Add support for Ernie NLP model to the Slim QAT (#22506)

* a test for Ernie QAT INT8 accuracy check

test=develop

* Remove NLP comparison test to split PRs

test=develop

* Fix typo and tabs, delete commented lines

test=develop

* re-combine the 2 PRs, test=develop
Co-authored-by: NMichał Gallus <sand3r@interia.eu>
Co-authored-by: Nbingyanghuang <33643817+bingyanghuang@users.noreply.github.com>

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -226,15 +226,18 @@ if(WITH_MKLDNN)
 
   set(INT8_DATA_DIR "${INFERENCE_DEMO_INSTALL_DIR}/int8v2")
 
-  ### Image classification tests
-  set(IMAGENET_DATA_PATH "${INT8_DATA_DIR}/data.bin")
-  set(INT8_IMG_CLASS_TEST_APP "test_analyzer_int8_image_classification")
-  set(INT8_IMG_CLASS_TEST_APP_SRC "analyzer_int8_image_classification_tester.cc")
+  ## Image classification models
 
-  # download dataset if necessary
-  download_int8_data(${INT8_DATA_DIR} "imagenet_val_100_tail.tar.gz")
+  # ImageNet small dataset
+  # May be already downloaded for INT8 QAT unit tests
+  set(IMAGENET_DATA_ARCHIVE "imagenet_val_100_tail.tar.gz")
+  set(IMAGENET_DATA_DIR "${INFERENCE_DEMO_INSTALL_DIR}/imagenet")
+  set(IMAGENET_DATA_PATH "${IMAGENET_DATA_DIR}/data.bin")
+  download_int8_data(${IMAGENET_DATA_DIR} ${IMAGENET_DATA_ARCHIVE})
 
   # build test binary to be used in subsequent tests
+  set(INT8_IMG_CLASS_TEST_APP "test_analyzer_int8_image_classification")
+  set(INT8_IMG_CLASS_TEST_APP_SRC "analyzer_int8_image_classification_tester.cc")
   inference_analysis_api_test_build(${INT8_IMG_CLASS_TEST_APP} ${INT8_IMG_CLASS_TEST_APP_SRC})
 
   # resnet50 int8
@@ -296,7 +299,7 @@ if(WITH_MKLDNN)
 
   ### optimized FP32 vs. QAT INT8 tests
   
-  set(QAT_DATA_DIR "${INFERENCE_DEMO_INSTALL_DIR}/int8v2")
+  set(QAT_DATA_DIR "${INFERENCE_DEMO_INSTALL_DIR}/qat")
   set(QAT_IMG_CLASS_TEST_APP "test_analyzer_qat_image_classification")
   set(QAT_IMG_CLASS_TEST_APP_SRC "analyzer_qat_image_classification_tester.cc")
 
@@ -304,8 +307,8 @@ if(WITH_MKLDNN)
   inference_analysis_api_test_build(${QAT_IMG_CLASS_TEST_APP} ${QAT_IMG_CLASS_TEST_APP_SRC})
 
   # MobileNet FP32 vs. QAT INT8
+  # The FP32 model should already be downloaded for slim QAT unit tests
   set(QAT2_MobileNet_MODEL_DIR "${QAT_DATA_DIR}/MobileNet_qat_perf")
-  download_qat_data(${QAT2_MobileNet_MODEL_DIR} "MobileNet_qat_perf.tar.gz")
   set(QAT2_INT8_MobileNet_MODEL_DIR "${QAT_DATA_DIR}/MobileNet_qat_perf_int8")
   download_qat_data(${QAT2_INT8_MobileNet_MODEL_DIR} "MobileNet_qat_perf_int8.tar.gz")
   inference_analysis_api_qat_test_run(test_analyzer_qat_performance_benchmark ${QAT_IMG_CLASS_TEST_APP} ${QAT2_MobileNet_MODEL_DIR}/MobileNet_qat_perf/float ${QAT2_INT8_MobileNet_MODEL_DIR}/MobileNet_qat_perf_int8 ${IMAGENET_DATA_PATH})

paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc

@@ -479,7 +479,7 @@ GetPrimitiveFactory(const MKLDNNDeviceContext& dev_ctx,
                     const Tensor* weights,
                     const mkldnn::engine& mkldnn_engine) {
   const std::string key = platform::CreateKey(
-      platform::ThreadIDasStr(), input->format(),
+      platform::ThreadIDasStr(), input->format(), input->dims()[0],
       framework::vectorize<int>(weights->dims()), ctx.OutputName("Out"));
 
   auto prim_creator =

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

@@ -17,10 +17,10 @@ from .... import core
 from ....framework import IrGraph
 from ....framework import IrNode
 
