tracel-ai · antimora · Nov 22, 2023 · Nov 22, 2023 · Nov 22, 2023 · Nov 22, 2023
diff --git a/burn-import/SUPPORTED-ONNX-OPS.md b/burn-import/SUPPORTED-ONNX-OPS.md
@@ -178,7 +178,7 @@ represent the corresponding Burn Op.
 | [SpaceToDepth][172]              |       ❌       |      ❌      |
 | [Split][173]                     |       ❌       |      ❌      |
 | [SplitToSequence][174]           |       ❌       |      ❌      |
-| [Sqrt][175]                      |       ❌       |      ✅      |
+| [Sqrt][175]                      |       ✅       |      ✅      |
 | [Squeeze][176]                   |       ❌       |      ✅      |
 | [STFT][177]                      |       ❌       |      ❌      |
 | [StringNormalizer][178]          |       ❌       |      ❌      |

diff --git a/burn-import/onnx-tests/README.md b/burn-import/onnx-tests/README.md
@@ -16,6 +16,10 @@ Here is the directory structure of this crate:
 - `build.rs`: This build script generates the ONNX models and is executed by `cargo test` before
   running the actual tests.
 
+## Setting up your python environment
+
+You need to install `onnx==1.15.0` and `torch-2.1.1` in your python environment to add a new test
+
 ## Adding new tests
 
 Here are the steps to add a new test:

diff --git a/burn-import/onnx-tests/build.rs b/burn-import/onnx-tests/build.rs
@@ -32,6 +32,7 @@ fn main() {
         .input("tests/reshape/reshape.onnx")
         .input("tests/sigmoid/sigmoid.onnx")
         .input("tests/softmax/softmax.onnx")
+        .input("tests/sqrt/sqrt.onnx")
         .input("tests/sub/sub_int.onnx")
         .input("tests/sub/sub.onnx")
         .input("tests/tanh/tanh.onnx")

diff --git a/burn-import/onnx-tests/tests/onnx_tests.rs b/burn-import/onnx-tests/tests/onnx_tests.rs
@@ -39,6 +39,7 @@
     reshape,
     sigmoid,
     softmax,
+    sqrt,
     sub_int,
     sub,
     tanh,
@@ -326,6 +327,17 @@
         assert_eq!(output.to_data(), expected);
     }
 
+    #[test]
+    fn sqrt() {
+        let model: sqrt::Model<Backend> = sqrt::Model::new();
+
+        let input = Tensor::<Backend, 2>::from_floats([[0.0, 4.0, 9.0], [25.0, 36.0, 49.0]]);
+        let output = model.forward(input);
+        let expected = Data::from([[0.0, 2.0, 3.0], [5.0, 6.0, 7.0]]);
+
+        assert_eq!(output.to_data(), expected);
+    }
+
     #[test]
     fn maxpool2d() {
         // Initialize the model without weights (because the exported file does not contain them)

diff --git a/burn-import/onnx-tests/tests/sqrt/sqrt.onnx b/burn-import/onnx-tests/tests/sqrt/sqrt.onnx
@@ -0,0 +1,16 @@
+pytorch2.1.1:q
+
+onnx::Sqrt_01/Sqrt"Sqrt
+main_graphZ&
+onnx::Sqrt_0
+
+
+
+	
+b
+1
+
+
+
+	
+B
diff --git a/burn-import/onnx-tests/tests/sqrt/sqrt.py b/burn-import/onnx-tests/tests/sqrt/sqrt.py
@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+
+# used to generate model: onnx-tests/tests/sqrt/sqrt.onnx
+
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, x):
+        return torch.sqrt(x)
+
+
+def main():
+    # Set random seed for reproducibility
+    torch.manual_seed(0)
+
+    # Export to onnx
+    model = Model()
+    model.eval()
+    device = torch.device("cpu")
+    onnx_name = "sqrt.onnx"
+    dummy_input = torch.randn(1, 4, 9, 25, device=device)
+
+    torch.onnx.export(model, (dummy_input), onnx_name,
+                      verbose=False, opset_version=16)
+
+    print("Finished exporting model to {}".format(onnx_name))
+
+    # Output some test data for use in the test
+    test_input = torch.tensor([[[[1.0, 4.0, 9.0, 25.0]]]])
+
+    print("Test input data: {}".format(test_input))
+    output = model.forward(test_input)
+    print("Test output data: {}".format(output))
+
+
+if __name__ == '__main__':
+    main()
diff --git a/burn-import/src/burn/node/unary.rs b/burn-import/src/burn/node/unary.rs
@@ -28,6 +28,7 @@ pub enum UnaryNodeKind {
     Relu,
     Reciprocal,
     Sigmoid,
+    Sqrt,
     Tanh,
     Transpose,
 }
@@ -43,6 +44,7 @@ impl UnaryNodeKind {
             Self::Relu => "relu",
             Self::Reciprocal => "reciprocal",
             Self::Sigmoid => "sigmoid",
+            Self::Sqrt => "sqrt",
             Self::Tanh => "tanh",
             Self::Transpose => "transpose",
         }
@@ -133,6 +135,11 @@ impl UnaryNode {
         Self::new(input, output, UnaryNodeKind::Softmax, Rc::new(function))
     }
 
+    pub(crate) fn sqrt(input: Type, output: Type) -> Self {
+        let function = move |input| quote! { burn::tensor::activation::sqrt(#input)};
+        Self::new(input, output, UnaryNodeKind::Sqrt, Rc::new(function))
+    }
+
     pub(crate) fn tanh(input: Type, output: Type) -> Self {
         let function = move |input| quote! { burn::tensor::activation::tanh(#input)};
         Self::new(input, output, UnaryNodeKind::Tanh, Rc::new(function))

diff --git a/burn-import/src/onnx/to_burn.rs b/burn-import/src/onnx/to_burn.rs
@@ -247,6 +247,7 @@ impl ONNXGraph {
                 NodeType::GatherElements => graph.register(Self::gather_conversion(node)),
                 NodeType::LogSoftmax => graph.register(Self::log_softmax_conversion(node)),
                 NodeType::Softmax => graph.register(Self::softmax_conversion(node)),
+                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::Reshape => graph.register(Self::reshape_conversion(node)),
@@ -471,6 +472,13 @@ impl ONNXGraph {
         UnaryNode::softmax(input, output, dim)
     }
 
+    fn sqrt_conversion(node: Node) -> UnaryNode {
+        let input = node.inputs.get(0).unwrap().to_type();
+        let output = node.outputs.get(0).unwrap().to_type();
+
+        UnaryNode::sqrt(input, output)
+    }
+
     fn tanh_conversion(node: Node) -> UnaryNode {
         let input = node.inputs.get(0).unwrap().to_type();
         let output = node.outputs.get(0).unwrap().to_type();