Onnx op topk

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

@@ -192,7 +192,7 @@ represent the corresponding Burn Op.
 | [TfIdfVectorizer][183]           |       ❌       |      ❌      |
 | [ThresholdedRelu][184]           |       ❌       |      ❌      |
 | [Tile][185]                      |       ✅       |      ✅      |
-| [TopK][186]                      |       ❌       |      ✅      |
+| [TopK][186]                      |       ✅       |      ✅      |
 | [Transpose][187]                 |       ✅       |      ✅      |
 | [Trilu][188]                     |       ❌       |      ✅      |
 | [Unique][189]                    |       ❌       |      ❌      |

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

@@ -107,6 +107,7 @@ fn main() {
         .input("tests/sum/sum_int.onnx")
         .input("tests/tanh/tanh.onnx")
         .input("tests/tile/tile.onnx")
+        .input("tests/top_k/top_k.onnx")
         .input("tests/transpose/transpose.onnx")
         .input("tests/unsqueeze/unsqueeze.onnx")
         .input("tests/unsqueeze/unsqueeze_opset11.onnx")

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

@@ -116,6 +116,7 @@ include_models!(
     sum_int,
     tanh,
     tile,
+    top_k_opset_1,
     transpose,
     unsqueeze,
     unsqueeze_opset11,
@@ -128,7 +129,7 @@ mod tests {
 
     use super::*;
 
-    use burn::tensor::{Bool, Int, Shape, Tensor, TensorData};
+    use burn::tensor::{cast::ToElement, Bool, Int, Shape, Tensor, TensorData};
 
     use float_cmp::ApproxEq;
 
@@ -2125,4 +2126,30 @@ mod tests {
         assert!(i_output.equal(i_expected).all().into_scalar());
         assert!(b_output.equal(b_expected).all().into_scalar());
     }
+
+    #[test]
+    fn top_k_opset_1() {
+        // Initialize the model
+        let device = Default::default();
+        let model = top_k_opset1::Model::<Backend>::new(&device);
+
+        // Run the model
+        let input = Tensor::<Backend, 2>::from_floats(
+            [[1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0]],
+            &device,
+        );
+        let (values_tensor, indices_tensor) = model.forward(input);
+
+        // expected results
+        let expected_values_tensor =
+            TensorData::from([[4.0, 3.0, 2.to_f32()], [4.0, 3.0, 2.to_f32()]]);
+        let expected_indices_tensor = TensorData::from([[3, 2, 1], [3, 2, 1]]);
+
+        values_tensor
+            .to_data()
+            .assert_eq(&expected_values_tensor, true);
+        indices_tensor
+            .to_data()
+            .assert_eq(&expected_indices_tensor, true);
+    }
 }

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

@@ -0,0 +1,75 @@
+import numpy as np
+import onnx
+from onnx import helper, TensorProto
+
+# Define the input tensor
+X = np.array([[0, 1, 2, 3],
+              [4, 5, 6, 7],
+              [8, 9, 10, 11]], dtype=np.float32)
+
+# Define the value of K
+k = 3
+K = np.array([k], dtype=np.int64)
+axis = 1
+new_dims = [X.shape[0], k]
+
+def create_model(op_set_version: int):
+    input_tensors = [helper.make_tensor_value_info('X', TensorProto.FLOAT, X.shape)]
+
+    output_tensors = [
+        helper.make_tensor_value_info('Values', TensorProto.FLOAT, new_dims),
+        helper.make_tensor_value_info('Indices', TensorProto.INT32, new_dims)
+    ]
+
+    # Create the TopK node
+    if op_set_version > 1:
+        node = helper.make_node(
+            'TopK',
+            inputs=['X', 'K'],
+            outputs=['Values', 'Indices'],
+            axis=axis,  # Axis along which to find the top K elements
+        )
+        input_tensors.append(helper.make_tensor_value_info('K', TensorProto.INT32, K.shape))
+    else:
+        node = helper.make_node(
+            'TopK',
+            inputs=['X'],
+            outputs=['Values', 'Indices'],
+            axis=axis,  # Axis along which to find the top K elements
+            k=k
+        )
+
+    # Create the graph
+    graph = helper.make_graph(
+        nodes = [node],
+        name = 'TopKGraph',
+        inputs = input_tensors,
+        outputs = output_tensors,
+        # Unconmment when initializers are supported. Currently we can't test opset 10/11 since the code will require a k value to be initialized for testing.
+        #initializer = [
+        #    helper.make_tensor('X', TensorProto.FLOAT, X.shape, X),
+        #    helper.make_tensor('K', TensorProto.INT64, [1], [k]),
+        #]
+    )
+
+    # Create the model
+    model = helper.make_model(
+        graph,
+        ir_version=8,
+        opset_imports=[onnx.helper.make_operatorsetid("", op_set_version)]
+    )
+    # Check the model
+    onnx.checker.check_model(model)
+
+    # Save the model to a file
+    onnx.save(model, f'top_k_opset_{op_set_version}.onnx')
+    print(f"Model saved to top_k_opset_{op_set_version}.onnx")
+
+def main():
+    # Unconmment when initializers are supported.
+    for op_set_version in [1, 10, 11]:
+        create_model(op_set_version)
+
+
+if __name__ == "__main__":
+    main()

