Fix type parse error about AdaptiveMaxPool (apache#15016)

* fix type parse error about max_pool

Fix the bug when the output_size=(3, None).
Crash message: Check failed: (!checked_type.defined()) is false: Expected Array[PrimExpr], but got Array[index 1: relay.Constant]

* add new test case to caputure bug in adaptive_max_pool

python/tvm/relay/frontend/pytorch.py

@@ -1124,6 +1124,10 @@ def func(x):
     def adaptive_max_pool(self, op, inputs, input_types):
         data = inputs[0]
         output_size = inputs[1]
+        for i, item in enumerate(output_size):
+            if isinstance(item, tvm.relay.expr.Constant):
+                # convert Constant to int
+                output_size[i] = item.data.numpy()[()]
         # returns dummy indices too
         return op(data, output_size=output_size), None
 

tests/python/frontend/pytorch/test_forward.py

@@ -886,6 +886,8 @@ def test_forward_adaptive_avgpool():
 
     input_data = torch.rand([1, 3, 5, 6]).float()
     verify_model(torch.nn.AdaptiveAvgPool2d([3, None]).eval(), input_data=input_data)
+    input_data = torch.rand([1, 1, 3, 5, 6]).float()
+    verify_model(torch.nn.AdaptiveAvgPool3d([3, None, None]).eval(), input_data=input_data)
 
 
 @tvm.testing.uses_gpu
@@ -901,6 +903,11 @@ def test_forward_adaptive_maxpool():
     verify_model(torch.nn.AdaptiveMaxPool1d([1]).eval(), input_data=input_data)
     verify_model(torch.nn.AdaptiveMaxPool1d([5]).eval(), input_data=input_data)
 
+    input_data = torch.rand([1, 3, 5, 6]).float()
+    verify_model(torch.nn.AdaptiveMaxPool2d([3, None]).eval(), input_data=input_data)
+    input_data = torch.rand([1, 1, 3, 5, 6]).float()
+    verify_model(torch.nn.AdaptiveMaxPool3d([3, None, None]).eval(), input_data=input_data)
+
 
 @tvm.testing.uses_gpu
 def test_forward_maxpool2d():