refactor: Centralizing sigmoid implementation (FX Converter Refactor [2/N]) <Target: converter_reorg_proto> (#1868)

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>

Author: narendasan

Diff for: py/torch_tensorrt/fx/converters/__init__.py

@@ -2,7 +2,6 @@
 import tensorrt as trt
 
 if hasattr(trt, "__version__"):
-    from .activation import *  # noqa: F401 F403
     from .adaptive_avgpool import *  # noqa: F401 F403
     from .add import *  # noqa: F401 F403
     from .batchnorm import *  # noqa: F401 F403

Diff for: py/torch_tensorrt/fx/converters/acc_ops_converters.py

@@ -3192,21 +3192,13 @@ def acc_ops_sigmoid(
     kwargs: Dict[str, Argument],
     name: str,
 ) -> Union[TRTTensor, Sequence[TRTTensor]]:
-    input_val = kwargs["input"]
 
-    if not isinstance(input_val, TRTTensor):
-        raise RuntimeError(
-            f"Sigmoid received input {input_val} that is not part "
-            "of the TensorRT region!"
-        )
-
-    return activation.convert_activation(
+    return activation.sigmoid(
         network,
         target,
         SourceIR.ACC,
         name,
-        trt.ActivationType.SIGMOID,
-        input_val,
+        kwargs["input"],
     )
 
 

Diff for: py/torch_tensorrt/fx/converters/activation.py

@@ -484,3 +484,21 @@ def aten_ops_sym_size(
     )
     set_layer_name(slice_layer, target, "_slice_layer")
     return slice_layer.get_output(0)
+
+
+@tensorrt_converter(torch.ops.aten.sigmoid.default)
+def aten_ops_sigmoid(
+    network: TRTNetwork,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> Union[TRTTensor, Sequence[TRTTensor]]:
+
+    return activation.sigmoid(
+        network,
+        target,
+        SourceIR.ATEN,
+        name,
+        args[0],
+    )

Diff for: py/torch_tensorrt/fx/converters/aten_ops_converters.py

@@ -90,3 +90,30 @@ def relu_dyn_range_fn(dyn_range):
         input_val,
         dyn_range_fn=relu_dyn_range_fn,
     )
+
+
+def sigmoid(
+    network: TRTNetwork,
+    target: Target,
+    source_ir: Optional[SourceIR],
+    name: str,
+    input_val: TRTTensor,
+):
+    operation_type = trt.ActivationType.SIGMOID
+
+    def sigmoid_dyn_range_fn(dyn_range):
+        def sigmoid_fn(x):
+            # TODO: Can this just call torch.nn.functional.sigmoid?
+            return 1 / (1 + np.exp(-x))
+
+        return sigmoid_fn(dyn_range[0]), sigmoid_fn(dyn_range[1])
+
+    return convert_activation(
+        network,
+        target,
+        source_ir,
+        name,
+        operation_type,
+        input_val,
+        dyn_range_fn=sigmoid_dyn_range_fn,
+    )

Diff for: py/torch_tensorrt/fx/converters/impl/activation.py

@@ -22,3 +22,17 @@ def relu(network, submod, args, kwargs, layer_name):
         name=layer_name,
         input_val=kwargs["input"],
     )
+
+
+@tensorrt_converter(torch.nn.modules.activation.Sigmoid)
+def sigmoid(network, submod, args, kwargs, layer_name):
+    # args/kwargs should have already been normalized to kwargs
+    assert len(args) == 0
+
+    activation.sigmoid(
+        network=network,
+        target="torch.nn.modules.activation.Sigmoid",
+        source_ir=SourceIR.NN,
+        name=layer_name,
+        input_val=kwargs["input"],
+    )

Diff for: py/torch_tensorrt/fx/converters/nn_ops_converters.py

@@ -0,0 +1,67 @@
+import torch
+import torch.nn as nn
+from torch.testing._internal.common_utils import run_tests
+from torch_tensorrt.fx.utils import LowerPrecision
+from torch_tensorrt.fx.tools.common_fx2trt import DispatchTestCase, InputTensorSpec
+
+
+class TestSigmoidConverter(DispatchTestCase):
+    def test_sigmoid(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.sigmoid(x)
+
+        inputs = [torch.randn(1, 10)]
+        self.run_test(
+            TestModule(), inputs, expected_ops={torch.ops.aten.sigmoid.default}
+        )
+
+    def test_sigmoid_with_dynamic_shape(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.sigmoid(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.sigmoid.default}
+        )
+
+    def test_sigmoid_with_dynamic_shape_four_dimensions(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.sigmoid(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.sigmoid.default}
+        )
+
+    def test_sigmoid_fp16(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.sigmoid(x)
+
+        inputs = [torch.randn(1, 10)]
+        self.run_test(
+            TestModule(),
+            inputs,
+            expected_ops={torch.ops.aten.sigmoid.default},
+            precision=torch.half,
+        )
+
+
+if __name__ == "__main__":
+    run_tests()