Add deploy_backend in PTQ and QAT & update to the default onnx new format (#1601)

paddleslim/auto_compression/compressor.py

@@ -890,8 +890,7 @@ class AutoCompression:
                 test_program,
                 self._places,
                 self._quant_config,
-                scope=paddle.static.global_scope(),
-                save_clip_ranges_path=self.final_dir)
+                scope=paddle.static.global_scope())
 
         feed_vars = [
             test_program.global_block().var(name)

paddleslim/auto_compression/utils/fake_ptq.py

@@ -3,14 +3,17 @@ import paddle
 from paddle.fluid.framework import IrGraph
 from paddle.framework import core
 from paddle.static.quantization import QuantizationTransformPass, QuantizationTransformPassV2, AddQuantDequantPass, AddQuantDequantPassV2, QuantizationFreezePass, QuantWeightPass
+from paddle.static.quantization import utils
 
 try:
-    from paddle.static.quantization import utils
+    from paddle.static.quantization import quant_config
+    TRANSFORM_PASS_OP_TYPES = list(
+        quant_config.SUPPORT_WEIGHT_QUANTIZATION_OP_DICT.keys())
+    QUANT_DEQUANT_PASS_OP_TYPES = list(
+        quant_config.SUPPORT_ACT_QUANTIZATION_OP_DICT.keys())
+except:
     TRANSFORM_PASS_OP_TYPES = utils._weight_supported_quantizable_op_type
     QUANT_DEQUANT_PASS_OP_TYPES = utils._act_supported_quantizable_op_type
-except:
-    TRANSFORM_PASS_OP_TYPES = QuantizationTransformPass._supported_quantizable_op_type
-    QUANT_DEQUANT_PASS_OP_TYPES = AddQuantDequantPass._supported_quantizable_op_type
 
 from ...common.load_model import load_inference_model
 
@@ -155,7 +158,9 @@ def post_quant_fake(executor,
 
         for block_id in range(len(_program.blocks)):
             for op in _program.blocks[block_id].ops:
-                if op.type in (_quantizable_op_type + utils._out_scale_op_list):
+                if op.type in (
+                        _quantizable_op_type +
+                        list(quant_config.SUPPORT_QUANTIZATION_OP_DICT.keys())):
                     out_var_names = utils._get_op_output_var_names(op)
                     for var_name in out_var_names:
                         analysis_and_save_info(op, var_name)

paddleslim/dygraph/quant/qat.py

@@ -57,8 +57,8 @@ _quant_config_default = {
     'quantizable_layer_type': ['Conv2D', 'Linear'],
     # whether fuse conv and bn before QAT
     'fuse_conv_bn': False,
-    # Whether to export the quantized model with format of ONNX. Default is False.
-    'onnx_format': False,
+    # Whether to export the quantized model with format of ONNX. Default is True.
+    'onnx_format': True,
 }
 
 

paddleslim/quant/quant_aware_with_infermodel.py

@@ -311,9 +311,7 @@ def export_quant_infermodel(
     #    operators' order for the inference.
     #    The dtype of float_program's weights is float32, but in int8 range.
     ############################################################################################################
-    float_program, int8_program = convert(test_program, place, quant_config, \
-                                          scope=scope, \
-                                          save_int8=True)
+    float_program = convert(test_program, place, quant_config, scope=scope)
     ############################################################################################################
     # 4. Save inference model
     ############################################################################################################

paddleslim/quant/quanter.py

@@ -41,9 +41,10 @@ try:
     from paddle.static.quantization import AddQuantDequantPassV2
     from paddle.static.quantization import PostTrainingQuantizationProgram
     from paddle.static.quantization import AddQuantDequantForInferencePass
+    from paddle.static.quantization import quant_config
 except:
     _logger.warning(
-        "Some functions fail to import, please update PaddlePaddle version to 2.4+"
+        "Some functions failed to import, better to update PaddlePaddle to the latest develop version."
     )
 
