added support for logsigmoid op

python/tvm/relax/frontend/torch/base_fx_graph_translator.py

@@ -1083,7 +1083,15 @@ def _linear(self, node: fx.Node) -> relax.Var:
         weight = args[1]
         bias = args[2] if len(args) > 2 else None
         return self.block_builder.emit(relax.op.linear(x, weight, bias, "float32"))
-
+
+    def _logsigmoid(self, node: fx.Node) -> relax.Var:
+        x = self.env[node.args[0]]
+        neg_x = self.block_builder.emit(relax.op.negative(x))
+        exp_neg_x = self.block_builder.emit(relax.op.exp(neg_x))
+        add_one = self.block_builder.emit(relax.op.add(exp_neg_x, relax.const(1.0, dtype="float32")))
+        log_val = self.block_builder.emit(relax.op.log(add_one))
+        return self.block_builder.emit(relax.op.negative(log_val))
+
     def _max_pool1d_impl(
         self,
         x: relax.Expr,

python/tvm/relax/frontend/torch/exported_program_translator.py

@@ -354,6 +354,7 @@ def create_convert_map(
             "log1p.default": self._log1p,
             "logical_not.default": self._unary_op(relax.op.logical_not),
             "log_softmax.int": self._log_softmax,
+            "logsigmoid.default": self._logsigmoid,
             "neg.default": self._unary_op(relax.op.negative),
             "pad.default": self._pad,
             "pixel_shuffle.default": self._pixel_shuffle,

tests/python/relax/test_frontend_from_exported_program.py

@@ -734,6 +734,40 @@ def main(
     verify_model(LogSoftmax2(), example_args, {}, expected1)
 
 
+def test_logsigmoid():
+    class LogSigmoid(Module):
+        def __init__(self):
+            super().__init__()
+            self.ls = torch.nn.LogSigmoid()
+
+        def forward(self, input):
+            return self.ls(input)
+
+    class LogSigmoid2(Module):
+        def forward(self, input):
+            return torch.nn.functional.logsigmoid(input)
+
+    @tvm.script.ir_module
+    class expected_logsigmoid:
+        @R.function
+        def main(
+            input: R.Tensor((1, 3, 10, 10), dtype="float32")
+        ) -> R.Tuple(R.Tensor((1, 3, 10, 10), dtype="float32")):
+            with R.dataflow():
+                neg_input = R.negative(input)
+                exp_neg = R.exp(neg_input)
+                add_one = R.add(exp_neg, R.const(1.0, "float32"))
+                log_val = R.log(add_one)
+                result = R.negative(log_val)
+                gv: R.Tuple(R.Tensor((1, 3, 10, 10), dtype="float32")) = (result,)
+                R.output(gv)
+            return gv
+
+    example_args = (torch.randn(1, 3, 10, 10, dtype=torch.float32),)
+    verify_model(LogSigmoid(), example_args, {}, expected_logsigmoid)
+    verify_model(LogSigmoid2(), example_args, {}, expected_logsigmoid)
+
+
 def test_prelu():
     class Prelu1(Module):
         def __init__(self, num_parameters=1, alpha=0.25):