Add subtract tensor from scalar for ONNX sub op (#1964)

.gitignore

@@ -9,3 +9,6 @@ target
 .vs
 .fleet
 .ipynb_checkpoints/
+
+# Generated IR and Burn Graph from ONNX
+out

crates/burn-import/DEVELOPMENT.md

@@ -62,7 +62,7 @@ To extend `burn-import` with support for new ONNX operators, follow these steps:
    the Burn model in Rust code, and `my-model.json` includes the model data.
 
 7. **Add End-to-End Test**: Include the test in `./burn-import/onnx-tests/tests/onnx_tests.rs`.
-   Further details can be found in the [onnx-tests README](./burn-import/onnx-tests/README.md).
+   Further details can be found in the [onnx-tests README](./onnx-tests/README.md).
 
 ## Testing
 

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

@@ -147,7 +147,7 @@ mod tests {
         let input = Tensor::<Backend, 4>::from_floats([[[[1., 2., 3., 4.]]]], &device);
         let scalar = 3.0f64;
         let output = model.forward(input, scalar);
-        let expected = TensorData::from([[[[6f32, 7., 8., 9.]]]]);
+        let expected = TensorData::from([[[[-12f32, -13., -14., -15.]]]]);
 
         output.to_data().assert_eq(&expected, true);
     }
@@ -162,7 +162,7 @@ mod tests {
         let input = Tensor::<Backend, 4, Int>::from_ints([[[[1, 2, 3, 4]]]], &device);
         let scalar = 3;
         let output = model.forward(input, scalar);
-        let expected = TensorData::from([[[[6i64, 6, 6, 6]]]]);
+        let expected = TensorData::from([[[[-12i64, -12, -12, -12]]]]);
 
         output.to_data().assert_eq(&expected, true);
     }

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

@@ -26,6 +26,9 @@ class Model(nn.Module):
         # Sutract a scalar from a tensor
         x = x - d
 
+        # Sutract a tensor from a scalar
+        x = d - x
+
         return x
 
 
@@ -40,8 +43,9 @@ def main():
 
     scalar = 3.0
 
-    torch.onnx.export(model, (dummy_input, scalar), onnx_name,
-                      verbose=False, opset_version=16)
+    torch.onnx.export(
+        model, (dummy_input, scalar), onnx_name, verbose=False, opset_version=16
+    )
 
     print("Finished exporting model to {}".format(onnx_name))
 
@@ -53,5 +57,5 @@ def main():
     print("Test output data: {}".format(output))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

crates/burn-import/onnx-tests/tests/sub/sub_int.py

@@ -27,6 +27,9 @@ class Model(nn.Module):
         # Sutract a scalar from a tensor
         x = x - d
 
+        # Sutract a tensor from a scalar
+        x = d - x
+
         return x
 
 
@@ -41,8 +44,9 @@ def main():
     test_input = torch.tensor([[[[1, 2, 3, 4]]]], device=device)
     scalar = 3
 
-    torch.onnx.export(model, (test_input, scalar), onnx_name,
-                      verbose=False, opset_version=16)
+    torch.onnx.export(
+        model, (test_input, scalar), onnx_name, verbose=False, opset_version=16
+    )
 
     print("Finished exporting model to {}".format(onnx_name))
 
@@ -51,5 +55,5 @@ def main():
     print("Test output data: {}".format(output))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

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

@@ -131,6 +131,7 @@ impl BinaryNode {
             (Type::Tensor(_), Type::Tensor(_)) => move |lhs, rhs| quote! { #lhs.sub(#rhs) },
             (Type::Tensor(_), Type::Scalar(_)) => move |lhs, rhs| quote! { #lhs.sub_scalar(#rhs) },
             (Type::Scalar(_), Type::Scalar(_)) => move |lhs, rhs| quote! { #lhs - #rhs },
+            (Type::Scalar(_), Type::Tensor(_)) => move |lhs, rhs| quote! { -#rhs.sub_scalar(#lhs) },
             _ => panic!("Subtraction is supported for tensor and scalar only"),
         };
 

crates/onnx-ir/src/dim_inference.rs

@@ -69,7 +69,7 @@ pub fn dim_inference(node: &mut Node) {
         NodeType::Slice => slice_update_outputs(node),
         NodeType::Softmax => same_as_input(node),
         NodeType::Sqrt => same_as_input(node),
-        NodeType::Sub => same_as_input(node),
+        NodeType::Sub => sub_update_outputs(node),
         NodeType::Sum => same_as_input(node),
         NodeType::Tanh => same_as_input(node),
         NodeType::Transpose => same_as_input(node),
@@ -481,6 +481,20 @@ fn slice_update_outputs(node: &mut Node) {
     }
 }
 
+fn sub_update_outputs(node: &mut Node) {
+    node.outputs[0].ty = match (node.inputs[0].ty.clone(), node.inputs[1].ty.clone()) {
+        (ArgType::Scalar(_lhs), ArgType::Scalar(rhs)) => ArgType::Scalar(rhs),
+        (ArgType::Scalar(_lhs), ArgType::Tensor(rhs)) => ArgType::Tensor(rhs),
+        (ArgType::Tensor(lhs), ArgType::Scalar(_rhs)) => ArgType::Tensor(lhs),
+        // Support broadcasting for lhs/rhs
+        (ArgType::Tensor(lhs), ArgType::Tensor(rhs)) if lhs.dim > rhs.dim => ArgType::Tensor(lhs),
+        (ArgType::Tensor(lhs), ArgType::Tensor(rhs)) if lhs.dim <= rhs.dim => ArgType::Tensor(rhs),
+        _ => {
+            panic!("Only tensor-scalar inputs are valid.");
+        }
+    };
+}
+
 /// Update the output tensor dimension based on the "axes" attribute or the second input
 fn unsqueeze_update_output(node: &mut Node) {
     let axes = if node.inputs.len() == 2 {