValue("num_outputs exceeds dim size"));
+            }
+            let chunk_size = dim_size.div_ceil(n_splits);
+            range_chunks(0..dim_size, chunk_size)
+                .map(|chunk| input.slice_axis(axis, chunk).to_tensor_in(pool))
+                .collect()
+        }
+    };
 
     Ok(outputs)
 }
 
 #[derive(Debug)]
 pub struct Split {
     pub axis: isize,
+    pub num_outputs: Option<u32>,
 }
 
 impl Operator for Split {
@@ -50,11 +86,21 @@ impl Operator for Split {
 
     fn run(&self, pool: &TensorPool, inputs: InputList) -> Result<OutputList, OpError> {
         let input = inputs.require(0)?;
-        let splits = inputs.require_as::<i32>(1)?;
-        let splits = static_dims!(splits, 1)?;
+        let splits = inputs.get_as::<i32>(1)?;
+
+        let split_sizes = if let Some(splits) = splits {
+            let splits = static_dims!(splits, 1)?;
+            SplitSizes::Sizes(splits)
+        } else if let Some(num_outputs) = self.num_outputs {
+            SplitSizes::NumSplits(num_outputs)
+        } else {
+            return Err(OpError::InvalidValue(
+                "Either `num_outputs` or `splits` must be set",
+            ));
+        };
 
         map_input!(input, x, {
-            split(pool, x, self.axis, &splits)
+            split(pool, x, self.axis, split_sizes)
                 .map(|tensors| tensors.into_iter().map(|t| t.into()).collect())
         })
     }
@@ -64,60 +110,113 @@ impl Operator for Split {
 mod tests {
     use rten_tensor::prelude::*;
     use rten_tensor::Tensor;
+    use rten_testing::TestCases;
 
     use crate::ops::tests::new_pool;
     use crate::ops::{split, OpError};
 
+    use super::SplitSizes;
+
     #[test]
     fn test_split() {
-        let pool = new_pool();
-
         let input = Tensor::from([[0., 1.], [2., 3.], [4., 5.], [6., 7.], [8., 9.]]);
 
-        // Split with positive axis
-        let splits = &[1, 1];
-        let results = split(&pool, input.view(), 1, &splits.into()).unwrap();
-
-        assert_eq!(results.len(), 2);
-        assert_eq!(results[0].data().unwrap(), &[0., 2., 4., 6., 8.]);
-        assert_eq!(results[1].data().unwrap(), &[1., 3., 5., 7., 9.]);
-
-        // Split with negative axis
-        let splits = &[1, 1];
-        let results = split(&pool, input.view(), -1, &splits.into()).unwrap();
-
-        assert_eq!(results.len(), 2);
-        assert_eq!(results[0].data().unwrap(), &[0., 2., 4., 6., 8.]);
-        assert_eq!(results[1].data().unwrap(), &[1., 3., 5., 7., 9.]);
+        #[derive(Debug)]
+        struct Case<'a> {
+            axis: isize,
+            splits: SplitSizes<'a>,
+            expected: Vec<Tensor>,
+        }
+
+        let cases = [
+            // Positive axis
+            Case {
+                axis: 1,
+                splits: [1, 1].as_slice().into(),
+                expected: [
+                    Tensor::from([[0.], [2.], [4.], [6.], [8.]]),
+                    Tensor::from([[1.], [3.], [5.], [7.], [9.]]),
+                ]
+                .into(),
+            },
+            // Negative axis
+            Case {
+                axis: -1,
+                splits: [1, 1].as_slice().into(),
+                expected: [
+                    Tensor::from([[0.], [2.], [4.], [6.], [8.]]),
+                    Tensor::from([[1.], [3.], [5.], [7.], [9.]]),
+                ]
+                .into(),
+            },
+            // Splits specified as count
+            Case {
+                axis: 0,
+                splits: SplitSizes::NumSplits(3),
+                expected: [
+                    Tensor::from([[0., 1.], [2., 3.]]),
+                    Tensor::from([[4., 5.], [6., 7.]]),
+                    Tensor::from([[8., 9.]]),
+                ]
+                .into(),
+            },
+        ];
+
+        cases.test_each(|case| {
+            let pool = new_pool();
+            let results = split(&pool, input.view(), case.axis, case.splits.clone()).unwrap();
+            let expected_splits = match case.splits {
+                SplitSizes::NumSplits(n) => n as usize,
+                SplitSizes::Sizes(sizes) => sizes.len(),
+            };
+            assert_eq!(results.len(), expected_splits);
+            assert_eq!(results, case.expected);
+        })
     }
 
     #[test]
     fn test_split_invalid_inputs() {
-        let pool = new_pool();
-
         let input = Tensor::from([[0., 1.], [2., 3.], [4., 5.], [6., 7.], [8., 9.]]);
 
-        let splits = &[1, 1];
-        let result = split(&pool, input.view(), 2, &splits.into());
-        assert_eq!(result.err(), Some(OpError::InvalidValue("Axis is invalid")));
-
-        let result = split(&pool, input.view(), -3, &splits.into());
-        assert_eq!(result.err(), Some(OpError::InvalidValue("Axis is invalid")));
-
-        let splits = &[1, 2];
-        let result = split(&pool, input.view(), 1, &splits.into());
-        assert_eq!(
-            result.err(),
-            Some(OpError::InvalidValue(
-                "Split sizes do not sum to dimension size"
-            ))
-        );
-
-        let splits = &[1, -2];
-        let result = split(&pool, input.view(), 1, &splits.into());
-        assert_eq!(
-            result.err(),
-            Some(OpError::InvalidValue("Split sizes must be >= 0"))
-        );
+        #[derive(Debug)]
+        struct Case<'a> {
+            axis: isize,
+            splits: SplitSizes<'a>,
+            expected: OpError,
+        }
+
+        let cases = [
+            Case {
+                axis: 2,
+                splits: [1, 1].as_slice().into(),
+                expected: OpError::InvalidValue("Axis is invalid"),
+            },
+            Case {
+                axis: 1,
+                splits: [1, 2].as_slice().into(),
+                expected: OpError::InvalidValue("Split sizes do not sum to dimension size"),
+            },
+            Case {
+                axis: 1,
+                splits: [1, -2].as_slice().into(),
+                expected: OpError::InvalidValue("Split sizes must be >= 0"),
+            },
+            Case {
+                axis: 1,
+                splits: SplitSizes::NumSplits(0),
+                expected: OpError::InvalidValue("num_outputs must be > 0"),
+            },
+            Case {
+                axis: 1,
+                splits: SplitSizes::NumSplits(3),
+                expected: OpError::InvalidValue("num_outputs exceeds dim size"),
+            },
+        ];
+
+        cases.test_each(|case| {
+            let pool = new_pool();
+            let result = split(&pool, input.view(), case.axis, case.splits.clone());
+            assert_eq!(result.err().as_ref(), Some(&case.expected));
+        })
     }
 }

src/schema.fbs

@@ -485,6 +485,7 @@ table SoftmaxAttrs {
 
 table SplitAttrs {
   axis:int;
+  num_outputs:int = null;
 }
 
 table TopKAttrs {