QAT: support for new models (#23928) (#24121)

* QAT: support range-based quantization and scales from attribute

* added support for channelwise

test=release/2.0

python/paddle/fluid/contrib/slim/quantization/qat2_int8_mkldnn_pass.py

@@ -45,16 +45,14 @@ class Qat2Int8MkldnnPass(object):
         self._place = _place
         self._core = _core
         self._debug = _debug
-        self._quantize_types = [
+        self._fake_quantize_types = [
             'fake_quantize_moving_average_abs_max',
             'fake_quantize_range_abs_max',
             'fake_quantize_dequantize_moving_average_abs_max'
         ]
-        self._fake_quantize_types = [
-            'fake_quantize_moving_average_abs_max',
-            'fake_quantize_dequantize_moving_average_abs_max'
+        self._fake_dequantize_types = [
+            'fake_dequantize_max_abs', 'fake_channel_wise_dequantize_max_abs'
         ]
-        self._fake_dequantize_types = ['fake_dequantize_max_abs']
         self._quantized_ops = _quantized_ops
         self._scale_immutable_ops = [
             'transpose2', 'reshape2', 'pool2d', 'scale'
@@ -74,7 +72,9 @@ class Qat2Int8MkldnnPass(object):
         assert isinstance(graph,
                           IrGraph), 'graph must be the instance of IrGraph.'
 
-        graph = self._gather_scales(graph)
+        graph = self._gather_weight_scales_from_fake(graph)
+        graph = self._gather_output_scales_from_attr(graph)
+        graph = self._gather_input_scales_from_fake(graph)
         graph = self._remove_fake_ops(graph)
         graph = self._dequantize_weights(graph)
         graph = self._optimize_fp32_graph(graph)
@@ -83,6 +83,7 @@ class Qat2Int8MkldnnPass(object):
         graph = self._propagate_scales(graph)
         graph = self._set_dummy_out_scales(graph)
         graph = self._quantize_fp32_graph(graph)
+        graph = self._optimize_int8_graph(graph)
         graph = self._cleanup(graph)
         return graph
 
@@ -90,9 +91,6 @@ class Qat2Int8MkldnnPass(object):
         assert isinstance(graph,
                           IrGraph), 'graph must be the instance of IrGraph.'
 
-        graph = self._gather_scales(graph)
-        graph = self._remove_fake_ops(graph)
-        graph = self._dequantize_weights(graph)
         graph = self._optimize_fp32_graph(graph)
         graph = self._cleanup(graph)
         return graph
@@ -108,29 +106,61 @@ class Qat2Int8MkldnnPass(object):
     def _is_fc_quantized(self):
         return 'fc' in self._quantized_ops
 
-    def _gather_scales(self, graph):
+    def _gather_input_scales_from_fake(self, graph):
+        def _add_scale_for_vars(var_names, use_unsigned_int, lod_tensor):
+            scales = self._var_quant_scales
+            for var_name in var_names:
+                scales[var_name] = (use_unsigned_int, lod_tensor)
+
         for op in graph.all_op_nodes():
-            if op.name() in self._quantize_types:
+            if op.name() in self._fake_quantize_types:
                 bit_length = op.op().attr("bit_length")
                 assert bit_length == 8, 'Unsupported number quantization bits ({}). Only 8 is supported now.'.format(
                     bit_length)
 
                 input_name = op.input("X")[0]
                 scale_name = op.input("InScale")[0]
+                output_name = op.output("Out")[0]
                 # Gather new weights scale after folding batchnorm in convolution
                 scale = np.array(1.0 / self._load_param(
                     self._scope, scale_name)[0]).astype(np.float64)
                 lod_tensor = self._convert_scale2tensor(scale)
                 use_unsigned_int = False
-                self._var_quant_scales[input_name] = (use_unsigned_int,
-                                                      lod_tensor)
-                self._var_quant_scales[scale_name.replace(".scale", "")] = (
-                    use_unsigned_int, lod_tensor)
+                _add_scale_for_vars([input_name, output_name], use_unsigned_int,
+                                    lod_tensor)
+
+        return graph
 
