Add reduce mean ONNX op support (#1637)

* Add reduce mean onnx op support

* Fix comment

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

@@ -140,7 +140,7 @@ represent the corresponding Burn Op.
 | [ReduceLogSum][133]              |       ❌       |      ❌      |
 | [ReduceLogSumExp][134]           |       ❌       |      ❌      |
 | [ReduceMax][135]                 |       ❌       |      ✅      |
-| [ReduceMean][136]                |       ❌       |      ✅      |
+| [ReduceMean][136]                |       ✅       |      ✅      |
 | [ReduceMin][137]                 |       ❌       |      ✅      |
 | [ReduceProd][138]                |       ❌       |      ✅      |
 | [ReduceSum][139]                 |       ❌       |      ✅      |

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

@@ -35,6 +35,7 @@ fn main() {
         .input("tests/recip/recip.onnx")
         .input("tests/relu/relu.onnx")
         .input("tests/leaky_relu/leaky_relu.onnx")
+        .input("tests/reduce_mean/reduce_mean.onnx")
         .input("tests/reshape/reshape.onnx")
         .input("tests/sigmoid/sigmoid.onnx")
         .input("tests/sin/sin.onnx")

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

@@ -41,6 +41,7 @@ include_models!(
     mul,
     neg,
     recip,
+    reduce_mean,
     relu,
     reshape,
     sigmoid,
@@ -443,6 +444,22 @@ mod tests {
         output3.to_data().assert_approx_eq(&expected3, 3);
     }
 
+    #[test]
+    fn reduce_mean() {
+        let device = Default::default();
+        let model: reduce_mean::Model<Backend> = reduce_mean::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 = Data::from([9.75]);
+        let expected = Data::from([[[[9.75]]]]);
+
+        assert_eq!(output_scalar.to_data(), expected_scalar);
+        assert_eq!(output_tensor.to_data(), input.to_data());
+        assert_eq!(output_value.to_data(), expected);
+    }
+
     #[test]
     fn reshape() {
         // Initialize the model without weights (because the exported file does not contain them)

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

@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+
+# used to generate model: onnx-tests/tests/reduce_mean/reduce_mean.onnx
+
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, x):
+        return (
+            # ReduceMean, keepdims=0, axes=None
+            torch.mean(x),
+            # ReduceMean, keepdims=1, axes=[1]
+            torch.mean(x, dim=1, keepdim=True),
+            # ReduceMean, keepdims=1, axes=[-1]
+            torch.mean(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")
+    onnx_name = "reduce_mean.onnx"
+    test_input = torch.tensor([[[[1.0, 4.0, 9.0, 25.0]]]], device=device)
+
+    torch.onnx.export(model, test_input, onnx_name, verbose=False, opset_version=16)
+
+    print(f"Finished exporting model to {onnx_name}")
+
+    # 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

@@ -29,6 +29,7 @@ pub enum UnaryNodeKind {
     Log,
     LogSoftmax,
     Neg,
+    ReduceMean,
     Reciprocal,
     LeakyRelu,
     Relu,
@@ -52,6 +53,7 @@ impl UnaryNodeKind {
             Self::Log => "log",
             Self::LogSoftmax => "log_softmax",
             Self::Neg => "neg",
+            Self::ReduceMean => "reduce_mean",
             Self::Reciprocal => "reciprocal",
             Self::LeakyRelu => "leaky_relu",
             Self::Relu => "relu",
@@ -253,6 +255,32 @@ impl UnaryNode {
             _ => panic!("output must be a tensor"),
         }
     }
+
+    pub(crate) fn reduce_mean(input: Type, output: Type, dim: Option<usize>) -> Self {
+        // ReduceMean is constrained to numeric tensors, so no need to check for bool.
+        if let Type::Tensor(_) = output {
+            if let Some(dim) = dim {
+                // ReduceMean, keepdims=1, axes=[dim]
+                let dim = dim.to_tokens();
+                Self::new(
+                    input,
+                    output,
+                    UnaryNodeKind::ReduceMean,
+                    Rc::new(move |input| quote! { #input.mean_dim(#dim) }),
+                )
+            } else {
+                // ReduceMean, keepdims=0, axes=None
+                Self::new(
+                    input,
+                    output,
+                    UnaryNodeKind::ReduceMean,
+                    Rc::new(move |input| quote! { #input.mean() }),
+                )
+            }
+        } else {
+            panic!("ReduceMean only supports tensor output");
+        }
+    }
 }
 
