Adding selu converter

Author: apbose

py/torch_tensorrt/fx/converters/acc_ops_converters.py

@@ -1064,16 +1064,15 @@ def acc_ops_selu(
     kwargs: Dict[str, Argument],
     name: str,
 ) -> Union[TRTTensor, Sequence[TRTTensor]]:
-    input_val = kwargs["input"]
-    operation_type = trt.ActivationType.SELU
-    return activation.convert_activation(
+
+    return activation.selu(
         network,
         target,
         SourceIR.ACC,
         name,
-        operation_type,
-        input_val,
+        kwargs["input"],
     )
+    
 
 
 @tensorrt_converter(acc_ops.softsign)

py/torch_tensorrt/fx/converters/aten_ops_converters.py

@@ -179,6 +179,14 @@ def aten_ops_elu(
     name: str,
 ) -> Union[TRTTensor, Sequence[TRTTensor]]:
 
+    if (len(args) > 2) :
+        return activation.selu(
+            network,
+            target,
+            SourceIR.ATEN,
+            name,
+            args[0],
+        )
     return activation.elu(
         network,
         target,

py/torch_tensorrt/fx/converters/impl/activation.py

@@ -114,4 +114,28 @@ def elu_dyn_range_fn(dyn_range):
         input_val,
         alpha,
         dyn_range_fn=elu_dyn_range_fn,
+    )
+
+
+def selu(
+    network: TRTNetwork,
+    target: Target,
+    source_ir: Optional[SourceIR],
+    name: str,
+    input_val: TRTTensor,
+    alpha: Optional[Any],
+):
+    operation_type = trt.ActivationType.SELU
+
+    def elu_dyn_range_fn(dyn_range):
+        return (torch.nn.SELU(dyn_range[0]), torch.nn.ELU(dyn_range[1]))
+
+    return convert_activation(
+        network,
+        target,
+        source_ir,
+        name,
+        operation_type,
+        input_val,
+        dyn_range_fn=elu_dyn_range_fn,
     )

py/torch_tensorrt/fx/converters/nn_ops_converters.py

@@ -26,15 +26,30 @@ def relu(network, submod, args, kwargs, layer_name):
 
 @tensorrt_converter(torch.nn.functional.elu)
 @tensorrt_converter(torch.nn.modules.activation.ELU)
-def relu(network, submod, args, kwargs, layer_name):
+def elu(network, submod, args, kwargs, layer_name):
     # args/kwargs should have already been normalized to kwargs
     assert len(args) == 0
 
-    return activation.relu(
+    return activation.elu(
         network=network,
         target="torch.nn.functional.elu",
         source_ir=SourceIR.NN,
         name=layer_name,
         input_val=kwargs["input"],
     )
 
+
+@tensorrt_converter(torch.nn.functional.selu)
+@tensorrt_converter(torch.nn.modules.activation.SELU)
+def selu(network, submod, args, kwargs, layer_name):
+    # args/kwargs should have already been normalized to kwargs
+    assert len(args) == 0
+
+    return activation.selu(
+        network=network,
+        target="torch.nn.functional.selu",
+        source_ir=SourceIR.NN,
+        name=layer_name,
+        input_val=kwargs["input"],
+        alpha = kwargs["alpha"]
+    )

py/torch_tensorrt/fx/test/converters/aten_op/test_selu_aten.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+from torch.testing._internal.common_utils import run_tests
+from torch_tensorrt.fx.tools.common_fx2trt import DispatchTestCase, InputTensorSpec
+
+
+class TestSeLUConverter(DispatchTestCase):
+    def test_selu(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.selu(x)
+
+        inputs = [torch.randn(1, 10)]
+        self.run_test(TestModule(), inputs, expected_ops={torch.ops.aten.elu.default})
+
+    def test_selu_with_dynamic_shape(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.selu(x)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1), (1, 2, 3), (3, 3, 3))],
+            ),
+        ]
+        self.run_test_with_dynamic_shape(
+            TestModule(), input_specs, expected_ops={torch.ops.aten.elu.default}
+        )
+
+    def test_selu_with_dynamic_shape_four_dimensions(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.selu(x)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 5), (1, 2, 3, 5), (3, 3, 3, 5))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            TestModule(), input_specs, expected_ops={torch.ops.aten.elu.default}
+        )
+
+
+if __name__ == "__main__":
+    run_tests()