[cherry-pick] Fix the bug of exporting model in dygraph QAT

python/paddle/fluid/contrib/slim/quantization/imperative/qat.py

@@ -72,7 +72,8 @@ def __init__(self,
                  weight_preprocess_layer=None,
                  act_preprocess_layer=None,
                  weight_quantize_layer=None,
-                 act_quantize_layer=None):
+                 act_quantize_layer=None,
+                 onnx_format=False):
         """
         The constructor for ImperativeQuantAware.
 
@@ -124,6 +125,8 @@ def __init__(self,
                 activation and returns dequantized activation. 
                 If None, will use quantization op defined by 'activation_quantize_type'.
                 Default is None.
+            onnx_format (bool, optional): Whether to export the quantized model
+                with format of ONNX. Default is False.
 
         Note:
             If user sets attribute 'skip_quant' to a Layer that support dynamic
@@ -224,7 +227,7 @@ def forward(self, inputs):
         self._quantize_inputs = ImperativeQuantizeInputs(**kwargs)
 
         self._quantize_outputs = ImperativeQuantizeOutputs(
-            moving_rate, activation_bits)
+            moving_rate, activation_bits, onnx_format)
 
     def quantize(self, model):
         """
@@ -413,7 +416,7 @@ class ImperativeQuantizeOutputs(object):
     Calculate the output scales for target layers.
     """
 
-    def __init__(self, moving_rate=0.9, activation_bits=8):
+    def __init__(self, moving_rate=0.9, activation_bits=8, onnx_format=False):
         """
         The constructor for ImperativeQuantizeOutputs.
 
@@ -425,6 +428,7 @@ def __init__(self, moving_rate=0.9, activation_bits=8):
         super(ImperativeQuantizeOutputs, self).__init__()
         self._moving_rate = moving_rate
         self._activation_bits = activation_bits
+        self._onnx_format = onnx_format
 
     def apply(self, model):
         """
@@ -461,12 +465,7 @@ def apply(self, model):
 
             setattr(parent_layer, sub_name, cur_quant_layer)
 
-    def save_quantized_model(self,
-                             model,
-                             path,
-                             input_spec=None,
-                             onnx_format=False,
-                             **config):
+    def save_quantized_model(self, model, path, input_spec=None, **config):
         """
         Save the quantized model for the inference.
 
@@ -479,8 +478,6 @@ def save_quantized_model(self,
                 InputSpec or example Tensor. If None, all input variables of 
                 the original Layer's forward method would be the inputs of
                 the saved model. Default None.
-            onnx_format (bool, optional): Whether to export the quantized model 
-                with format of ONNX. Default is False.
             **config (dict, optional): Other save configuration options for
                 compatibility. We do not recommend using these configurations,
                 they may be removed in the future. If not necessary, DO NOT use
@@ -521,7 +518,7 @@ def save_quantized_model(self,
                                    model_filename=model_filename,
                                    params_filename=params_filename))
 
-        if not onnx_format:
+        if not self._onnx_format:
             self._gather_scales(infer_program, scope, fetch_targets)
 
             # Remove `moving_average_abs_max_scale` node in sub graphs.
@@ -540,10 +537,14 @@ def save_quantized_model(self,
             graph = IrGraph(core.Graph(infer_program.desc), for_test=False)
             transform_pass = ReplaceFakeQuantDequantPass(
                 scope, place, quant_bits=self._activation_bits)
-            transform_pass.apply(graph)
+            for sub_graph in graph.all_sub_graphs():
+                sub_graph._for_test = True
+                transform_pass.apply(sub_graph)
 
             quant_weight_pass = QuantWeightPass(scope, place)
-            quant_weight_pass.apply(graph)
+            for sub_graph in graph.all_sub_graphs():
+                sub_graph._for_test = True
+                quant_weight_pass.apply(sub_graph)
 
             infer_program = graph.to_program()
 
@@ -565,18 +566,24 @@ def _is_target_layer(self, layer):
         """
         Whether the layer needs to calculate output scales.
         """
+        # exclude fake_quant ops in quant_layers file
+        if not isinstance(layer, dygraph.Layer):
+            return False
+
+        if self._onnx_format:
+            return True if isinstance(layer, tuple(
+                utils.fake_quant_wrap_layers)) else False
+
         flag = False
-        if isinstance(layer, dygraph.Layer):
-            # exclude fake_quant ops in quant_layers file
-            if utils.is_leaf_layer(layer) and \
-                not isinstance(layer, tuple(utils.fake_quant_leaf_layers)):
-                flag = True
+        if utils.is_leaf_layer(layer) and \
+            not isinstance(layer, tuple(utils.fake_quant_leaf_layers)):
+            flag = True
 
-            if isinstance(layer, tuple(utils.fake_quant_wrap_layers)):
-                flag = True
+        if isinstance(layer, tuple(utils.fake_quant_wrap_layers)):
+            flag = True
 
-            if isinstance(layer, paddle.nn.quant.FloatFunctionalLayer):
-                flag = True
+        if isinstance(layer, paddle.nn.quant.FloatFunctionalLayer):
+            flag = True
 
         return flag
 

python/paddle/fluid/contrib/slim/quantization/quantization_pass.py

@@ -59,6 +59,7 @@
 _fake_quant_dequant_op_list = [
     'fake_quantize_dequantize_moving_average_abs_max',
     "fake_channel_wise_quantize_dequantize_abs_max",
+    "fake_quantize_dequantize_abs_max",
 ]
 
 _conv_ops = ['conv2d', 'depthwise_conv2d', 'conv2d_transpose']

python/paddle/fluid/contrib/slim/quantization/utils.py

@@ -332,9 +332,11 @@ def _clip(x, scale):
         x[x < -scale] = -scale
         return x
 
-    assert quant_axis in [0, 1], 'quant_axis should be 0 or 1 for now.'
     bnt = (1 << (weight_bits - 1)) - 1
+    if isinstance(scale, list) and len(scale) == 1:
+        scale = scale[0]
     if isinstance(scale, list):
+        assert quant_axis in [0, 1], 'quant_axis should be 0 or 1 for now.'
         for i, s in enumerate(scale):
             if s == 0.0:
                 s = 1e-8

python/paddle/fluid/contrib/slim/tests/test_imperative_qat.py

@@ -68,7 +68,8 @@ def func_qat(self):
         imperative_qat = ImperativeQuantAware(
             weight_quantize_type=self.weight_quantize_type,
             activation_quantize_type=self.activation_quantize_type,
-            fuse_conv_bn=self.fuse_conv_bn)
+            fuse_conv_bn=self.fuse_conv_bn,
+            onnx_format=self.onnx_format)
 
         with fluid.dygraph.guard():
             # For CI coverage
@@ -187,8 +188,7 @@ def func_qat(self):
                 input_spec=[
                     paddle.static.InputSpec(shape=[None, 1, 28, 28],
                                             dtype='float32')
-                ],
-                onnx_format=self.onnx_format)
+                ])
             print('Quantized model saved in %s' % tmpdir)
 
             if core.is_compiled_with_cuda():