+    def _gather_weight_scales_from_fake(self, graph):
+        for op in graph.all_op_nodes():
             if op.name() in self._fake_dequantize_types:
                 input_name = op.input("X")[0]
-                _max_range = op.op().attr("max_range")
-                self._weight_scales[input_name] = _max_range
+                if op.op().has_attr("max_range"):
+                    _max_range = np.array(op.op().attr("max_range")).astype(
+                        np.float64)
+                    self._weight_scales[input_name] = _max_range
+                else:
+                    scale_name = op.input("Scales")[0]
+                    scale = np.array(
+                        self._s8_max * self._s8_max / self._load_param(
+                            self._scope, scale_name)).astype(np.float64)
+                    self._weight_scales[input_name] = scale
+
+        return graph
+
+    def _gather_output_scales_from_attr(self, graph):
+        for op in graph.all_op_nodes():
+            if op.op().has_attr("out_threshold"):
+                attr_scale = op.op().attr("out_threshold")
+                if attr_scale == 0.0: continue
+                scale = np.array(1.0 / attr_scale).astype(np.float64)
+                scale_lod_tensor = self._convert_scale2tensor(scale)
+                use_unsigned_int = False
+                for output_name in op.op().outputs():
+                    for out_var_name in op.op().output(output_name):
+                        self._var_quant_scales[out_var_name] = (
+                            use_unsigned_int, scale_lod_tensor)
+
         return graph
 
     def _propagate_scales(self, graph):
@@ -274,29 +304,24 @@ class Qat2Int8MkldnnPass(object):
     def _dequantize_weights(self, graph):
         for op in graph.all_op_nodes():
             if op.name() in self._conv_ops:
-                self._dequantize_conv_weights(graph, op)
+                self._dequantize_op_weights(graph, op, "Filter", "Output")
             elif self._is_fc_quantized() and op.name() in self._mul_ops:
-                self._dequantize_mul_weights(graph, op)
+                self._dequantize_op_weights(graph, op, "Y", "Out")
         return graph
 
-    def _dequantize_conv_weights(self, graph, op_node):
-        weight_name = op_node.input("Filter")[0]
-        output_name = op_node.output("Output")[0]
+    def _dequantize_op_weights(self, graph, op_node, weight_name, output_name):
+        weight_var_name = op_node.input(weight_name)[0]
+        output_var_name = op_node.output(output_name)[0]
         # Convert int8 range weights to fp32 range weights
-        scales = self._weight_scales[output_name]
-        weight = self._load_param(self._scope, weight_name)
-        w_fp32 = np.divide(np.multiply(weight, self._s8_max), scales)
-        w_fp32 = w_fp32.reshape(weight.shape)
-        self._restore_var(weight_name, w_fp32)
-
-    def _dequantize_mul_weights(self, graph, op_node):
-        weight_name = op_node.input("Y")[0]
-        output_name = op_node.output("Out")[0]
-        scales = self._weight_scales[output_name]
-        weight = self._load_param(self._scope, weight_name)
-        w_fp32 = np.divide(np.multiply(weight, self._s8_max), scales)
-        w_fp32 = w_fp32.reshape(weight.shape)
-        self._restore_var(weight_name, w_fp32)
+        scales = self._weight_scales[output_var_name]
+        weight = self._load_param(self._scope, weight_var_name)
+        assert scales.size == 1 or scales.size == len(
+            weight
+        ), "The size of weight scales vector ({}) does not match the number of output channels ({}) in the weights tensor {}.".format(
+            scales.size, len(weight), weight_var_name)
+        w_fp32 = np.divide(np.multiply(weight, self._s8_max).T, scales.T).T
+        w_fp32 = w_fp32.reshape(weight.shape).astype(np.float32)
+        self._restore_var(weight_var_name, w_fp32)
 
     def _restore_var(self, name, array):
         tensor = self._scope.find_var(name).get_tensor()
@@ -366,11 +391,14 @@ class Qat2Int8MkldnnPass(object):
         self._remove_unused_var_nodes(graph)
         return graph
 