 #[cfg(test)]
@@ -437,6 +465,43 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_unary_codegen_reduce_mean() {
+        one_node_graph(
+            UnaryNode::reduce_mean(
+                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.mean_dim(1);
+
+                    tensor2
+                }
+            },
+            vec!["tensor1".to_string()],
+            vec!["tensor2".to_string()],
+        );
+
+        one_node_graph(
+            UnaryNode::reduce_mean(
+                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.mean();
+
+                    tensor2
+                }
+            },
+            vec!["tensor1".to_string()],
+            vec!["tensor2".to_string()],
+        );
+    }
+
     #[test]
     fn test_unary_codegen_reciprocal() {
         one_node_graph(

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

@@ -39,7 +39,7 @@ pub fn dim_inference(node: &mut Node, graph_io: &mut OnnxGraphIO) {
         NodeType::Mul => same_as_input(node),
         NodeType::Neg => same_as_input(node),
         NodeType::Reciprocal => same_as_input(node),
-        NodeType::ReduceMean => mean_update_outputs(node),
+        NodeType::ReduceMean => reduce_mean_update_outputs(node),
         NodeType::Relu => same_as_input(node),
         NodeType::Reshape => reshape_update_outputs(node),
         NodeType::Shape => shape_update_outputs(node),
@@ -206,12 +206,11 @@ fn reshape_update_outputs(node: &mut Node) {
     });
 }
 
-fn mean_update_outputs(node: &mut Node) {
+fn reduce_mean_update_outputs(node: &mut Node) {
     if node.inputs.len() != 1 {
         panic!("Mean: multiple inputs are not supported");
     }
 
-    // Extract the configuration of the linear layer (inputs are known)
     let node_input = &mut node.inputs[0];
     let tensor = match node_input.clone().ty {
         ArgType::Tensor(tensor) => tensor,
@@ -230,6 +229,9 @@ fn mean_update_outputs(node: &mut Node) {
     if dim_only {
         node.outputs[0].ty = ArgType::Tensor(tensor);
     } else {
+        // NOTE: ReduceMean w/o keepdims reduces to a scalar value, but Burn doesn't have
+        // 0-dim tensor so we can't track or perform other ops on that value
+        // node.outputs[0].ty = ArgType::Scalar(tensor.elem_type);
         node.outputs[0].ty = ArgType::Tensor(TensorType { dim: 1, ..tensor });
     }
 }

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

@@ -660,3 +660,47 @@ fn padding_config_1d(pads: &[i64]) -> PaddingConfig1d {
         panic!("Padding configuration ({:?}) not supported", pads);
     }
 }
+
+pub fn reduce_mean_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(),
+            _ => {}
+        }
+    }
+
+    if axes.len() > 1 {
+        panic!("ReduceMean: reducing on multiple dimensions is not supported")
+    }
+
+    if axes.is_empty() && keepdims == 1 {
+        panic!("ReduceMean: axes must be provided with keepdims")
+    }
+
+    if !axes.is_empty() && keepdims == 0 {
+        // Not supported in Burn
+        panic!("ReduceMean: 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)
+    }
+}

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

@@ -252,6 +252,7 @@ impl OnnxGraph {
                 NodeType::Sqrt => graph.register(Self::sqrt_conversion(node)),
                 NodeType::Tanh => graph.register(Self::tanh_conversion(node)),
                 NodeType::Constant => graph.register(Self::constant_conversion::<PS>(node)),
+                NodeType::ReduceMean => graph.register(Self::reduce_mean_conversion(node)),
                 NodeType::Reshape => graph.register(Self::reshape_conversion(node)),
                 NodeType::Reciprocal => graph.register(Self::reciprocal_conversion(node)),
                 NodeType::Sigmoid => graph.register(Self::sigmoid_conversion(node)),
@@ -463,6 +464,15 @@ impl OnnxGraph {
 
         ReshapeNode::new(input, output, shape)
     }
+
+    fn reduce_mean_conversion(node: Node) -> UnaryNode {
+        let input = node.inputs.first().unwrap().to_type();
+        let output = node.outputs.first().unwrap().to_type();
+        let dim = reduce_mean_config(&node);
+
+        UnaryNode::reduce_mean(input, output, dim)
+    }
+
     fn unsqueeze_conversion(node: Node) -> UnsqueezeNode {
         let input = node.inputs.first().unwrap().to_tensor_type();
         let output = node.outputs.first().unwrap().to_tensor_type();