Add reduce max ONNX op support (#1636)

* Add reduce max onnx op support

* Fix comments on tensor rank 1 result

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

@@ -139,7 +139,7 @@ represent the corresponding Burn Op.
 | [ReduceL][132]                   |       ❌       |      ❌      |
 | [ReduceLogSum][133]              |       ❌       |      ❌      |
 | [ReduceLogSumExp][134]           |       ❌       |      ❌      |
-| [ReduceMax][135]                 |       ❌       |      ✅      |
+| [ReduceMax][135]                 |       ✅       |      ✅      |
 | [ReduceMean][136]                |       ✅       |      ✅      |
 | [ReduceMin][137]                 |       ❌       |      ✅      |
 | [ReduceProd][138]                |       ❌       |      ✅      |

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

@@ -37,6 +37,7 @@ fn main() {
         .input("tests/recip/recip.onnx")
         .input("tests/relu/relu.onnx")
         .input("tests/leaky_relu/leaky_relu.onnx")
+        .input("tests/reduce_max/reduce_max.onnx")
         .input("tests/reduce_mean/reduce_mean.onnx")
         .input("tests/reshape/reshape.onnx")
         .input("tests/shape/shape.onnx")

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

@@ -45,6 +45,7 @@ include_models!(
     neg,
     not,
     recip,
+    reduce_max,
     reduce_mean,
     relu,
     reshape,
@@ -449,6 +450,22 @@ mod tests {
         output3.to_data().assert_approx_eq(&expected3, 3);
     }
 
+    #[test]
+    fn reduce_max() {
+        let device = Default::default();
+        let model: reduce_max::Model<Backend> = reduce_max::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([25.]);
+        let expected = Data::from([[[[25.]]]]);
+
+        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 reduce_mean() {
         let device = Default::default();

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

@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+
+# used to generate model: onnx-tests/tests/reduce_max/reduce_max.onnx
+
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, x):
+        return (
+            # ReduceMax, keepdims=0, axes=None
+            torch.max(x),
+            # ReduceMax, keepdims=1, axes=[1]
+            torch.max(x, dim=1, keepdim=True).values,
+            # ReduceMax, keepdims=1, axes=[-1]
+            torch.max(x, dim=-1, keepdim=True).values,
+        )
+
+
+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_max.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

@@ -27,13 +27,14 @@ pub enum UnaryNodeKind {
     Exp,
     Flatten,
     Gelu,
+    LeakyRelu,
     Log,
     LogSoftmax,
     Neg,
     Not,
+    ReduceMax,
     ReduceMean,
     Reciprocal,
-    LeakyRelu,
     Relu,
     Shape,
     Sigmoid,
@@ -53,13 +54,14 @@ impl UnaryNodeKind {
             Self::Exp => "exp",
             Self::Flatten => "flatten",
             Self::Gelu => "gelu",
+            Self::LeakyRelu => "leaky_relu",
             Self::Log => "log",
             Self::LogSoftmax => "log_softmax",
             Self::Neg => "neg",
             Self::Not => "not",
+            Self::ReduceMax => "reduce_max",
             Self::ReduceMean => "reduce_mean",
             Self::Reciprocal => "reciprocal",
-            Self::LeakyRelu => "leaky_relu",
             Self::Relu => "relu",
             Self::Shape => "shape",
             Self::Sigmoid => "sigmoid",
@@ -294,6 +296,36 @@ impl UnaryNode {
         }
     }
 