 WEIGHT_QUANTIZATION_TYPES = [
@@ -61,12 +62,14 @@ ACTIVATION_QUANTIZATION_TYPES_TENSORRT = [
 
 VALID_DTYPES = ['int8']
 try:
+    TRANSFORM_PASS_OP_TYPES = list(
+        quant_config.SUPPORT_WEIGHT_QUANTIZATION_OP_DICT.keys())
+    QUANT_DEQUANT_PASS_OP_TYPES = list(
+        quant_config.SUPPORT_ACT_QUANTIZATION_OP_DICT.keys())
+except:
     from paddle.static.quantization import utils
     TRANSFORM_PASS_OP_TYPES = utils._weight_supported_quantizable_op_type
     QUANT_DEQUANT_PASS_OP_TYPES = utils._act_supported_quantizable_op_type
-except:
-    TRANSFORM_PASS_OP_TYPES = QuantizationTransformPass._supported_quantizable_op_type
-    QUANT_DEQUANT_PASS_OP_TYPES = AddQuantDequantPass._supported_quantizable_op_type
 
 TENSORRT_OP_TYPES = [
     'mul', 'conv2d', 'pool2d', 'depthwise_conv2d', 'elementwise_add',
@@ -99,11 +102,13 @@ _quant_config_default = {
     # if True, 'quantoze_op_types' will be TRANSFORM_PASS_OP_TYPES + QUANT_DEQUANT_PASS_OP_TYPES 
     'is_full_quantize': False,
     # if True, use onnx format to quant.
-    'onnx_format': False,
+    'onnx_format': True,
     # quant post to get initial scale for quant_aware
     'quant_post_first': False,
     # whether scale can be train
-    'scale_trainable': True
+    'scale_trainable': True,
+    # Deploy backend, it could be: None, TensorRT, MKLDNN, ARM
+    'deploy_backend': None
 }
 
 
@@ -195,6 +200,32 @@ def _parse_configs(user_config):
     assert isinstance(configs['moving_rate'], float), \
         "moving_rate must be float value, The decay coefficient of moving average, default is 0.9."
 
+    deploy_backend = configs['deploy_backend']
+    assert not deploy_backend or deploy_backend.lower() in [
+        'tensorrt', 'mkldnn', 'arm'
+    ], "Deploy Backend {} not support, please choose None, tensorrt or mkldnn.".format(
+        deploy_backend)
+    try:
+        if not deploy_backend:
+            configs['quant_config'] = quant_config.BaseQuantizer(
+                quantizable_op_type=configs['quantize_op_types'],
+                quant_bits=configs['weight_bits'], )
+        elif deploy_backend.lower() == "tensorrt":
+            configs['quant_config'] = quant_config.TensorRTQuantizer(
+                quantizable_op_type=configs['quantize_op_types'],
+                quant_bits=configs['weight_bits'], )
+        elif deploy_backend.lower() == "mkldnn":
+            configs['quant_config'] = quant_config.MKLDNNQuantizer(
+                quantizable_op_type=configs['quantize_op_types'],
+                quant_bits=configs['weight_bits'], )
+        elif deploy_backend.lower() == "arm":
+            configs['quant_config'] = quant_config.ARMCPUQuantizer(
+                quantizable_op_type=configs['quantize_op_types'],
+                quant_bits=configs['weight_bits'], )
+    except:
+        _logger.warning(
+            "Set deploy_backend failed, Please update to PaddlePaddle Develop.")
+
     return configs
 
