Fixed a couple of dynamic shape detection issues (#996)

Summary:
Pull Request resolved: #996

This diff fixed a couple of issues for dynamic shape detection:

* handle cases where sample tensors may not have any dynamic dimension

* added two lowering configs to guide dynamic shape detection

  (1) can_last_dim_be_dynamic: specifies if the last dimension can be dynamic
  (2) can_value_one_be_dynamic: specifies if dimension value one is allowed
      to appear at any dynamic dimension

Reviewed By: hl475

Differential Revision: D54831007

fbshipit-source-id: cc17db2dd386748fec5e4d491fc58ae975b16b7f

fx2ait/fx2ait/tensor_spec.py

@@ -470,7 +470,12 @@ def from_input_list_with_batch_size_jagged_tensor(
 
     @classmethod
     # pyre-ignore [2]: Parameter `sample_input` must have a type other than `Any`
-    def find_batch_size_dim(cls, inputs: Any) -> []:
+    def find_batch_size_dim(
+        cls,
+        inputs: Any,
+        can_non_first_dim_be_dynamic: bool = True,
+        can_dim_value_one_be_dynamic: bool = True,
+    ) -> []:
         if isinstance(inputs, torch.Tensor) or len(inputs) <= 1:
             return [0]
         shapes = [i.shape for i in inputs]
@@ -484,7 +489,9 @@ def find_batch_size_dim(cls, inputs: Any) -> []:
             # Dedup shape value for single tensor
             first_dims.add(shape[0])
             seen_dims = set()
-            for i, dim in enumerate(shape):
+            valid_len = len(shape) if can_non_first_dim_be_dynamic else 1
+            for i in range(valid_len):
+                dim = shape[i]
                 if dim not in seen_dims:
                     frequency_map[dim] = frequency_map.get(dim, 0) + 1
                     position_scores[dim] = position_scores.get(dim, 0) + i
@@ -501,7 +508,18 @@ def find_batch_size_dim(cls, inputs: Any) -> []:
                 frequency_map.items(),
                 key=lambda x: (-x[1], position_scores[x[0]]),
             )
-            batch_size = sorted_frequency[0][0]
+            if len(sorted_frequency) > 1:
+                if not can_dim_value_one_be_dynamic and sorted_frequency[0][0] == 1:
+                    # It's often that dim value one indicates a non-dynamic dimension.
+                    # If the user says so, we pick the second most frequent value.
+                    batch_size = sorted_frequency[1][0]
+                else:
+                    batch_size = sorted_frequency[0][0]
+            else:
+                if not can_dim_value_one_be_dynamic and sorted_frequency[0][0] == 1:
+                    batch_size = -1
+                else:
+                    batch_size = sorted_frequency[0][0]
         else:
             # no dims to sort: no batch_size
             batch_size = -1
@@ -511,6 +529,8 @@ def find_batch_size_dim(cls, inputs: Any) -> []:
             # Default batch size dim = -1, indicate no batch_size
             dim = -1
             for index, val in enumerate(i.shape):
+                if not can_non_first_dim_be_dynamic and index > 0:
+                    break
                 if val == batch_size:
                     dim = index
                     break