+    pub(crate) fn reduce_max(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 {
+                    // Max is only implemented on numeric tensors
+                    panic!("ReduceMax is not supported for boolean");
+                }
+
+                // ReduceMax, keepdims=1, axes=[dim]
+                let dim = dim.to_tokens();
+                Self::new(
+                    input,
+                    output,
+                    UnaryNodeKind::ReduceMax,
+                    Rc::new(move |input| quote! { #input.max_dim(#dim) }),
+                )
+            } else {
+                // ReduceMax, keepdims=0, axes=None
+                Self::new(
+                    input,
+                    output,
+                    UnaryNodeKind::ReduceMax,
+                    Rc::new(move |input| quote! { #input.max() }),
+                )
+            }
+        } else {
+            panic!("ReduceMax only supports tensor output");
+        }
+    }
+
     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 {
@@ -519,6 +551,43 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_unary_codegen_reduce_max() {
+        one_node_graph(
+            UnaryNode::reduce_max(
+                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.max_dim(1);
+
+                    tensor2
+                }
+            },
+            vec!["tensor1".to_string()],
+            vec!["tensor2".to_string()],
+        );
+
+        one_node_graph(
+            UnaryNode::reduce_max(
+                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.max();
+
+                    tensor2
+                }
+            },
+            vec!["tensor1".to_string()],
+            vec!["tensor2".to_string()],
+        );
+    }
+
     #[test]
     fn test_unary_codegen_reduce_mean() {
         one_node_graph(

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

@@ -40,6 +40,7 @@ pub fn dim_inference(node: &mut Node, graph_io: &mut OnnxGraphIO) {
         NodeType::Neg => same_as_input(node),
         NodeType::Not => same_as_input(node),
         NodeType::Reciprocal => same_as_input(node),
+        NodeType::ReduceMax => reduce_max_update_outputs(node),
         NodeType::ReduceMean => reduce_mean_update_outputs(node),
         NodeType::Relu => same_as_input(node),
         NodeType::Reshape => reshape_update_outputs(node),
@@ -239,8 +240,10 @@ fn reduce_mean_update_outputs(node: &mut Node) {
         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
+        // 0-dim tensor so we can't track or perform other ops on that value if we call
+        // `.into_scalar()` on the result of `tensor.max()`
         // node.outputs[0].ty = ArgType::Scalar(tensor.elem_type);
+        // Instead, we return a tensor of rank 1 (the result of `tensor.max()`)
         node.outputs[0].ty = ArgType::Tensor(TensorType { dim: 1, ..tensor });
     }
 }
@@ -438,3 +441,36 @@ fn conv_transpose2d_update_outputs(node: &mut Node) {
         panic!("Only tensor input is valid");
     }
 }
+
+/// Infers the shape of a ReduceMax node and replaces the shape of the output tensor.
+fn reduce_max_update_outputs(node: &mut Node) {
+    if node.inputs.len() != 1 {
+        panic!("ReduceMax: 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 {
+        // NOTE: ReduceMax 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 if we call
+        // `.into_scalar()` on the result of `tensor.max()`
+        // node.outputs[0].ty = ArgType::Scalar(tensor.elem_type);
+        // Instead, we return a tensor of rank 1 (the result of `tensor.max()`)
+        node.outputs[0].ty = ArgType::Tensor(TensorType { dim: 1, ..tensor });
+    }
+}

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

@@ -661,6 +661,50 @@ fn padding_config_1d(pads: &[i64]) -> PaddingConfig1d {
     }
 }
 
+pub fn reduce_max_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!("ReduceMax: reducing on multiple dimensions is not supported")
+    }
+
+    if axes.is_empty() && keepdims == 1 {
+        panic!("ReduceMax: axes must be provided with keepdims")
+    }
+
+    if !axes.is_empty() && keepdims == 0 {
+        // Not supported in Burn
+        panic!("ReduceMax: 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_mean_config(node: &Node) -> Option<usize> {
     let mut axes = Vec::new();
     let mut keepdims = 1;

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

@@ -253,6 +253,7 @@ 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::ReduceMax => graph.register(Self::reduce_max_conversion(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)),
@@ -467,6 +468,14 @@ impl OnnxGraph {
         ReshapeNode::new(input, output, shape)
     }
 
+    fn reduce_max_conversion(node: Node) -> UnaryNode {
+        let input = node.inputs.first().unwrap().to_type();
+        let output = node.outputs.first().unwrap().to_type();
+        let dim = reduce_max_config(&node);
+
+        UnaryNode::reduce_max(input, output, dim)
+    }
+
     fn reduce_mean_conversion(node: Node) -> UnaryNode {
         let input = node.inputs.first().unwrap().to_type();
         let output = node.outputs.first().unwrap().to_type();