 
@@ -388,11 +419,17 @@ def quant_aware(program,
         sub_graphs = [sub_graph for sub_graph in main_graph.all_sub_graphs()]
         transform_pass_ops = []
         quant_dequant_ops = []
-        for op_type in config['quantize_op_types']:
-            if op_type in TRANSFORM_PASS_OP_TYPES:
-                transform_pass_ops.append(op_type)
-            elif op_type in QUANT_DEQUANT_PASS_OP_TYPES:
-                quant_dequant_ops.append(op_type)
+        if 'quant_config' in config and config['quant_config']:
+            transform_pass_ops = config[
+                'quant_config'].weight_quant_operation_types
+            quant_dequant_ops = config[
+                'quant_config'].activation_quant_operation_types
+        else:
+            for op_type in config['quantize_op_types']:
+                if op_type in TRANSFORM_PASS_OP_TYPES:
+                    transform_pass_ops.append(op_type)
+                elif op_type in QUANT_DEQUANT_PASS_OP_TYPES:
+                    quant_dequant_ops.append(op_type)
         if len(transform_pass_ops) > 0:
             trannsform_func = 'QuantizationTransformPassV2' if config[
                 'onnx_format'] else 'QuantizationTransformPass'
@@ -486,8 +523,8 @@ def quant_post_static(executor,
                       hist_percent=0.9999,
                       bias_correction=False,
                       quantizable_op_type=[
-                          "conv2d", "depthwise_conv2d", "mul", "matmul",
-                          "matmul_v2"
+                          "conv2d", "depthwise_conv2d", "conv2d_transpose",
+                          "mul", "matmul", "matmul_v2"
                       ],
                       is_full_quantize=False,
                       weight_bits=8,
@@ -495,10 +532,9 @@ def quant_post_static(executor,
                       activation_quantize_type='range_abs_max',
                       weight_quantize_type='channel_wise_abs_max',
                       optimize_model=False,
-                      onnx_format=False,
+                      onnx_format=True,
                       skip_tensor_list=None,
-                      is_use_cache_file=False,
-                      cache_dir="./temp_post_training"):
+                      deploy_backend=None):
     """
     The function utilizes static post training quantization method to
     quantize the fp32 model. It uses calibrate data to calculate the
@@ -568,10 +604,11 @@ def quant_post_static(executor,
         optimize_model(bool, optional): If set optimize_model as True, it applies some 
                 passes to optimize the model before quantization. So far, the place of
                 executor must be cpu it supports fusing batch_norm into convs.
-        onnx_format(bool): Whether to export the quantized model with format of ONNX. Default is False.
+        onnx_format(bool): Whether to export the quantized model with format of ONNX. Default is True.
         skip_tensor_list(list): List of skip quant tensor name.
-        is_use_cache_file(bool): This param is deprecated.
-        cache_dir(str): This param is deprecated.
+        deploy_backend(str): Deploy backend, it could be None, TensorRT, MKLDNN, ARM.
+                Other backends will continue to expand, the default is None, which means to
+                use the default general quantization configuration.
     
     Returns:
         None
@@ -599,8 +636,10 @@ def quant_post_static(executor,
             activation_quantize_type=activation_quantize_type,
             weight_quantize_type=weight_quantize_type,
             onnx_format=onnx_format,
-            skip_tensor_list=skip_tensor_list,  # support in Paddle >= 2.3.1
-            optimize_model=optimize_model)
+            skip_tensor_list=skip_tensor_list,
+            optimize_model=optimize_model,
+            deploy_backend=deploy_backend,  # support at Paddle develop
+        )
     except:
         post_training_quantization = PostTrainingQuantization(
             executor=executor,
@@ -624,6 +663,7 @@ def quant_post_static(executor,
             activation_quantize_type=activation_quantize_type,
             weight_quantize_type=weight_quantize_type,
             onnx_format=onnx_format,
+            skip_tensor_list=skip_tensor_list,
             optimize_model=optimize_model)
 
     post_training_quantization.quantize()
@@ -639,12 +679,7 @@ def quant_post_static(executor,
 quant_post = quant_post_static
 
 
-def convert(program,
-            place,
-            config=None,
-            scope=None,
-            save_int8=False,
-            save_clip_ranges_path='./'):
+def convert(program, place, config=None, scope=None, save_int8=False):
     """
     convert quantized and well-trained ``program`` to final  quantized
     ``program``that can be used to  save ``inference model``.
@@ -666,7 +701,6 @@ def convert(program,
         save_int8: Whether to return ``program`` which model parameters'
                 dtype is ``int8``. This parameter can only be used to
                 get model size. Default: ``False``.
-        save_clip_ranges_path: If config.onnx_format=True, quantization clip ranges will be saved locally.
 
     Returns:
         Tuple : freezed program which can be used for inference.
@@ -732,6 +766,9 @@ def convert(program,
             persistables.extend(_op.input('X'))
             _op.desc.set_input("X", persistables)
 
+    assert not (
+        save_int8 and config['onnx_format']
+    ), "When onnx_format=True, already saved int8 weight,so you can't set save_int8=True."
     if save_int8:
         convert_int8_pass = ConvertToInt8Pass(scope=scope, place=place)
         for sub_graph in test_graph.all_sub_graphs():

paddleslim/quant/reconstruction_quantization.py

@@ -180,10 +180,10 @@ class ReconstructionQuantization(PostTrainingQuantization):
         # save out_threshold for quantized ops.
         self._save_output_threshold()
 
-        if any(op_type in self._quantizable_op_type
+        if any(op_type in self.quant_config.activation_quant_operation_types
                for op_type in self._dynamic_quantize_op_type):
             self._collect_dynamic_quantize_op_threshold(
-                self._dynamic_quantize_op_type, )
+                self._dynamic_quantize_op_type)
 