-__all__ = ['FakeQAT2MkldnnINT8KernelPass', 'FakeQAT2MkldnnINT8PerfPass']
+__all__ = ['QatInt8MkldnnPass', 'Qat2Int8MkldnnPass']
 
 
-class FakeQAT2MkldnnINT8KernelPass(object):
+class QatInt8MkldnnPass(object):
     """
     Convert QuantizationFreezePass generated IrGraph to MKL-DNN supported INT8
     IrGraph. Following transformations did in this pass:
@@ -48,13 +48,13 @@ class FakeQAT2MkldnnINT8KernelPass(object):
             # The original graph will be rewrite.
             import paddle.fluid as fluid
             from paddle.fluid.contrib.slim.quantization \
-                import FakeQAT2MkldnnINT8KernelPass
+                import QatInt8MkldnnPass
             from paddle.fluid.framework import IrGraph
             from paddle.fluid import core
 
             graph = IrGraph(core.Graph(fluid.Program().desc), for_test=False)
             place = fluid.CPUPlace()
-            mkldnn_pass = FakeQAT2MkldnnINT8KernelPass(fluid.global_scope(),
+            mkldnn_pass = QatInt8MkldnnPass(fluid.global_scope(),
             place)
             mkldnn_pass.apply(graph)
         """
@@ -276,7 +276,7 @@ class FakeQAT2MkldnnINT8KernelPass(object):
         graph.safe_remove_nodes(all_unused_vars)
 
 
-class FakeQAT2MkldnnINT8PerfPass(object):
+class Qat2Int8MkldnnPass(object):
     """
     Transform a QAT model IrGraph into MKL-DNN supported INT8 IrGraph.
     The pass consists of the following transformations:
@@ -290,7 +290,12 @@ class FakeQAT2MkldnnINT8PerfPass(object):
             passes (`cpu_quantize_pass`, `cpu_quantize_squash_pass`).
     """
 
-    def __init__(self, _scope=None, _place=None, _core=None, _debug=False):
+    def __init__(self,
+                 _quantized_ops,
+                 _scope=None,
+                 _place=None,
+                 _core=None,
+                 _debug=False):
         self._scope = _scope
         self._place = _place
         self._core = _core
@@ -305,6 +310,10 @@ class FakeQAT2MkldnnINT8PerfPass(object):
             'fake_quantize_dequantize_moving_average_abs_max'
         ]
         self._fake_dequantize_types = ['fake_dequantize_max_abs']
+        self._quantized_ops = _quantized_ops
+        self._scale_immutable_ops = [
+            'transpose2', 'reshape2', 'pool2d', 'scale'
+        ]
         self._conv_ops = ['conv2d', 'depthwise_conv2d']
         self._pool_ops = ['pool2d']
         self._mul_ops = ['mul']
@@ -324,8 +333,9 @@ class FakeQAT2MkldnnINT8PerfPass(object):
         graph = self._dequantize_weights(graph)
         graph = self._optimize_fp32_graph(graph)
         graph = self._compute_weight_scales(graph)
-        graph = self._update_conv_relu_scales(graph)
-        graph = self._update_pooling_scales(graph)
+        graph = self._update_relu_output_scales(graph)
+        graph = self._propagate_scales(graph)
+        graph = self._set_dummy_fc_out_scales(graph)
         graph = self._quantize_fp32_graph(graph)
         graph = self._remove_unused_var_nodes(graph)
         return graph
@@ -346,6 +356,12 @@ class FakeQAT2MkldnnINT8PerfPass(object):
         tensor.set(scale, core.CPUPlace())
         return tensor
 
+    def _is_conv_quantized(self):
+        return any(op_type in self._quantized_ops for op_type in self._conv_ops)
+
+    def _is_fc_quantized(self):
+        return 'fc' in self._quantized_ops
+
     def _gather_scales(self, graph):
         for op in graph.all_op_nodes():
             if op.name() in self._quantize_types:
@@ -371,34 +387,94 @@ class FakeQAT2MkldnnINT8PerfPass(object):
                 self._weight_scales[input_name] = _max_range
         return graph
 
