[Torch] Fix advanced indexing with NoneType index arguments (#13826)

[Torch] Fix advanced indexing with NoneType index

Author: padreofthegame

python/tvm/relay/frontend/pytorch.py

@@ -2330,19 +2330,49 @@ def one_hot(self, inputs, input_types):
 
     def index(self, inputs, input_types):
         data = inputs[0]
+        data_shape = self.infer_type(data).shape
+
+        axes_adv_idx = [i for i, v in enumerate(inputs[1]) if v is not None]
+        axes_rest = [i for i in range(len(data_shape)) if i not in axes_adv_idx]
+
+        # check if the adv_index axes are consecutive
+        # if consecutive, result must be transposed again at the end
+        consecutive = True
+        for curr, nxt in zip(axes_adv_idx[:-1], axes_adv_idx[1:]):
+            if nxt - curr != 1:
+                consecutive = False
+                break
+
         indices_list = []
+        axes_order = axes_adv_idx + axes_rest
 
-        for indices in inputs[1]:
-            if self.infer_type(indices).dtype == "bool":
+        for i in axes_adv_idx:
+            inp = inputs[1][i]
+            if self.infer_type(inp).dtype == "bool":
                 # adv_index does not support a mask as the index tensor (it will treat 0/1 as
                 # an index rather than a flag).
                 # So we use argwhere to turn the mask into indices, which will also take care
                 # of the dynamism in the indexing by mask.
-                indices_list.append(_op.squeeze(_op.transform.argwhere(indices), axis=[1]))
+                indices_list.append(_op.squeeze(_op.transform.argwhere(inp), axis=[1]))
             else:
-                indices_list.append(indices)
+                indices_list.append(inp)
+
+        data_after_adv_index = _op.adv_index([_op.transpose(data, axes=axes_order)] + indices_list)
 
-        return _op.adv_index([data] + indices_list)
+        if consecutive:
+            num_dims = len(self.infer_type(data_after_adv_index).shape)
+            num_new_dims = num_dims - len(axes_rest)
+
+            axes_final_order = list(range(num_dims))
+            axes_final_order = (
+                axes_final_order[num_new_dims : num_new_dims + axes_adv_idx[0]]
+                + axes_final_order[:num_new_dims]
+                + axes_final_order[num_new_dims + axes_adv_idx[0] :]
+            )
+
+            return _op.transpose(data_after_adv_index, axes=axes_final_order)
+        else:
+            return data_after_adv_index
 
     def meshgrid(self, inputs, input_types):
         data = inputs[0]

tests/python/frontend/pytorch/test_forward.py

@@ -4034,6 +4034,41 @@ def forward(self, x):
     input_data = torch.rand(input_shape).float()
     verify_model(Index1().eval(), input_data=input_data)
 
+    class Index2(Module):
+        def forward(self, x):
+            return x[None, [2, 2]]
+
+    input_data = torch.rand(input_shape).float()
+    verify_model(Index2().eval(), input_data=input_data)
+
+    class Index3(Module):
+        def forward(self, x):
+            return x[None, [0, 1, 2], 1, [2, 3, 4]]
+
+    input_data = torch.rand(input_shape).float()
+    verify_model(Index3().eval(), input_data=input_data)
+
+    class Index4(Module):
+        def forward(self, x):
+            return x[None, [0, 0], None, np.array([[0], [1], [2]]), None]
+
+    input_data = torch.rand(input_shape).float()
+    verify_model(Index4().eval(), input_data=input_data)
+
+    class Index5(Module):
+        def forward(self, x):
+            return x[None, None, [0, 0], np.array([[0], [1], [2]]), None]
+
+    input_data = torch.rand(input_shape).float()
+    verify_model(Index5().eval(), input_data=input_data)
+
+    class Index6(Module):
+        def forward(self, x):
+            return x[None, 1, None, [1, 2, 3]]
+
+    input_data = torch.rand(input_shape).float()
+    verify_model(Index6().eval(), input_data=input_data)
+
     def test_fn_bool_mask():
         return lambda data, mask: data[0, mask]