From 5993dde3e52bdb540ff443c41cdc29592f0437b5 Mon Sep 17 00:00:00 2001
From: cc <52520497+juncaipeng@users.noreply.github.com>
Date: Wed, 3 Jun 2020 10:48:52 +0800
Subject: [PATCH] [Cherry-pick] Fix some bugs for quantization (#24852)

* Update sigmoid output from Y to out, test=develop (#24765)

* Collecting concat output threshold, test=develop (#24742)

* Add output threshold for ops that have several output activations, test=develop (#24726)

* [Fix bug] Init scale node in OutScaleForTrainingPass and enable test_quantization_scale_pass UT (#24393)

* Init scale node in OutScaleForTrainingPass, test=develop
* Enable test_quantization_scale, test=develop
---
 .../slim/quantization/quantization_pass.py    | 43 +++++++++++--------
 .../fluid/contrib/slim/tests/CMakeLists.txt   |  3 --
 2 files changed, 24 insertions(+), 22 deletions(-)

diff --git a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
index cde41e687f..f04869156f 100644
--- a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
+++ b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
@@ -43,7 +43,7 @@ _fake_quant_dequant_op_list = [
 _out_scale_op_list = [
     "conv2d", "depthwise_conv2d", "mul", "matmul", "relu", "leaky_relu",
     "relu6", "sigmoid", "tanh", "prelu", "swish", "softmax", "batch_norm",
-    "elementwise_add", "pool2d", "reshape2", "transpose2"
+    "elementwise_add", "pool2d", "reshape2", "transpose2", "concat"
 ]
 
 # list op real input and output names, to avoid processing input such as AxisTensor.
@@ -83,7 +83,7 @@ _op_real_in_out_name = {
     "swish": [["X"], ["Out"]],
     "dropout": [["X"], ["Out"]],
     "batch_norm": [["X"], ["Y"]],
-    "sigmoid": [["X"], ["Y"]],
+    "sigmoid": [["X"], ["Out"]],
 }
 
 
@@ -1156,20 +1156,27 @@ class OutScaleForTrainingPass(object):
         assert isinstance(graph,
                           IrGraph), 'graph must be the instance of IrGraph.'
         self._is_test = graph.is_test()
-        ops = graph.all_op_nodes()
-        for op_node in ops:
-            name = op_node.name()
-            if name in self._teller_set:
-                if len(op_node.output_arg_names()) != 1:
-                    continue
-                in_node = graph._find_node_by_name(
-                    op_node.outputs, op_node.output_arg_names()[0])
+        target_ops = []
+        for op in graph.all_op_nodes():
+            if op.name() in self._teller_set:
+                target_ops.append(op)
+        for op in target_ops:
+            for output_var_name in _get_op_output_var_names(op):
+                in_node = graph._find_node_by_name(op.outputs, output_var_name)
                 out_node = graph.create_var_node_from_desc(in_node.var())
                 scale_node = graph.create_persistable_node(
                     name=self._scale_name(in_node.name()),
                     var_type=core.VarDesc.VarType.LOD_TENSOR,
                     shape=[1],
                     var_dtype=in_node.dtype())
+                data_type = 'float64' if in_node.dtype() \
+                    == core.VarDesc.VarType.FP64 else 'float32'
+                _init_var_node(
+                    scale_node,
+                    np.ones(
+                        [1], dtype=data_type),
+                    self._scope,
+                    self._place)
                 ins = {'X': in_node}
                 outs = {'Out': out_node, 'OutScale': scale_node}
                 if not self._is_test:
@@ -1178,8 +1185,6 @@ class OutScaleForTrainingPass(object):
                         var_type=core.VarDesc.VarType.LOD_TENSOR,
                         var_dtype=in_node.dtype(),
                         shape=[1])
-                    data_type = 'float64' if in_node.dtype(
-                    ) == core.VarDesc.VarType.FP64 else 'float32'
                     _init_var_node(
                         state_in_node,
                         np.ones(
@@ -1257,13 +1262,13 @@ class OutScaleForInferencePass(object):
         """
         assert isinstance(graph,
                           IrGraph), 'graph must be the instance of IrGraph.'
-        ops = graph.all_op_nodes()
-        for op_node in ops:
-            name = op_node.name()
-            if name in self._teller_set:
-                if len(op_node.output_arg_names()) != 1:
-                    continue
-                scale_name = self._scale_name(op_node.output_arg_names()[0])
+        op_nodes = graph.all_op_nodes()
+        for op_node in op_nodes:
+            if op_node.name() in self._teller_set:
+                output_var_name = _get_op_output_var_names(op_node)
+                assert len(output_var_name) == 1, "Only support collecting " \
+                    "output for op that only has an activation output for now."
+                scale_name = self._scale_name(output_var_name[0])
                 scale_v = np.array(
                     self._scope.find_var(scale_name).get_tensor())[0]
                 op_node.op()._set_attr("out_threshold", float(scale_v))
diff --git a/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt b/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt
index 737d67051a..5aa05a158d 100644
--- a/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt
+++ b/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt
@@ -114,9 +114,6 @@ if(WIN32)
     list(REMOVE_ITEM TEST_OPS test_weight_quantization_mobilenetv1)
 endif()
 
-# Disable unittest for random error temporary
-list(REMOVE_ITEM TEST_OPS test_quantization_scale_pass)
-
 if(LINUX AND WITH_MKLDNN)
 
 	#### Image classification dataset: ImageNet (small)
-- 
GitLab