-    def _update_pooling_scales(self, graph):
+    def _propagate_scales(self, graph):
+        def _update_scale_op_in_scale(op, input, output):
+            unsigned, tensor = self._var_quant_scales[output]
+            scale = np.array(tensor) * op.op().attr("scale")
+            new_tensor = self._convert_scale2tensor(scale.astype(np.float64))
+            self._var_quant_scales[input] = (unsigned, new_tensor)
+
+        def _update_scales(graph):
+            waiting_for_scale = set()
+            for op in graph.all_op_nodes():
+                if op.name() in self._scale_immutable_ops:
+                    input_name = op.input("X")[0]
+                    output_name = op.output("Out")[0]
+                    tensor_names = [input_name, output_name]
+
+                    # Scale is not quantized, so if it doesn't have any scales
+                    # to propagate, its tensors won't be added to the waiting list.
+                    if all(name not in self._var_quant_scales for name in tensor_names) \
+                            and op.name() != 'scale':
+                        waiting_for_scale.update(tensor_names)
+                        continue
+
+                    if input_name in self._var_quant_scales:
+                        self._var_quant_scales[
+                            output_name] = self._var_quant_scales[input_name]
+                    elif output_name in self._var_quant_scales:
+                        if op.name() == 'scale':
+                            _update_scale_op_in_scale(op, input_name,
+                                                      output_name)
+                        else:
+                            self._var_quant_scales[
+                                input_name] = self._var_quant_scales[
+                                    output_name]
+            return waiting_for_scale
+
+        waiting_for_scale = _update_scales(graph)
+
+        while len(waiting_for_scale) != 0:
+            waiting_for_scale = _update_scales(graph)
+
+        return graph
+
+    def _set_dummy_fc_out_scales(self, graph):
+        '''
+        For the output tensors of FC that do not have an assigned scale,
+        assign a dummy scale (same scale as input), so that the quantize pass
+        won't fail. In the end these scales aren't used, since FCs that
+        have an unassigend output scale will have a force_fp32_output attr
+        set to True.
+        '''
         for op in graph.all_op_nodes():
-            if op.name() in self._pool_ops:
-                input_name = op.input("X")[0]
+            if op.name() in self._fc_ops:
+                input_name = op.input("Input")[0]
                 output_name = op.output("Out")[0]
-                if input_name in self._var_quant_scales:
+                if input_name in self._var_quant_scales and \
+                    output_name not in self._var_quant_scales:
+                    # use input scale as a "dummy" scale
                     self._var_quant_scales[
                         output_name] = self._var_quant_scales[input_name]
+
         return graph
 
     def _load_param(self, scope, param_name):
         return np.array(scope.find_var(param_name).get_tensor())
 
     def _remove_fake_ops(self, graph):
+        '''
+        When FC isn't quantized:
+        Remove fake (de)quantize ops that do not surround mul.
+        When FC is quantized:
+        Remove all fake (de)quantize ops.
+        '''
+        is_fc_quantized = self._is_fc_quantized()
         for op in graph.all_op_nodes():
             if op.name() in self._fake_quantize_types:
                 op_out = graph._find_node_by_name(op.outputs,
                                                   op.output("Out")[0])
                 next_op = op_out.outputs[0]
-                if next_op.name() not in self._mul_ops:
+                if next_op.name() not in self._mul_ops or is_fc_quantized:
                     self._remove_fake_quantize(graph, op)
 
         for op in graph.all_op_nodes():
             if op.name() in self._fake_dequantize_types:
                 op_in = graph._find_node_by_name(op.inputs, op.input("X")[0])
                 prev_op = op_in.inputs[0]
-                if prev_op.name() not in self._mul_ops:
+                if prev_op.name() not in self._mul_ops or is_fc_quantized:
                     self._remove_fake_dequantize(graph, op)
+
         return graph
 
     def _remove_fake_quantize(self, graph, op):
@@ -444,6 +520,8 @@ class FakeQAT2MkldnnINT8PerfPass(object):
         for op in graph.all_op_nodes():
             if op.name() in self._conv_ops:
                 self._dequantize_conv_weights(graph, op)
+            elif self._is_fc_quantized() and op.name() in self._mul_ops:
+                self._dequantize_mul_weights(graph, op)
         return graph
 
     def _dequantize_conv_weights(self, graph, op_node):
@@ -472,13 +550,20 @@ class FakeQAT2MkldnnINT8PerfPass(object):
     def _optimize_fp32_graph(self, graph):
         graph = self._apply_pass(graph, 'mkldnn_placement_pass',
                                  ['mkldnn_enabled_op_types'], [set()])
-        graph = self._apply_pass(graph, 'depthwise_conv_mkldnn_pass')
-        graph = self._apply_pass(graph, 'conv_bn_fuse_pass')
-        graph = self._apply_pass(graph, 'conv_eltwiseadd_bn_fuse_pass')
-        graph = self._apply_pass(graph, 'conv_bias_mkldnn_fuse_pass')
-        graph = self._apply_pass(graph, 'conv_elementwise_add_mkldnn_fuse_pass')
-        graph = self._apply_pass(graph, 'conv_relu_mkldnn_fuse_pass')
-        graph = self._apply_pass(graph, 'conv_relu6_mkldnn_fuse_pass')
+        if self._is_conv_quantized():
+            graph = self._apply_pass(graph, 'depthwise_conv_mkldnn_pass')
+            graph = self._apply_pass(graph, 'conv_bn_fuse_pass')
+            graph = self._apply_pass(graph, 'conv_eltwiseadd_bn_fuse_pass')
+            graph = self._apply_pass(graph, 'conv_bias_mkldnn_fuse_pass')
+            graph = self._apply_pass(graph,
+                                     'conv_elementwise_add_mkldnn_fuse_pass')
+            graph = self._apply_pass(graph, 'conv_relu_mkldnn_fuse_pass')
+            graph = self._apply_pass(graph, 'conv_relu6_mkldnn_fuse_pass')
+        if self._is_fc_quantized():
+            graph = self._apply_pass(graph, 'fc_fuse_pass',
+                                     ['use_gpu', 'use_fc_padding'],
+                                     [False, False])
+            graph = self._apply_pass(graph, 'fc_mkldnn_pass')
         return graph
 
