Temporary fix for TFT analyzers which do not reduce instance dims to work with numpy 1.24

PiperOrigin-RevId: 508126642

Author: zoyahav

tensorflow_transform/analyzer_nodes.py

@@ -433,9 +433,17 @@ def encode_cache(self, accumulator):
     return tf.compat.as_bytes(json.dumps(primitive_accumulator))
 
   def decode_cache(self, encoded_accumulator):
-    return np.array(
-        json.loads(tf.compat.as_text(encoded_accumulator)), dtype=self._dtype
-    )
+    # TODO(b/268341036): Set dtype correctly for combiners for numpy 1.24.
+    try:
+      return np.array(
+          json.loads(tf.compat.as_text(encoded_accumulator)), dtype=self._dtype
+      )
+    except ValueError:
+      if self._dtype != object:
+        return np.array(
+            json.loads(tf.compat.as_text(encoded_accumulator)), dtype=object
+        )
+      raise
 
 
 class AnalyzerDef(nodes.OperationDef, metaclass=abc.ABCMeta):

tensorflow_transform/beam/analyzer_cache_test.py

@@ -81,7 +81,7 @@ def test_validate_dataset_keys(self):
           coder=analyzer_nodes._VocabularyAccumulatorCoder(),
           value=[b'\x8a', 29]),
       dict(
-          testcase_name='_VocabularyAccumulatorCoderClassAccumulator',
+          testcase_name='_WeightedMeanAndVarAccumulatorPerKey',
           coder=analyzer_nodes._VocabularyAccumulatorCoder(),
           value=[
               b'A',
@@ -92,6 +92,19 @@ def test_validate_dataset_keys(self):
                   weight=np.array(0.),
               )
           ]),
+      dict(
+          testcase_name='_WeightedMeanAndVarAccumulatorKeepDims',
+          coder=analyzer_nodes.JsonNumpyCacheCoder(),
+          # TODO(b/268341036): Remove this complication once np 1.24 issue is
+          # fixed.
+          value=analyzer_nodes.JsonNumpyCacheCoder(object).decode_cache(
+              analyzer_nodes.JsonNumpyCacheCoder().encode_cache(
+                  analyzers._WeightedMeanAndVarAccumulator(
+                      count=np.array(0),
+                      mean=np.array([], dtype=np.float64),
+                      variance=np.array([], dtype=np.float64),
+                      weight=np.array(0.0))))
+          ),
       dict(
           testcase_name='_QuantilesAccumulatorCoderClassAccumulator',
           coder=analyzers._QuantilesSketchCacheCoder(),