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

@@ -11,7 +11,8 @@ use super::{
     max_pool1d::MaxPool1dNode, max_pool2d::MaxPool2dNode, mean::MeanNode, pad::PadNode,
     prelu::PReluNode, random_normal::RandomNormalNode, random_uniform::RandomUniformNode,
     range::RangeNode, reshape::ReshapeNode, resize::ResizeNode, slice::SliceNode,
-    squeeze::SqueezeNode, sum::SumNode, tile::TileNode, unary::UnaryNode, unsqueeze::UnsqueezeNode,
+    squeeze::SqueezeNode, sum::SumNode, tile::TileNode, top_k::TopKNode, unary::UnaryNode,
+    unsqueeze::UnsqueezeNode,
 };
 use crate::burn::{BurnImports, Scope, Type};
 use burn::backend::NdArray;
@@ -114,6 +115,7 @@ pub enum Node<PS: PrecisionSettings> {
     Squeeze(SqueezeNode),
     Sum(SumNode),
     Tile(TileNode),
+    TopK(TopKNode),
     Unary(UnaryNode),
     Unsqueeze(UnsqueezeNode),
     Where(WhereNode),
@@ -162,6 +164,7 @@ macro_rules! match_all {
             Node::Squeeze(node) => $func(node),
             Node::Sum(node) => $func(node),
             Node::Tile(node) => $func(node),
+            Node::TopK(node) => $func(node),
             Node::Unary(node) => $func(node),
             Node::Unsqueeze(node) => $func(node),
             Node::Where(node) => $func(node),
@@ -218,6 +221,7 @@ impl<PS: PrecisionSettings> Node<PS> {
             Node::Squeeze(_) => "squeeze",
             Node::Sum(_) => "add",
             Node::Tile(_) => "tile",
+            Node::TopK(_) => "top_k",
             Node::Unary(unary) => unary.kind.as_str(),
             Node::Unsqueeze(_) => "unsqueeze",
             Node::Where(_) => "where",

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

@@ -37,6 +37,7 @@ pub(crate) mod slice;
 pub(crate) mod squeeze;
 pub(crate) mod sum;
 pub(crate) mod tile;
+pub(crate) mod top_k;
 pub(crate) mod unary;
 pub(crate) mod unsqueeze;
 pub(crate) use base::*;