     def _apply_pass(self, graph, pass_name, attrs=None, attr_values=None):
@@ -528,6 +613,7 @@ class FakeQAT2MkldnnINT8PerfPass(object):
                         np.abs(weights.reshape(weights.shape[0], -1)).astype(
                             np.float64),
                         axis=axis)
+                    scales[scales == np.Inf] = 0.0
 
                     lod_tensor = self._convert_scale2tensor(scales)
                     use_unsigned_int = False
@@ -546,46 +632,41 @@ class FakeQAT2MkldnnINT8PerfPass(object):
                     ids.append(op.id())
         return set(ids) if len(ids) else set([-1])
 
-    def _transform_to_quantize_mkldnn(self, graph, op_node):
-        """
-        Transform fake_quantize_xx op to quantize mkldnn op in the graph.
-        """
-        input_var_node = graph._find_node_by_name(op_node.inputs,
-                                                  op_node.input("X")[0])
-        output_var_node = graph._find_node_by_name(op_node.outputs,
-                                                   op_node.output("Out")[0])
-        scale_in = self._s8_max / self._load_param(
-            self._scope, op_node.input("InScale")[0])[0]
-        quant_op_node = graph.create_op_node(
-            op_type='quantize',
-            attrs={
-                'data_format': 'MKLDNNLAYOUT',
-                'use_mkldnn': 1,
-                'Scale': scale_in,
-                'is_negative_input': 1
-            },
-            inputs={'Input': input_var_node},
-            outputs={'Output': output_var_node})
-        graph.link_to(input_var_node, quant_op_node)
-        graph.link_to(quant_op_node, output_var_node)
-        graph.safe_remove_nodes(op_node)
-        return quant_op_node
+    def _update_relu_output_scales(self, graph):
+        def _update_scale(graph, ops, op_out_name, predicate):
+            '''
+            Sets the type of an output scale of a passed op type(s) to 'unsigned int8' if the
+            predicate applied on op passes. Typically, the predicate checks if op's
+            activation is set to relu.
+            '''
+            for op in graph.all_op_nodes():
+                if op.name() in ops:
+                    out_name = op.output(op_out_name)[0]
+                    if out_name in self._var_quant_scales and predicate(op.op(
+                    )):
+                        _, tensor = self._var_quant_scales[out_name]
+                        self._var_quant_scales[out_name] = (True, tensor)
+            return graph
+
+        if self._is_conv_quantized():
+            conv_predicate = lambda op: op.attr("fuse_activation") == 'relu' and \
+                op.attr("fuse_residual_connection") == False
+            graph = _update_scale(graph, self._conv_ops, "Output",
+                                  conv_predicate)
+
+        if self._is_fc_quantized():
+            fc_predicate = lambda op: op.attr("activation_type") == 'relu'
+            graph = _update_scale(graph, self._fc_ops, "Out", fc_predicate)
 
-    def _update_conv_relu_scales(self, graph):
-        for op in graph.all_op_nodes():
-            if op.name() in self._conv_ops:
-                out_name = op.output("Output")[0]
-                if out_name in self._var_quant_scales and \
-                   op.op().attr("fuse_activation") == 'relu' and \
-                   op.op().attr("fuse_residual_connection") == False:
-                    _, tensor = self._var_quant_scales[out_name]
-                    self._var_quant_scales[out_name] = (True, tensor)
         return graph
 