-    def _cleanup(self, graph):
+    def _optimize_int8_graph(self, graph):
         # remove dropout ops
         graph = self._apply_pass(graph, 'simplify_with_basic_ops_pass')
         # make some MKL-DNN ops working inplace
         graph = self._apply_pass(graph, 'mkldnn_inplace_pass')
+        return graph
+
+    def _cleanup(self, graph):
         graph = self._remove_unused_var_nodes(graph)
         graph = self._set_op_role_forward(graph)
         return graph

python/paddle/fluid/contrib/slim/tests/CMakeLists.txt

@@ -68,7 +68,7 @@ function(inference_qat2_int8_image_classification_test target qat_model_dir fp32
                  --batch_size 10
                  --batch_num 2
                  --acc_diff_threshold 0.1
-		 --quantized_ops ${quantized_ops})
+                 --quantized_ops ${quantized_ops})
 endfunction()
 
 # set batch_size 10 for UT only (avoid OOM). For whole dataset, use batch_size 20 
@@ -215,13 +215,28 @@ if(LINUX AND WITH_MKLDNN)
 
 	set(QAT2_IC_QUANTIZED_OPS "conv2d,pool2d")
 
-	# QAT2 ResNet50
+	# QAT2 ResNet50 with input/output scales in `fake_quantize_moving_average_abs_max` operators,
+	# with weight scales in `fake_dequantize_max_abs` operators
         set(QAT2_RESNET50_MODEL_DIR "${QAT_INSTALL_DIR}/ResNet50_qat_perf")
 	set(FP32_RESNET50_MODEL_DIR "${INT8_INSTALL_DIR}/resnet50")
 	set(QAT2_RESNET50_MODEL_ARCHIVE "ResNet50_qat_perf.tar.gz")
 	download_qat_model(${QAT2_RESNET50_MODEL_DIR} ${QAT2_RESNET50_MODEL_ARCHIVE})
 	inference_qat2_int8_image_classification_test(test_qat2_int8_resnet50_mkldnn ${QAT2_RESNET50_MODEL_DIR}/ResNet50_qat_perf/float ${FP32_RESNET50_MODEL_DIR}/model ${IMAGENET_DATA_PATH} ${QAT2_IC_QUANTIZED_OPS})
 
+	# QAT2 ResNet50 with input/output scales in `fake_quantize_range_abs_max` operators and the `out_threshold` attributes,
+	# with weight scales in `fake_dequantize_max_abs` operators
+	set(QAT2_RESNET50_RANGE_MODEL_DIR "${QAT_INSTALL_DIR}/ResNet50_qat_range")
+	set(QAT2_RESNET50_RANGE_MODEL_ARCHIVE "ResNet50_qat_range.tar.gz")
+	download_qat_model(${QAT2_RESNET50_RANGE_MODEL_DIR} ${QAT2_RESNET50_RANGE_MODEL_ARCHIVE})
+	inference_qat2_int8_image_classification_test(test_qat2_int8_resnet50_range_mkldnn ${QAT2_RESNET50_RANGE_MODEL_DIR}/ResNet50_qat_range ${FP32_RESNET50_MODEL_DIR}/model ${IMAGENET_DATA_PATH} ${QAT2_IC_QUANTIZED_OPS})
+
+	# QAT2 ResNet50 with input/output scales in `fake_quantize_range_abs_max` operators and the `out_threshold` attributes,
+	# with weight scales in `fake_channel_wise_dequantize_max_abs` operators
+	set(QAT2_RESNET50_CHANNELWISE_MODEL_DIR "${QAT_INSTALL_DIR}/ResNet50_qat_channelwise")
+	set(QAT2_RESNET50_CHANNELWISE_MODEL_ARCHIVE "ResNet50_qat_channelwise.tar.gz")
+	download_qat_model(${QAT2_RESNET50_CHANNELWISE_MODEL_DIR} ${QAT2_RESNET50_CHANNELWISE_MODEL_ARCHIVE})
+	inference_qat2_int8_image_classification_test(test_qat2_int8_resnet50_channelwise_mkldnn ${QAT2_RESNET50_CHANNELWISE_MODEL_DIR}/ResNet50_qat_channelwise ${FP32_RESNET50_MODEL_DIR}/model ${IMAGENET_DATA_PATH} ${QAT2_IC_QUANTIZED_OPS})
+
 	# QAT2 MobileNetV1
         set(QAT2_MOBILENETV1_MODEL_DIR "${QAT_INSTALL_DIR}/MobileNet_qat_perf")
 	set(FP32_MOBILENETV1_MODEL_DIR "${INT8_INSTALL_DIR}/mobilenetv1")