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

@@ -0,0 +1,114 @@
+use super::{Node, NodeCodegen};
+use crate::burn::{Scope, TensorType, Type};
+use burn::config::Config;
+use burn::record::PrecisionSettings;
+use proc_macro2::TokenStream;
+use quote::{quote, ToTokens};
+
+#[derive(Config, Debug)]
+pub struct TopKConfig {
+    pub axis: usize,
+    pub k: usize,
+    pub largest: usize,
+}
+
+#[derive(Debug, Clone, new)]
+pub struct TopKNode {
+    pub input: TensorType,
+    pub outputs: Vec<TensorType>,
+    pub config: TopKConfig,
+}
+
+impl<PS: PrecisionSettings> NodeCodegen<PS> for TopKNode {
+    fn output_types(&self) -> Vec<Type> {
+        self.outputs
+            .iter()
+            .map(|t| Type::Tensor(t.clone()))
+            .collect()
+    }
+
+    fn input_types(&self) -> Vec<Type> {
+        vec![Type::Tensor(self.input.clone())]
+    }
+
+    fn forward(&self, scope: &mut Scope, node_position: usize) -> TokenStream {
+        let axis = self.config.axis.to_token_stream();
+        let k = self.config.k.to_token_stream();
+        let largest = self.config.largest.to_token_stream();
+
+        let input = scope.tensor_use_owned(&self.input, node_position);
+        let values_output = &self.outputs[0].name;
+        let indices_output = &self.outputs[1].name;
+
+        quote! {
+            let (#values_output, #indices_output) = #input.topk_with_indices(#k, #axis, #largest);
+        }
+    }
+
+    fn into_node(self) -> Node<PS> {
+        Node::TopK(self)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use burn::record::FullPrecisionSettings;
+
+    use super::*;
+    use crate::burn::{
+        graph::BurnGraph,
+        node::{test::assert_tokens, top_k::TopKNode},
+        TensorType,
+    };
+
+    #[test]
+    fn test_codegen_nodes() {
+        let mut graph = BurnGraph::<FullPrecisionSettings>::default();
+        let config = TopKConfig::new(1, 3, 1);
+
+        graph.register(TopKNode::new(
+            TensorType::new_float("input_tensor", 4),
+            vec![
+                TensorType::new_float("values_tensor", 4),
+                TensorType::new_int("indices_tensor", 4),
+            ],
+            config,
+        ));
+
+        graph.register_input_output(
+            vec!["input_tensor".to_string()],
+            vec!["values_tensor".to_string(), "indices_tensor".to_string()],
+        );
+
+        let expected = quote! {
+            use burn::tensor::Int;
+            use burn::{
+                module::Module,
+                tensor::{backend::Backend, Tensor},
+            };
+
+            #[derive(Module, Debug)]
+            pub struct Model<B: Backend> {
+                phantom: core::marker::PhantomData<B>,
+                device: burn::module::Ignored<B::Device>,
+            }
+
+            impl<B: Backend> Model <B> {
+                #[allow(unused_variables)]
+                pub fn new(device: &B::Device) -> Self {
+                    Self {
+                        phantom: core::marker::PhantomData,
+                        device: burn::module::Ignored(device.clone()),
+                    }
+                }
+                #[allow(clippy::let_and_return, clippy::approx_constant)]
+                pub fn forward(&self, input_tensor: Tensor<B, 4>) -> (Tensor<B, 4>, Tensor<B, 4, Int>) {
+                    let (values_tensor, indices_tensor) = input_tensor.topk_with_indices(3usize, 1usize, 1usize);
+                    (values_tensor, indices_tensor)
+                }
+            }
+        };
+
+        assert_tokens(graph.codegen(), expected);
+    }
+}

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

@@ -7,9 +7,16 @@ use burn::nn::{
     PaddingConfig2d, PaddingConfig3d,
 };
 
-use crate::burn::node::{expand::ExpandShape, pad::PadConfig, tile::TileConfig};
+use crate::burn::node::{expand::ExpandShape, pad::PadConfig, tile::TileConfig, top_k::TopKConfig};
 use onnx_ir::ir::{ArgType, AttributeValue, Data, ElementType, Node};
 
+/// Extract and convert a given attribute to i64
+fn extract_attr_value_i64(node: &Node, key: &str) -> i64 {
+    let error_msg = format!("Expected the following attribute key: {:?}", key);
+    let value = node.attrs.get(key).expect(&error_msg).clone().into_i64();
+    value
+}
+
 /// Create a Conv1dConfig from the attributes of the node
 pub fn conv1d_config(curr: &Node) -> Conv1dConfig {
     let mut kernel_shape = Vec::new(); // TODO default inferred from weight tensor per spec
@@ -795,6 +802,38 @@ pub fn tile_config(node: &Node) -> TileConfig {
     TileConfig::new(repeat)
 }
 
+/// Create a TopKConfig from the attributes of the node.
+pub fn top_k_config(node: &Node) -> TopKConfig {
+    // extract the shape of the input data tensor
+    let data_tensor = match node.inputs.first().unwrap().clone().ty {
+        ArgType::Tensor(tensor) => tensor,
+        _ => panic!("Only tensor input is valid"),
+    };
+
+    let k = match node.inputs.get(1) {
+        Some(k_tensor) => k_tensor
+            .clone()
+            .value
+            .expect("Expecting K tensor to have a value.")
+            .into_i64s()[0],
+        _ => extract_attr_value_i64(node, "k"),
+    };
+
+    let mut axis: i64 = extract_attr_value_i64(node, "axis");
+
+    // if axis is negative, it is counted from the end
+    if axis < 0 {
+        axis += data_tensor.dim as i64;
+    }
+
+    let largest = match node.attrs.get("largest") {
+        Some(val) => val.clone().into_i64(),
+        _ => 1,
+    };
+
+    TopKConfig::new(axis as usize, k as usize, largest as usize)
+}
+
 /// Create a PadConfig from the attributes of the node
 pub fn pad_config(node: &Node) -> PadConfig {
     fn get_pads_input(node: &Node) -> Vec<i64> {