+    def _get_data_layout(self):
+        return 'NHWC' if self._is_conv_quantized() else 'NCHW'
+
     def _quantize_fp32_graph(self, graph):
         ir_pass = self._core.get_pass('cpu_quantize_placement_pass')
         cpp_graph = graph.graph
-        ir_pass.set('quantize_enabled_op_types', {'conv2d', 'pool2d'})
+        ir_pass.set('quantize_enabled_op_types', self._quantized_ops)
         ir_pass.set('quantize_excluded_op_ids',
                     self._find_avg_pooling_ids(graph))
         ir_pass.apply(cpp_graph)
@@ -593,8 +674,8 @@ class FakeQAT2MkldnnINT8PerfPass(object):
             graph.draw('.', 'qat_int8_{}'.format(ir_pass.type()),
                        graph.all_op_nodes())
 
-        graph = self._apply_pass(graph, 'cpu_quantize_pass',
-                                 ['quant_var_scales'],
-                                 [self._var_quant_scales])
+        graph = self._apply_pass(
+            graph, 'cpu_quantize_pass', ['quant_var_scales', 'data_layout'],
+            [self._var_quant_scales, self._get_data_layout()])
         graph = self._apply_pass(graph, 'cpu_quantize_squash_pass')
         return graph

python/paddle/fluid/contrib/slim/tests/qat_int8_comparison.py → python/paddle/fluid/contrib/slim/tests/qat_int8_image_classification_comparison.py

@@ -24,8 +24,8 @@ import time
 import paddle
 import paddle.fluid as fluid
 from paddle.fluid.framework import IrGraph
-from paddle.fluid.contrib.slim.quantization import FakeQAT2MkldnnINT8KernelPass
-from paddle.fluid.contrib.slim.quantization import FakeQAT2MkldnnINT8PerfPass
+from paddle.fluid.contrib.slim.quantization import QatInt8MkldnnPass
+from paddle.fluid.contrib.slim.quantization import Qat2Int8MkldnnPass
 from paddle.fluid import core
 
 logging.basicConfig(format='%(asctime)s-%(levelname)s: %(message)s')
@@ -53,10 +53,6 @@ def parse_args():
         action='store_true',
         help='If used, the QAT model is treated as a second generation model for performance optimization.'
     )
-    parser.add_argument(
-        '--save_model',
-        action='store_true',
-        help='If used, the QAT model will be saved after all transformations')
     parser.add_argument('--infer_data', type=str, default='', help='Data file.')
     parser.add_argument(
         '--batch_num',
@@ -68,15 +64,20 @@ def parse_args():
         type=float,
         default=0.01,
         help='Accepted accuracy difference threshold.')
+    parser.add_argument(
+        '--quantized_ops',
+        type=str,
+        default='',
+        help='A comma separated list of quantized operators.')
 
     test_args, args = parser.parse_known_args(namespace=unittest)
 
     return test_args, sys.argv[:1] + args
 
 
-class TestQatInt8Comparison(unittest.TestCase):
+class QatInt8ImageClassificationComparisonTest(unittest.TestCase):
     """
-    Test for accuracy comparison of QAT FP32 and INT8 inference.
+    Test for accuracy comparison of QAT FP32 and INT8 Image Classification inference.
     """
 
     def _reader_creator(self, data_file='data.bin'):
@@ -182,14 +183,15 @@ class TestQatInt8Comparison(unittest.TestCase):
                 graph.draw('.', 'qat_orig', graph.all_op_nodes())
             if (transform_to_int8):
                 if (test_case_args.qat2):
-                    transform_to_mkldnn_int8_pass = FakeQAT2MkldnnINT8PerfPass(
+                    transform_to_mkldnn_int8_pass = Qat2Int8MkldnnPass(
+                        self._quantized_ops,
                         _scope=inference_scope,
                         _place=place,
                         _core=core,
                         _debug=self._debug)
                     graph = transform_to_mkldnn_int8_pass.apply(graph)
                 else:
-                    mkldnn_int8_pass = FakeQAT2MkldnnINT8KernelPass(
+                    mkldnn_int8_pass = QatInt8MkldnnPass(
                         _scope=inference_scope, _place=place)
                     graph = mkldnn_int8_pass.apply(graph)
 
@@ -256,12 +258,6 @@ class TestQatInt8Comparison(unittest.TestCase):
             _logger.info('Total inference run time: {:.2f} s'.format(
                 infer_total_time))
 
-            if test_case_args.save_model:
-                with fluid.scope_guard(inference_scope):
-                    fluid.io.save_inference_model(
-                        'transformed_qat_int8_model', feed_target_names,
-                        fetch_targets, exe, inference_program)
-
             return outputs, acc1_avg, acc5_avg, fps_avg, latency_avg
 
     def _summarize_performance(self, fp32_fps, fp32_lat, int8_fps, int8_lat):
@@ -298,6 +294,7 @@ class TestQatInt8Comparison(unittest.TestCase):
         skip_batch_num = test_case_args.skip_batch_num
         acc_diff_threshold = test_case_args.acc_diff_threshold
         self._debug = test_case_args.debug
+        self._quantized_ops = set(test_case_args.quantized_ops.split(','))
 
