Support quantization of layers with a conv/matmul followed by an activation, with no bias or batch norm between the conv/matmul and the activation.

PiperOrigin-RevId: 207621304

2 files changed, 29 insertions, 1 deletions

diff --git a/tensorflow/contrib/quantize/python/quantize.py b/tensorflow/contrib/quantize/python/quantize.py
index 903faeff11..cb66fd1f76 100644
--- a/tensorflow/contrib/quantize/python/quantize.py
+++ b/tensorflow/contrib/quantize/python/quantize.py
@@ -198,7 +198,7 @@ def _FindLayersToQuantize(graph):
             |
     [post_conv_correction]
             |
-     biasadd|folded_bias
+     [biasadd|folded_bias]
             |
          [bypass]
             |
@@ -320,6 +320,7 @@ def _FindLayersToQuantize(graph):
               folded_bias_add_pattern,
               batch_norm_identity,
               bypass_pattern,
+              layer_pattern,
           ])
       ])
 
diff --git a/tensorflow/contrib/quantize/python/quantize_test.py b/tensorflow/contrib/quantize/python/quantize_test.py
index 98209fffb9..06ebcdfee1 100644
--- a/tensorflow/contrib/quantize/python/quantize_test.py
+++ b/tensorflow/contrib/quantize/python/quantize_test.py
@@ -194,6 +194,33 @@ class QuantizeTest(test_util.TensorFlowTestCase):
 
         self.assertNotIn('test/relu6', [c.name for c in consumers])
 
+  def testLayerActivationQuantized(self):
+    self._RunTestOverParameters(self._TestLayerActivationQuantized)
+
+  def _TestLayerActivationQuantized(self, is_training):
+    graph = ops.Graph()
+    with graph.as_default():
+      batch_size, height, width, depth = 5, 128, 128, 3
+      input1 = array_ops.zeros((batch_size, height, width, depth))
+      _ = conv2d(
+          input1,
+          32, [5, 5],
+          stride=2,
+          padding='SAME',
+          weights_initializer=self._WeightInit(0.09),
+          activation_fn=nn_ops.relu6,
+          biases_initializer=None,
+          scope='test')
+      # Ensure that both weights and output of activations are quantized
+      # when we have a conv->relu6 with no bias add
+      quantize.Quantize(graph, is_training, weight_bits=8, activation_bits=8)
+      activation_op = graph.get_operation_by_name('test/Relu6')
+      conv_op = graph.get_operation_by_name('test/Conv2D')
+      self.assertTrue('test/weights_quant/FakeQuantWithMinMaxVars:0' in
+                      [tensor_in.name for tensor_in in conv_op.inputs])
+      self.assertTrue('FakeQuantWithMinMaxVars' in
+                      [op.type for op in activation_op.outputs[0].consumers()])
+
   def testFinalLayerQuantized(self):
     self._RunTestOverParameters(self._TestFinalLayerQuantized)