         # Move sub blocks persistable var to global block
         global_block = self._program.global_block()

tests/test_quant_aware.py

@@ -11,6 +11,7 @@
 # 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 sys
 sys.path.append("../")
 import unittest
@@ -23,70 +24,7 @@ from layers import conv_bn_layer
 import numpy as np
 
 
-class TestQuantAwareCase1(StaticCase):
-    def get_model(self):
-        image = paddle.static.data(
-            name='image', shape=[None, 1, 28, 28], dtype='float32')
-        label = paddle.static.data(name='label', shape=[None, 1], dtype='int64')
-        model = MobileNet()
-        out = model.net(input=image, class_dim=10)
-        cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
-        avg_cost = paddle.mean(x=cost)
-        startup_prog = paddle.static.default_startup_program()
-        train_prog = paddle.static.default_main_program()
-        return startup_prog, train_prog
-
-    def get_op_number(self, prog):
-
-        graph = paddle.fluid.framework.IrGraph(
-            paddle.framework.core.Graph(prog.desc), for_test=False)
-        quant_op_nums = 0
-        op_nums = 0
-        for op in graph.all_op_nodes():
-            if op.name() in ['conv2d', 'depthwise_conv2d', 'mul']:
-                op_nums += 1
-            elif 'fake_' in op.name():
-                quant_op_nums += 1
-        return op_nums, quant_op_nums
-
-    def test_quant_op(self):
-        startup_prog, train_prog = self.get_model()
-        place = paddle.CUDAPlace(0) if paddle.is_compiled_with_cuda(
-        ) else paddle.CPUPlace()
-        exe = paddle.static.Executor(place)
-        exe.run(startup_prog)
-        config_1 = {
-            'weight_quantize_type': 'channel_wise_abs_max',
-            'activation_quantize_type': 'moving_average_abs_max',
-            'quantize_op_types': ['depthwise_conv2d', 'mul', 'conv2d'],
-        }
-
-        quant_prog_1 = quant_aware(
-            train_prog, place, config=config_1, for_test=True)
-        op_nums_1, quant_op_nums_1 = self.get_op_number(quant_prog_1)
-        convert_prog_1 = convert(quant_prog_1, place, config=config_1)
-        convert_op_nums_1, convert_quant_op_nums_1 = self.get_op_number(
-            convert_prog_1)
-
-        config_1['not_quant_pattern'] = ['last_fc']
-        quant_prog_2 = quant_aware(
-            train_prog, place, config=config_1, for_test=True)
-        op_nums_2, quant_op_nums_2 = self.get_op_number(quant_prog_2)
-        convert_prog_2 = convert(quant_prog_2, place, config=config_1)
-        convert_op_nums_2, convert_quant_op_nums_2 = self.get_op_number(
-            convert_prog_2)
-
-        self.assertTrue(op_nums_1 == op_nums_2)
-        # test quant_aware op numbers
-        self.assertTrue(op_nums_1 * 4 == quant_op_nums_1)
-        # test convert op numbers
-        self.assertTrue(convert_op_nums_1 * 2 == convert_quant_op_nums_1)
-        # test skip_quant
-        self.assertTrue(quant_op_nums_1 - 4 == quant_op_nums_2)
-        self.assertTrue(convert_quant_op_nums_1 - 2 == convert_quant_op_nums_2)
-
-
-class TestQuantAwareCase2(StaticCase):
+class TestQuantAwareCase(StaticCase):
     def test_accuracy(self):
         image = paddle.static.data(
             name='image', shape=[None, 1, 28, 28], dtype='float32')
@@ -103,7 +41,7 @@ class TestQuantAwareCase2(StaticCase):
             weight_decay=paddle.regularizer.L2Decay(4e-5))
         optimizer.minimize(avg_cost)
         main_prog = paddle.static.default_main_program()
-        val_prog = main_prog.clone(for_test=True)
+        val_prog = paddle.static.default_main_program().clone(for_test=True)
 
         place = paddle.CUDAPlace(0) if paddle.is_compiled_with_cuda(
         ) else paddle.CPUPlace()
@@ -173,14 +111,61 @@ class TestQuantAwareCase2(StaticCase):
         }
         quant_train_prog = quant_aware(main_prog, place, config, for_test=False)
         quant_eval_prog = quant_aware(val_prog, place, config, for_test=True)
+        op_nums_1, quant_op_nums_1 = self.get_op_number(quant_eval_prog)
+        # test quant_aware op numbers
+        self.assertTrue(op_nums_1 * 2 == quant_op_nums_1)
+
         train(quant_train_prog)
-        quant_eval_prog, int8_prog = convert(
-            quant_eval_prog, place, config, save_int8=True)
-        top1_2, top5_2 = test(quant_eval_prog)
+        convert_eval_prog = convert(quant_eval_prog, place, config)
+
+        top1_2, top5_2 = test(convert_eval_prog)
         # values before quantization and after quantization should be close
         print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))
         print("after quantization: top1: {}, top5: {}".format(top1_2, top5_2))
 