         _logger.info('QAT FP32 & INT8 prediction run.')
         _logger.info('QAT model: {0}'.format(qat_model_path))
@@ -305,6 +302,7 @@ class TestQatInt8Comparison(unittest.TestCase):
         _logger.info('Batch size: {0}'.format(batch_size))
         _logger.info('Batch number: {0}'.format(batch_num))
         _logger.info('Accuracy drop threshold: {0}.'.format(acc_diff_threshold))
+        _logger.info('Quantized ops: {0}.'.format(self._quantized_ops))
 
         _logger.info('--- QAT FP32 prediction start ---')
         val_reader = paddle.batch(

python/paddle/fluid/contrib/slim/tests/qat_int8_nlp_comparison.py

+#   copyright (c) 2020 paddlepaddle authors. all rights reserved.
+#
+# 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.
+
+import unittest
+import os
+import sys
+import argparse
+import logging
+import struct
+import six
+import numpy as np
+import time
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.framework import IrGraph
+from paddle.fluid.contrib.slim.quantization import Qat2Int8MkldnnPass
+from paddle.fluid import core
+
+logging.basicConfig(format='%(asctime)s-%(levelname)s: %(message)s')
+_logger = logging.getLogger(__name__)
+_logger.setLevel(logging.INFO)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch_size', type=int, default=1, help='Batch size.')
+    parser.add_argument(
+        '--skip_batch_num',
+        type=int,
+        default=0,
+        help='Number of the first minibatches to skip in performance statistics.'
+    )
+    parser.add_argument(
+        '--debug',
+        action='store_true',
+        help='If used, the graph of QAT model is drawn.')
+    parser.add_argument(
+        '--qat_model', type=str, default='', help='A path to a QAT model.')
+    parser.add_argument(
+        '--save_model',
+        action='store_true',
+        help='If used, the QAT model will be saved after all transformations')
+    parser.add_argument('--infer_data', type=str, default='', help='Data file.')
+    parser.add_argument(
+        '--labels', type=str, default='', help='File with labels.')
+    parser.add_argument(
+        '--batch_num',
+        type=int,
+        default=1,
+        help='Number of batches to process. 0 or less means all.')
+    parser.add_argument(
+        '--acc_diff_threshold',
+        type=float,
+        default=0.01,
+        help='Accepted accuracy difference threshold.')
+    parser.add_argument(
+        '--quantized_ops',
+        type=str,
+        default='',
+        help='A comma separated list of quantized operators.')
+
+    test_args, args = parser.parse_known_args(namespace=unittest)
+
+    return test_args, sys.argv[:1] + args
+
+
+class QatInt8NLPComparisonTest(unittest.TestCase):
+    """
+    Test for accuracy comparison of QAT FP32 and INT8 NLP inference.
+    """
+
+    def _reader_creator(self, data_file=None, labels_file=None):
+        assert data_file, "The dataset file is missing."
+        assert labels_file, "The labels file is missing."
+
+        def reader():
+            with open(data_file, 'r') as df:
+                with open(labels_file, 'r') as lf:
+                    data_lines = df.readlines()
+                    labels_lines = lf.readlines()
+                    assert len(data_lines) == len(
+                        labels_lines
+                    ), "The number of labels does not match the length of the dataset."
+
+                    for i in range(len(data_lines)):
+                        data_fields = data_lines[i].split(';')
+                        assert len(
+                            data_fields
+                        ) >= 2, "The number of data fields in the dataset is less than 2"
+                        buffers = []
+                        shape = []
+                        for j in range(2):
+                            data = data_fields[j].split(':')
+                            assert len(
+                                data
+                            ) >= 2, "Size of data in the dataset is less than 2"
+                            # Shape is stored under index 0, while data under 1
+                            shape = data[0].split()
+                            shape.pop(0)
+                            shape_np = np.array(shape).astype("int64")
+                            buffer_i = data[1].split()
+                            buffer_np = np.array(buffer_i).astype("int64")
+                            buffer_np.shape = tuple(shape_np)
+                            buffers.append(buffer_np)
+                        label = labels_lines[i]
+                        yield buffers[0], buffers[1], int(label)
+
+        return reader
+
+    def _get_batch_correct(self, batch_output=None, labels=None):
+        total = len(batch_output)
+        assert total > 0, "The batch output is empty."
+        correct = 0
+        for n, output in enumerate(batch_output[0]):
+            max_idx = np.where(output == output.max())
+            if max_idx == labels[n]:
+                correct += 1
+        return correct
+
+    def _predict(self,
+                 test_reader=None,
+                 model_path=None,
+                 batch_size=1,
+                 batch_num=1,
+                 skip_batch_num=0,
+                 transform_to_int8=False):
+        place = fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        inference_scope = fluid.executor.global_scope()
+        with fluid.scope_guard(inference_scope):
+            if os.path.exists(os.path.join(model_path, '__model__')):
+                [inference_program, feed_target_names,
+                 fetch_targets] = fluid.io.load_inference_model(model_path, exe)
+            else:
+                [inference_program, feed_target_names,
+                 fetch_targets] = fluid.io.load_inference_model(
+                     model_path, exe, 'model', 'params')
+
+            graph = IrGraph(core.Graph(inference_program.desc), for_test=True)
+            if (self._debug):
+                graph.draw('.', 'qat_orig', graph.all_op_nodes())
+            if (transform_to_int8):
