ai-edge-quantizer-nightly 0.0.1.dev20241122__py3-none-any.whl → 0.0.1.dev20241124__py3-none-any.whl

ai_edge_quantizer/transformations/quantize_tensor.py

@@ -22,19 +22,19 @@ from ai_edge_quantizer.transformations import transformation_utils
 from ai_edge_litert import schema_py_generated  # pylint: disable=g-direct-tensorflow-import
 
 
-# TODO: b/335014051 - support distinguishing INT, FLOAT & UINT, BFLOAT
+# TODO: b/335014051 - Support distinguishing INT, FLOAT & UINT, BFLOAT.
 def quant_params_to_tflite_type(
     bitwidth: int,
 ) -> Optional[schema_py_generated.TensorType]:
-  """Given specifications from quant param return the corresponding tflite dtype.
+  """Given specifications from quant param return the corresponding TFLite dtype.
 
   Args:
-    bitwidth: bitwidth from UniformQuantParams
+    bitwidth: Bit width from UniformQuantParams.
 
   Returns:
-    the corresponding tflite tensortype
+    The corresponding TFLite tensor type.
   """
-  if bitwidth <= 4:
+  if bitwidth == 4:
     return schema_py_generated.TensorType.INT4
   elif bitwidth <= 8:
     return schema_py_generated.TensorType.INT8
@@ -68,19 +68,19 @@ def nonlinear_quant_params_to_tflite_type(
 
 
 def _pack_data(bitwidth: int, flattened_data: np.ndarray) -> np.ndarray:
-  """Pack the data to the corresponding bitwidth.
+  """Pack the data to the corresponding bit width.
 
-  If no packing is needed, the original data is returned. Any bitwidth equal or
-  less than 4 bits will be packed to 4 bits.
+  Currently only support 4 bits. If no packing is needed, the original data is
+  returned.
 
   Args:
-    bitwidth: Bitwidth from NonLinearQuantParams.
+    bitwidth: Bit width from NonLinearQuantParams.
     flattened_data: The data to be packed.
 
   Returns:
     Packed data.
   """
-  if bitwidth <= 4:
+  if bitwidth == 4:
     even_data = flattened_data[::2] & 0x0F
     odd_data = np.left_shift(flattened_data[1::2], 4).astype(np.uint8)
     if odd_data.shape[0] == even_data.shape[0] - 1:

ai_edge_quantizer/transformations/quantize_tensor_test.py

@@ -18,6 +18,7 @@
 import os
 import numpy as np
 from tensorflow.python.platform import googletest
+from absl.testing import parameterized
 from ai_edge_quantizer import qtyping
 from ai_edge_quantizer.transformations import quantize_tensor
 from ai_edge_quantizer.transformations import transformation_utils
@@ -28,7 +29,7 @@ from ai_edge_litert import schema_py_generated  # pylint: disable=g-direct-tenso
 TEST_DATA_PREFIX_PATH = test_utils.get_path_to_datafile("..")
 
 
-class QuantizeTensorTest(googletest.TestCase):
+class QuantizeTensorTest(parameterized.TestCase):
 
   def setUp(self):
     super().setUp()
@@ -179,40 +180,44 @@ class QuantizeTensorTest(googletest.TestCase):
     np.testing.assert_array_equal(quant_param.zeroPoint, [1])
     self.assertEqual(quant_param.quantizedDimension, 0)
 
-  def test_int5_constant_not_packed(self):
+  @parameterized.named_parameters(
+      dict(
+          testcase_name="int5",
+          num_bits=5,
+      ),
+      dict(
+          testcase_name="int2",
+          num_bits=2,
+      ),
+  )
+  def test_int_constant_not_packed(self, num_bits):
     subgraph = self._model.subgraphs[0]
     model = self._model
-    data = np.array(
-        [
-            0x0,
-            0x1,
-            0x2,
-            0x3,
-            0x4,
-            0x5,
-            0x6,
-            0x7,
-        ],
-        dtype=np.int8,
-    )
+    tensor_id = 7
+    data = np.array([0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7], dtype=np.int8)
     expected = np.array([0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7])
     ret = quantize_tensor.quantize_tensor(
         transformation_utils.TransformationInput(
-            tensor_id=7,
+            tensor_id=tensor_id,
             op_codes=model.operatorCodes,
             buffers=model.buffers,
             subgraph=subgraph,
             producer=-1,
             consumers=[4],
             quant_params=qtyping.UniformQuantParams(
-                5, None, np.ones(1), np.ones(1), True, data
+                num_bits=num_bits,
+                quantized_dimension=None,
+                scale=np.ones(1),
+                zero_point=np.ones(1),
+                symmetric=True,
+                quantized_data=data,
             ),
         )
     )
     self.assertEqual(ret.op_id, 0)
     self.assertEqual(ret.num_ops_added, 0)
     np.testing.assert_array_equal(model.buffers[8].data, expected)
-    quant_param = subgraph.tensors[7].quantization
+    quant_param = subgraph.tensors[tensor_id].quantization
     np.testing.assert_array_equal(quant_param.scale, [1])
     np.testing.assert_array_equal(quant_param.zeroPoint, [1])
     self.assertEqual(quant_param.quantizedDimension, 0)

{ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info → ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-quantizer-nightly
-Version: 0.0.1.dev20241122
+Version: 0.0.1.dev20241124
 Summary: A quantizer for advanced developers to quantize converted AI Edge models.
 Home-page: https://github.com/google-ai-edge/ai-edge-quantizer
 Keywords: On-Device ML,AI,Google,TFLite,Quantization,LLMs,GenAI

{ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info → ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info}/RECORD

@@ -42,8 +42,8 @@ ai_edge_quantizer/transformations/emulated_subchannel.py,sha256=HVaRxoC8PCAvy3xe
 ai_edge_quantizer/transformations/emulated_subchannel_test.py,sha256=gZP6u9NdPXl7s19qB_Un8evou9ZZV6I9Gy0E1rdobHM,7722
 ai_edge_quantizer/transformations/quant_insert.py,sha256=jn6HsJaV-sqBiFPY-Aqbd64t8zgcYVkEkZI375x_FWY,3958
 ai_edge_quantizer/transformations/quant_insert_test.py,sha256=X9ptPDvJCFkR5tejKnD1SlHFGPazQTW-wNNMV9MEAuw,10107
-ai_edge_quantizer/transformations/quantize_tensor.py,sha256=6lLJHA0G7tf9nrydnSPnWj1rYRN17dH-x5aCkEy3YDQ,5464
-ai_edge_quantizer/transformations/quantize_tensor_test.py,sha256=QAyV3IrvCc9puIWdDz-iONNCuKob7ZejgTbSvT5K3YA,7335
+ai_edge_quantizer/transformations/quantize_tensor.py,sha256=KsJbvhoyBu3D1G5R4nkl54w0TbdYPyit6JfABwlvtbw,5437
+ai_edge_quantizer/transformations/quantize_tensor_test.py,sha256=xfbVNdMbvfJXQcl0vPtmyqKhifVxNZlhu_Xq7RLL2NI,7638
 ai_edge_quantizer/transformations/transformation_utils.py,sha256=BaKy5LYWgqli62XGo3AGRDNtHjwpBNp5VF5XgFbfVmg,4298
 ai_edge_quantizer/transformations/transformation_utils_test.py,sha256=ks81nNvruOC88Tjdk3_qwku0V8p54p3gOqfObzNhWMM,5371
 ai_edge_quantizer/utils/__init__.py,sha256=lpq1g2ayg3lCPLy79t2VicYcnGKw64FfYIj1V7J-4m8,676
@@ -56,8 +56,8 @@ ai_edge_quantizer/utils/tfl_interpreter_utils.py,sha256=GzrsaL3fkOXN5iPRJv7lqhNI
 ai_edge_quantizer/utils/tfl_interpreter_utils_test.py,sha256=Op3JxtOqlrjzmYF18jnnstL1k9xiY9kKJ8S2vklKGkc,11327
 ai_edge_quantizer/utils/validation_utils.py,sha256=oYw33Sg547AqtGw-choPUJmp9SAKkV46J_ddqSsum2Q,3950
 ai_edge_quantizer/utils/validation_utils_test.py,sha256=V_qNDikPD4OPB-siOLQCWNVWTAu87h2IgNYt7teFd-o,2934
-ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info/METADATA,sha256=SPsCOtbr7DB1RDOdpKSMifClBKnsnlrreRe9-iQGZ9s,1484
-ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info/WHEEL,sha256=bFJAMchF8aTQGUgMZzHJyDDMPTO3ToJ7x23SLJa1SVo,92
-ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
-ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info/RECORD,,
+ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info/METADATA,sha256=vwBpgUIxXbDMs0CweYWXOgsDLAi_DXSDc_c7pVl7RCM,1484
+ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
+ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info/RECORD,,

{ai_edge_quantizer_nightly-0.0.1.dev20241122.dist-info → ai_edge_quantizer_nightly-0.0.1.dev20241124.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: bdist_wheel (0.45.0)
+Generator: bdist_wheel (0.45.1)
 Root-Is-Purelib: true
 Tag: py3-none-any