No public description

PiperOrigin-RevId: 726648660
Change-Id: I35836e5b0083e65758f070d9d2886cbfab07f39f

Author: lishanok

qkeras/qtools/qtools_util.py

@@ -248,6 +248,16 @@ def get_weights(layer, model_weights_already_quantized=True):
   return out
 
 
+def get_scale_from_quantized_bits_with_auto_po2(quantizer):
+  """Get scale from quantized_bits with alpha=auto_po2."""
+  if hasattr(quantizer.scale, "numpy"):
+    return quantizer.scale.numpy()
+  elif isinstance(quantizer.scale, np.ndarray):
+    return quantizer.scale
+  else:
+    return None
+
+
 def adjust_multiplier_for_auto_po2(multiplier, qkeras_weight_quantizer):
   """Adjust multiplier when weight quantizer is auto_po2 type.
 
@@ -267,11 +277,9 @@ def adjust_multiplier_for_auto_po2(multiplier, qkeras_weight_quantizer):
       qkeras_weight_quantizer.alpha == "auto_po2"):
     bits = output_quantizer.bits
     int_bits = output_quantizer.int_bits
-    scale = qkeras_weight_quantizer.scale
-    if hasattr(scale, "numpy"):
-      # if scale doesn't have numpy() function, it means the quantizer has
-      # never being called. Therfore we skip the following steps
-      scale = scale.numpy()
+    scale = get_scale_from_quantized_bits_with_auto_po2(
+        qkeras_weight_quantizer)
+    if scale is not None:
       if isinstance(scale, np.ndarray):
         scale = np.squeeze(scale)
         max_shift = int(np.log2(np.max(scale)))
@@ -293,6 +301,8 @@ def adjust_multiplier_for_auto_po2(multiplier, qkeras_weight_quantizer):
       output_quantizer.bits = total_bits
       output_quantizer.int_bits = max_int_bits
     else:
+      # If scale is None, it means the quantizer has
+      # never being called. Therfore we skip the bitwidth adjustment steps
       print("[WARNING] The weight quantizer is never called even though it has "
             "alpha=auto_po2. In this case we do not adjust the multiplier and "
             "accumulator bit width since we don't know the exact values of "

tests/qtools_util_test.py

@@ -0,0 +1,95 @@
+# Copyright 2020 Google LLC
+#
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for qtools_util module."""
+
+import json
+
+import numpy as np
+import pytest
+import tensorflow.keras as keras
+import tensorflow as tf
+
+from qkeras import quantizers
+from qkeras.qtools import qtools_util
+
+from qkeras.qtools import quantized_operators
+from qkeras.qtools.quantized_operators import quantizer_factory as quantizer_factory_module
+
+
+@pytest.mark.parametrize(
+    "w_bits, w_int_bits, weight_quantizer_scale_type, "
+    "expected_bits_before_adjustment, expected_int_bits_before_adjustment, "
+    "expected_bits_after_adjustment, expected_int_bits_after_adjustment",
+    [
+        (8, 0, "1.0", 11, 2, 11, 2),
+        (4, 2, "auto_po2", 7, 4, 10, 5),
+        (4, 0, "post_training_scale", 7, 2, 10, 5),
+    ],
+)
+def test_adjust_multiplier_for_auto_po2(
+    w_bits, w_int_bits, weight_quantizer_scale_type,
+    expected_bits_before_adjustment, expected_int_bits_before_adjustment,
+    expected_bits_after_adjustment, expected_int_bits_after_adjustment):
+  """Test adjust_multiplier_for_auto_po2 with auto_po2 weight quantizer."""
+
+  multiplier_factory = quantized_operators.MultiplierFactory()
+  quantizer_factory = quantizer_factory_module.QuantizerFactory()
+
+  qkeras_input_quantizer = quantizers.quantized_bits(4, 2, 1)
+
+  # Generate the weight quantizer.
+  if weight_quantizer_scale_type in ["auto_po2", "post_training_scale"]:
+    # Compute the scale for auto_po2 quantizer.
+    qkeras_weight_quantizer = quantizers.quantized_bits(
+        bits=w_bits, integer=w_int_bits, keep_negative=True,
+        symmetric=True, alpha="auto_po2")
+    weight_arr = np.array([1.07, -1.7, 3.06, 1.93, 0.37, -2.43, 6.3, -2.9]
+                          ).reshape((2, 4))
+    qkeras_weight_quantizer(weight_arr)
+
+    if weight_quantizer_scale_type == "post_training_scale":
+      # Set the post_training_scale as fixed scale.
+      auto_po2_scale = qkeras_weight_quantizer.scale.numpy()
+      qkeras_weight_quantizer = quantizers.quantized_bits(
+          bits=w_bits, integer=w_int_bits, alpha="auto_po2",
+          post_training_scale=auto_po2_scale)
+  else:
+    qkeras_weight_quantizer = quantizers.quantized_bits(w_bits, w_int_bits)
+
+  input_quantizer = quantizer_factory.make_quantizer(
+      qkeras_input_quantizer)
+  weight_quantizer = quantizer_factory.make_quantizer(
+      qkeras_weight_quantizer)
+
+  multiplier = multiplier_factory.make_multiplier(
+      weight_quantizer, input_quantizer)
+
+  np.testing.assert_equal(multiplier.output.bits,
+                          expected_bits_before_adjustment)
+  np.testing.assert_equal(multiplier.output.int_bits,
+                          expected_int_bits_before_adjustment)
+
+  qtools_util.adjust_multiplier_for_auto_po2(
+      multiplier, qkeras_weight_quantizer)
+  print(f"after adjustment: {multiplier.output.bits}, {multiplier.output.int_bits}")
+  np.testing.assert_equal(multiplier.output.bits,
+                          expected_bits_after_adjustment)
+  np.testing.assert_equal(multiplier.output.int_bits,
+                          expected_int_bits_after_adjustment)
+
+
+if __name__ == "__main__":
+  pytest.main([__file__])