+                transform_to_mkldnn_int8_pass = Qat2Int8MkldnnPass(
+                    self._quantized_ops,
+                    _scope=inference_scope,
+                    _place=place,
+                    _core=core,
+                    _debug=self._debug)
+                graph = transform_to_mkldnn_int8_pass.apply(graph)
+
+            inference_program = graph.to_program()
+
+            total_correct = 0
+            total_samples = 0
+            batch_times = []
+            ppses = []  # predictions per second
+            iters = 0
+            infer_start_time = time.time()
+            for data in test_reader():
+                if batch_num > 0 and iters >= batch_num:
+                    break
+                if iters == skip_batch_num:
+                    total_samples = 0
+                    infer_start_time = time.time()
+                input0 = np.array([x[0] for x in data]).astype('int64')
+                input1 = np.array([x[1] for x in data]).astype('int64')
+                labels = np.array([x[2] for x in data]).astype('int64')
+
+                start = time.time()
+                out = exe.run(inference_program,
+                              feed={
+                                  feed_target_names[0]: input0,
+                                  feed_target_names[1]: input1
+                              },
+                              fetch_list=fetch_targets)
+                batch_time = (time.time() - start) * 1000  # in miliseconds
+                batch_times.append(batch_time)
+                batch_correct = self._get_batch_correct(out, labels)
+                batch_len = len(data)
+                total_samples += batch_len
+                total_correct += batch_correct
+                batch_acc = float(batch_correct) / float(batch_len)
+                pps = batch_len / batch_time * 1000
+                ppses.append(pps)
+                latency = batch_time / batch_len
+                iters += 1
+                appx = ' (warm-up)' if iters <= skip_batch_num else ''
+                _logger.info(
+                    'batch {0}{4}, acc: {1:.4f}, latency: {2:.4f} ms, predictions per sec: {3:.2f}'
+                    .format(iters, batch_acc, latency, pps, appx))
+
+            # Postprocess benchmark data
+            infer_total_time = time.time() - infer_start_time
+            batch_latencies = batch_times[skip_batch_num:]
+            batch_latency_avg = np.average(batch_latencies)
+            latency_avg = batch_latency_avg / batch_size
+            ppses = ppses[skip_batch_num:]
+            pps_avg = np.average(ppses)
+            acc_avg = float(np.sum(total_correct)) / float(total_samples)
+            _logger.info('Total inference run time: {:.2f} s'.format(
+                infer_total_time))
+
+            return acc_avg, pps_avg, latency_avg
+
+    def _summarize_performance(self, fp32_pps, fp32_lat, int8_pps, int8_lat):
+        _logger.info('--- Performance summary ---')
+        _logger.info(
+            'FP32: avg predictions per sec: {0:.2f}, avg latency: {1:.4f} ms'.
+            format(fp32_pps, fp32_lat))
+        _logger.info(
+            'INT8: avg predictions per sec: {0:.2f}, avg latency: {1:.4f} ms'.
+            format(int8_pps, int8_lat))
+
+    def _compare_accuracy(self, fp32_acc, int8_acc, threshold):
+        _logger.info('--- Accuracy summary ---')
+        _logger.info(
+            'Accepted accuracy drop threshold: {0}. (condition: (FP32_acc - INT8_acc) <= threshold)'
+            .format(threshold))
+        _logger.info('FP32: avg accuracy: {0:.6f}'.format(fp32_acc))
+        _logger.info('INT8: avg accuracy: {0:.6f}'.format(int8_acc))
+        # Random outputs give accuracy about 0.33, we assume valid accuracy to be at least 0.5
+        assert fp32_acc > 0.5
+        assert int8_acc > 0.5
+        assert fp32_acc - int8_acc <= threshold
+
+    def test_graph_transformation(self):
+        if not fluid.core.is_compiled_with_mkldnn():
+            return
+
+        qat_model_path = test_case_args.qat_model
+        data_path = test_case_args.infer_data
+        labels_path = test_case_args.labels
+        batch_size = test_case_args.batch_size
+        batch_num = test_case_args.batch_num
+        skip_batch_num = test_case_args.skip_batch_num
+        acc_diff_threshold = test_case_args.acc_diff_threshold
+        self._debug = test_case_args.debug
+        self._quantized_ops = set(test_case_args.quantized_ops.split(','))
+
+        _logger.info('QAT FP32 & INT8 prediction run.')
+        _logger.info('QAT model: {0}'.format(qat_model_path))
+        _logger.info('Dataset: {0}'.format(data_path))
+        _logger.info('Labels: {0}'.format(labels_path))
+        _logger.info('Batch size: {0}'.format(batch_size))
+        _logger.info('Batch number: {0}'.format(batch_num))
+        _logger.info('Accuracy drop threshold: {0}.'.format(acc_diff_threshold))
+        _logger.info('Quantized ops: {0}.'.format(self._quantized_ops))
+
+        _logger.info('--- QAT FP32 prediction start ---')
+        val_reader = paddle.batch(
+            self._reader_creator(data_path, labels_path), batch_size=batch_size)
+        fp32_acc, fp32_pps, fp32_lat = self._predict(
+            val_reader,
+            qat_model_path,
+            batch_size,
+            batch_num,
+            skip_batch_num,
+            transform_to_int8=False)
+        _logger.info('--- QAT INT8 prediction start ---')
+        val_reader = paddle.batch(
+            self._reader_creator(data_path, labels_path), batch_size=batch_size)
+        int8_acc, int8_pps, int8_lat = self._predict(
+            val_reader,
+            qat_model_path,
+            batch_size,
+            batch_num,
+            skip_batch_num,
+            transform_to_int8=True)
+
+        self._summarize_performance(fp32_pps, fp32_lat, int8_pps, int8_lat)
+        self._compare_accuracy(fp32_acc, int8_acc, acc_diff_threshold)
+
+
+if __name__ == '__main__':
+    global test_case_args
+    test_case_args, remaining_args = parse_args()
+    unittest.main(argv=remaining_args)

