Add ReduceProd ONNX Import (#1955)

* Preliminary ReduceProd Support

* Add comma to keep formatter happy

* Give test results a 0.001 tolerance to account for floating-point multiplication

* Reformat assersions

* Correctly mark panic conditions in op_configuration

crates/burn-import/SUPPORTED-ONNX-OPS.md

@@ -142,7 +142,7 @@ represent the corresponding Burn Op.
 | [ReduceMax][135]                 |       ✅       |      ✅      |
 | [ReduceMean][136]                |       ✅       |      ✅      |
 | [ReduceMin][137]                 |       ✅       |      ✅      |
-| [ReduceProd][138]                |       ❌       |      ✅      |
+| [ReduceProd][138]                |       ✅       |      ✅      |
 | [ReduceSum][139]                 |       ✅       |      ✅      |
 | [ReduceSumSquare][140]           |       ❌       |      ❌      |
 | [Relu][141]                      |       ✅       |      ✅      |

crates/burn-import/onnx-tests/build.rs

@@ -54,6 +54,7 @@ fn main() {
         .input("tests/reduce_max/reduce_max.onnx")
         .input("tests/reduce_min/reduce_min.onnx")
         .input("tests/reduce_mean/reduce_mean.onnx")
+        .input("tests/reduce_prod/reduce_prod.onnx")
         .input("tests/reduce_sum/reduce_sum_opset13.onnx")
         .input("tests/reduce_sum/reduce_sum_opset11.onnx")
         .input("tests/reshape/reshape.onnx")

crates/burn-import/onnx-tests/tests/onnx_tests.rs

@@ -64,6 +64,7 @@ include_models!(
     reduce_max,
     reduce_min,
     reduce_mean,
+    reduce_prod,
     reduce_sum_opset13,
     reduce_sum_opset11,
     relu,
@@ -761,6 +762,27 @@ mod tests {
         output_value.to_data().assert_eq(&expected, true);
     }
 
+    #[test]
+    fn reduce_prod() {
+        let device = Default::default();
+        let model: reduce_prod::Model<Backend> = reduce_prod::Model::new(&device);
+
+        // Run the model
+        let input = Tensor::<Backend, 4>::from_floats([[[[1.0, 4.0, 9.0, 25.0]]]], &device);
+        let (output_scalar, output_tensor, output_value) = model.forward(input.clone());
+        let expected_scalar = TensorData::from([900f32]);
+        let expected = TensorData::from([[[[900f32]]]]);
+
+        // Tolerance of 0.001 since floating-point multiplication won't be perfect
+        output_scalar
+            .to_data()
+            .assert_approx_eq(&expected_scalar, 3);
+        output_tensor
+            .to_data()
+            .assert_approx_eq(&input.to_data(), 3);
+        output_value.to_data().assert_approx_eq(&expected, 3);
+    }
+
     #[test]
     fn reduce_sum_opset11() {
         let device = Default::default();

crates/burn-import/onnx-tests/tests/reduce_prod/reduce_prod.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+
+# used to generate model: onnx-tests/tests/reduce_sum/reduce_sum.onnx
+
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, x):
+        return (
+            # ReduceProd, keepdims=0, axes=None
+            torch.prod(x),
+            # ReduceProd, keepdims=1, axes=[1]
+            torch.prod(x, dim=1, keepdim=True),
+            # ReduceProd, keepdims=1, axes=[-1]
+            torch.prod(x, dim=-1, keepdim=True),
+        )
+
+
+def main():
+    # Set random seed for reproducibility
+    torch.manual_seed(0)
+
+    # Export to onnx
+    model = Model()
+    model.eval()
+    device = torch.device("cpu")
+    test_input = torch.tensor([[[[1.0, 4.0, 9.0, 25.0]]]], device=device)
+
+    torch.onnx.export(model, test_input, "reduce_prod.onnx", verbose=False, opset_version=16)
+
+    print("Finished exporting model")
+
+    # Output some test data for use in the test
+    print(f"Test input data: {test_input}")
+    output = model.forward(*test_input)
+    print(f"Test output data: {output}")
+
+
+if __name__ == "__main__":
+    main()

crates/burn-import/src/burn/node/unary.rs

@@ -35,6 +35,7 @@ pub enum UnaryNodeKind {
     ReduceMax,
     ReduceMin,
     ReduceMean,
+    ReduceProd,
     ReduceSum,
     Reciprocal,
     Relu,
@@ -65,6 +66,7 @@ impl UnaryNodeKind {
             Self::ReduceMax => "reduce_max",
             Self::ReduceMin => "reduce_min",
             Self::ReduceMean => "reduce_mean",
+            Self::ReduceProd => "reduce_prod",
             Self::ReduceSum => "reduce_sum",
             Self::Reciprocal => "reciprocal",
             Self::Relu => "relu",
@@ -388,6 +390,36 @@ impl UnaryNode {
         }
     }
 
+    pub(crate) fn reduce_prod(input: Type, output: Type, dim: Option<usize>) -> Self {
+        if let Type::Tensor(ref tensor) = output {
+            if let Some(dim) = dim {
+                if tensor.kind == TensorKind::Bool {
+                    // Prod is only implemented on numeric tensors
+                    panic!("ReduceProd is not supported for boolean");
+                }
+
+                // ReduceProd, keepdims=1, axes=[dim]
+                let dim = dim.to_tokens();
+                Self::new(
+                    input,
+                    output,
+                    UnaryNodeKind::ReduceProd,
+                    Rc::new(move |input| quote! { #input.prod_dim(#dim) }),
+                )
+            } else {
+                // ReduceProd, keepdims=0, axes=None
+                Self::new(
+                    input,
+                    output,
+                    UnaryNodeKind::ReduceProd,
+                    Rc::new(move |input| quote! { #input.prod() }),
+                )
+            }
+        } else {
+            panic!("ReduceProd only supports tensor output");
+        }
+    }
+
     pub(crate) fn reduce_sum(input: Type, output: Type, dim: Option<usize>) -> Self {
         if let Type::Tensor(ref tensor) = output {
             if let Some(dim) = dim {
@@ -733,6 +765,43 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_unary_codegen_reduce_prod() {
+        one_node_graph(
+            UnaryNode::reduce_prod(
+                Type::Tensor(TensorType::new_float("tensor1", 4)),
+                Type::Tensor(TensorType::new_float("tensor2", 4)),
+                Some(1),
+            ),
+            quote! {
+                pub fn forward(&self, tensor1: Tensor<B, 4>) -> Tensor<B, 4> {
+                    let tensor2 = tensor1.prod_dim(1);
+
+                    tensor2
+                }
+            },
+            vec!["tensor1".to_string()],
+            vec!["tensor2".to_string()],
+        );
+
+        one_node_graph(
+            UnaryNode::reduce_prod(
+                Type::Tensor(TensorType::new_float("tensor1", 4)),
+                Type::Tensor(TensorType::new_float("tensor2", 1)),
+                None,
+            ),
+            quote! {
+                pub fn forward(&self, tensor1: Tensor<B, 4>) -> Tensor<B, 1> {
+                    let tensor2 = tensor1.prod();
+
+                    tensor2
+                }
+            },
+            vec!["tensor1".to_string()],
+            vec!["tensor2".to_string()],
+        );
+    }
+
     #[test]
     fn test_unary_codegen_reduce_sum() {
         one_node_graph(

crates/burn-import/src/onnx/dim_inference.rs

@@ -57,6 +57,7 @@ pub fn dim_inference(node: &mut Node) {
         NodeType::ReduceMax => reduce_max_update_outputs(node),
         NodeType::ReduceMin => reduce_min_update_outputs(node),
         NodeType::ReduceMean => reduce_mean_update_outputs(node),
+        NodeType::ReduceProd => reduce_prod_update_outputs(node),
         NodeType::ReduceSum => reduce_sum_update_outputs(node),
         NodeType::Relu => same_as_input(node),
         NodeType::Reshape => reshape_update_outputs(node),
@@ -741,6 +742,33 @@ fn reduce_min_update_outputs(node: &mut Node) {
     }
 }
 
+/// Infers the shape of a ReduceProd node and replaces the shape of the output tensor.
+fn reduce_prod_update_outputs(node: &mut Node) {
+    if node.inputs.len() != 1 {
+        panic!("ReduceProd: multiple inputs are not supported");
+    }
+    let node_input = &mut node.inputs[0];
+    let tensor = match node_input.clone().ty {
+        ArgType::Tensor(tensor) => tensor,
+        _ => panic!("Only tensor input is valid"),
+    };
+
+    let dim_only = match node.attrs.get("axes") {
+        Some(value) => match &value {
+            AttributeValue::Int64(_) => true,
+            AttributeValue::Int64s(ints) => ints.len() == 1,
+            _ => false,
+        },
+        None => false,
+    };
+
+    if dim_only {
+        node.outputs[0].ty = ArgType::Tensor(tensor);
+    } else {
+        node.outputs[0].ty = ArgType::Tensor(TensorType { dim: 1, ..tensor });
+    }
+}
+
 /// Infers the shape of a ReduceSum node and replaces the shape of the output tensor.
 fn reduce_sum_update_outputs(node: &mut Node) {
     let node_input = &mut node.inputs[0];

crates/burn-import/src/onnx/op_configuration.rs

@@ -988,6 +988,51 @@ pub fn reduce_mean_config(node: &Node) -> Option<usize> {
     }
 }
 
+pub fn reduce_prod_config(node: &Node) -> Option<usize> {
+    let mut axes = Vec::new();
+    let mut keepdims = 1;
+
+    let tensor = match node.inputs.first().unwrap().clone().ty {
+        ArgType::Tensor(tensor) => tensor,
+        _ => panic!("Only tensor input is valid"),
+    };
+
+    // Extract the attributes
+    for (key, value) in node.attrs.iter() {
+        match key.as_str() {
+            "axes" => axes = value.clone().into_i64s(),
+            "keepdims" => keepdims = value.clone().into_i64(),
+            // TODO: handle noop_with_empty_axes (opset 18)
+            _ => {}
+        }
+    }
+
+    if axes.len() > 1 {
+        panic!("ReduceProd: reducing on multiple dimensions is not supported")
+    }
+
+    if axes.is_empty() && keepdims == 1 {
+        panic!("ReduceProd: axes must be provided with keepdims")
+    }
+
+    if !axes.is_empty() && keepdims == 0 {
+        // Not supported in Burn
+        panic!("ReduceProd: the reduce operation must preserve the reduced dimension")
+    }
+
+    if axes.is_empty() {
+        None
+    } else {
+        let mut dim = axes[0];
+
+        if dim < 0 {
+            // Accepted range is [-r, r-1] where r = rank(data) but Burn only supports positive dim
+            dim += tensor.dim as i64;
+        }
+        Some(dim as usize)
+    }
+}
+
 pub fn reduce_sum_config(node: &Node) -> Option<usize> {
     let mut axes = Vec::new();
     let mut keepdims = 1;

crates/burn-import/src/onnx/to_burn.rs

@@ -291,6 +291,7 @@ impl OnnxGraph {
                 NodeType::ReduceMax => graph.register(Self::reduce_max_conversion(node)),
                 NodeType::ReduceMin => graph.register(Self::reduce_min_conversion(node)),
                 NodeType::ReduceMean => graph.register(Self::reduce_mean_conversion(node)),
+                NodeType::ReduceProd => graph.register(Self::reduce_prod_conversion(node)),
                 NodeType::ReduceSum => graph.register(Self::reduce_sum_conversion(node)),
                 NodeType::Reshape => graph.register(Self::reshape_conversion(node)),
                 NodeType::Resize => graph.register(Self::resize_conversion(node)),
@@ -655,6 +656,14 @@ impl OnnxGraph {
         UnaryNode::reduce_mean(input, output, dim)
     }
 
+    fn reduce_prod_conversion(node: Node) -> UnaryNode {
+        let input = node.inputs.first().unwrap().to_type();
+        let output = node.outputs.first().unwrap().to_type();
+        let dim = reduce_prod_config(&node);
+
+        UnaryNode::reduce_prod(input, output, dim)
+    }
+
     fn reduce_sum_conversion(node: Node) -> UnaryNode {
         let input = node.inputs.first().unwrap().to_type();
         let output = node.outputs.first().unwrap().to_type();