+        convert_op_nums_1, convert_quant_op_nums_1 = self.get_convert_op_number(
+            convert_eval_prog)
+        # test convert op numbers
+        self.assertTrue(convert_op_nums_1 + 25 == convert_quant_op_nums_1)
+
+        config['not_quant_pattern'] = ['last_fc']
+        quant_prog_2 = quant_aware(
+            main_prog, place, config=config, for_test=True)
+        op_nums_2, quant_op_nums_2 = self.get_op_number(quant_prog_2)
+        convert_prog_2 = convert(quant_prog_2, place, config=config)
+        convert_op_nums_2, convert_quant_op_nums_2 = self.get_convert_op_number(
+            convert_prog_2)
+
+        self.assertTrue(op_nums_1 == op_nums_2)
+        # test skip_quant
+        self.assertTrue(quant_op_nums_1 - 2 == quant_op_nums_2)
+        self.assertTrue(convert_quant_op_nums_1 == convert_quant_op_nums_2)
+
+    def get_op_number(self, prog):
+        graph = paddle.fluid.framework.IrGraph(
+            paddle.framework.core.Graph(prog.desc), for_test=False)
+        quant_op_nums = 0
+        op_nums = 0
+        for op in graph.all_op_nodes():
+            if op.name() in ['conv2d', 'depthwise_conv2d', 'mul']:
+                op_nums += 1
+            elif op.name() == 'quantize_linear':
+                quant_op_nums += 1
+        return op_nums, quant_op_nums
+
+    def get_convert_op_number(self, prog):
+        graph = paddle.fluid.framework.IrGraph(
+            paddle.framework.core.Graph(prog.desc), for_test=True)
+        quant_op_nums = 0
+        op_nums = 0
+        dequant_num = 0
+        for op in graph.all_op_nodes():
+            if op.name() not in ['quantize_linear', 'dequantize_linear']:
+                op_nums += 1
+            elif op.name() == 'quantize_linear':
+                quant_op_nums += 1
+        return op_nums, quant_op_nums
+
 
 if __name__ == '__main__':
     unittest.main()

tests/test_quant_aware_user_defined.py

@@ -153,8 +153,7 @@ class TestQuantAwareCase1(StaticCase):
 
         quant_eval_prog = quant_aware(val_prog, place, config, for_test=True)
         train(quant_train_prog_pact)
-        quant_eval_prog, int8_prog = convert(
-            quant_eval_prog, place, config, save_int8=True)
+        quant_eval_prog = convert(quant_eval_prog, place, config)
         top1_2, top5_2 = test(quant_eval_prog)
         # values before quantization and after quantization should be close
         print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))

tests/test_quant_aware_with_infermodel.py

@@ -130,7 +130,8 @@ class TestQuantAwareWithInferModelCase1(StaticCase):
             'weight_quantize_type': 'channel_wise_abs_max',
             'activation_quantize_type': 'moving_average_abs_max',
             'not_quant_pattern': ['skip_quant'],
-            'quantize_op_types': ['conv2d', 'depthwise_conv2d', 'mul']
+            'quantize_op_types': ['conv2d', 'depthwise_conv2d', 'mul'],
+            'onnx_format': False
         }
         train_config = {
             "num_epoch": 1,  # training epoch num

tests/test_quant_post_quant_aware.py

@@ -143,8 +143,7 @@ class TestQuantPostQuantAwareCase1(StaticCase):
             scale_dict=scale_dict,
             model_type='transformer')
         train(quant_train_prog)
-        quant_eval_prog, int8_prog = convert(
-            quant_eval_prog, place, config, save_int8=True)
+        quant_eval_prog = convert(quant_eval_prog, place, config)
         top1_2 = test(quant_eval_prog)
         # values before quantization and after quantization should be close
         print("before quantization: top1: {}".format(top1_1))