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

@@ -24,7 +24,7 @@ import time
 import paddle
 import paddle.fluid as fluid
 from paddle.fluid.framework import IrGraph
-from paddle.fluid.contrib.slim.quantization import FakeQAT2MkldnnINT8PerfPass
+from paddle.fluid.contrib.slim.quantization import Qat2Int8MkldnnPass
 from paddle.fluid import core
 
 
@@ -42,6 +42,11 @@ def parse_args():
         type=str,
         default='',
         help='Saved optimized and quantized INT8 model')
+    parser.add_argument(
+        '--quantized_ops',
+        type=str,
+        default='',
+        help='A comma separated list of quantized operators.')
 
     test_args, args = parser.parse_known_args(namespace=unittest)
     return test_args, sys.argv[:1] + args
@@ -60,8 +65,9 @@ def transform_and_save_model(original_path, save_path, save_type):
              fetch_targets] = fluid.io.load_inference_model(original_path, exe,
                                                             'model', 'params')
 
-        transform_to_mkldnn_int8_pass = FakeQAT2MkldnnINT8PerfPass(
-            _scope=inference_scope, _place=place, _core=core)
+        quantized_ops = set(test_args.quantized_ops.split(','))
+        transform_to_mkldnn_int8_pass = Qat2Int8MkldnnPass(
+            quantized_ops, _scope=inference_scope, _place=place, _core=core)
 
         graph = IrGraph(core.Graph(inference_program.desc), for_test=True)
         if save_type == 'FP32':

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

@@ -22,7 +22,7 @@ import paddle
 from paddle.fluid.framework import IrGraph
 from paddle.fluid.contrib.slim.quantization import QuantizationFreezePass
 from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass
-from paddle.fluid.contrib.slim.quantization import FakeQAT2MkldnnINT8KernelPass
+from paddle.fluid.contrib.slim.quantization import QatInt8MkldnnPass
 from paddle.fluid import core
 
 os.environ["CPU_NUM"] = "1"
@@ -149,8 +149,7 @@ class TestMKLDNNTransformBasedFreezePass(unittest.TestCase):
         freeze_pass.apply(test_graph)
 
         # Transform quantized graph for MKL-DNN INT8 inference
-        mkldnn_int8_pass = FakeQAT2MkldnnINT8KernelPass(
-            _scope=scope, _place=place)
+        mkldnn_int8_pass = QatInt8MkldnnPass(_scope=scope, _place=place)
         mkldnn_int8_pass.apply(test_graph)
         dev_name = '_cpu_'
         if not for_ci: