Квантизация PaddlePaddle

requirements_frameworks.txt

@@ -17,6 +17,7 @@ dglgo==0.0.2
 
 tflite
 
+paddleslim
 paddlepaddle==2.6.0
 --extra-index-url https://mirror.baidu.com/pypi/simple
 

src/configs/paddle_quantization_config_template.xml

@@ -0,0 +1,26 @@
+<?xml version="1.0" encoding="utf-8"?>
+<QuantizationConfigs>
+    <Config>
+        <Model>
+            <Name></Name>
+            <PathPrefix></PathPrefix>
+            <ModelDir></ModelDir>
+            <ModelFileName></ModelFileName>
+            <ParamsFileName></ParamsFileName>
+        </Model>
+        <Dataset>
+            <Name></Name>
+            <Path></Path>
+            <Mean></Mean>
+            <Std></Std>
+            <BatchSize></BatchSize>
+            <BatchNum></BatchNum>
+            <ResizeResolution></ResizeResolution>
+        </Dataset>
+        <QuantizationParameters>
+            <InputShape></InputShape>
+            <InputName></InputName>
+            <SaveDir></SaveDir>
+        </QuantizationParameters>
+    </Config>
+</QuantizationConfigs>

src/quantization/paddlepaddle/README.md

@@ -0,0 +1,40 @@
+# PaddlePaddle quantization script
+
+Script name:
+
+```bash
+quantization_paddlepaddle.py
+```
+
+Required arguments:
+
+- `-c / --config` is a path to the file containing information
+  about quantization process in the xml-format. Template of the configuration file
+  located [here][config_path].
+
+Description of parameters:
+
+`Model` contains information about model to be quantized:
+- `Name` is a name of the model.
+- `PathPrefix` is a path to the model files without the extensions (.pdmodel, .pdiparams).
+- `ModelDir` is a directory with the model.
+- `ModelFileName` is a file name of the model description.
+- `ParamsFileName` is a file name of the model parameters.
+
+`Dataset` contains information about dataset for the model calibration:
+- `Name` is a dataset name.
+- `Path` is a path to the directory that contains input data.
+- `Mean` is a mean value for preprocessing data.
+- `Std` is a scale value for preprocessing data.
+- `ResizeResolution` is an image size for preprocessing data. Example: 224, 224.
+- `BatchSize` is a batch size.
+- `BatchNum` is the total number of batches
+
+`QuantizationParameters` contains information about the model input layer:
+- `InputShape` is a shape of the model's input layer.
+- `InputName` is a name of the model's input layer.
+- `SaveDir` is a directory for the quantized model to be saved.
+
+
+<!-- LINKS -->
+[config_path]: ../../configs/paddle_quantization_config_template.xml

src/quantization/paddlepaddle/parameters.py

@@ -0,0 +1,138 @@
+import sys
+from pathlib import Path
+import random
+import numpy as np
+import paddle
+from paddle.io import Dataset
+import ast
+from paddle.io import DataLoader
+from paddleslim.quant import quant_post_static
+import importlib
+sys.path.append(str(Path(__file__).resolve().parents[1]))
+from utils import ArgumentsParser  # noqa: E402
+
+random.seed(0)
+np.random.seed(0)
+
+default_crop_scale = (0.08, 1.0)
+default_crop_ratio = (3. / 4., 4. / 3.)
+
+
+class PaddleDatasetReader(Dataset):
+    def __init__(self, args, log):
+        super(PaddleDatasetReader, self).__init__()
+        self.log = log
+        self.log.info('Parsing dataset arguments.')
+        self.cv2 = importlib.import_module('cv2')
+        self.data_dir = args['Path']
+
+        self.resize_size = ast.literal_eval(args['ResizeResolution'])
+        self.mean = np.array((np.asarray(ast.literal_eval(args['Mean']), dtype=np.float32)
+                              if args['Mean'] is not None else [0., 0., 0.])).reshape((3, 1, 1))
+        self.std = np.array((np.asarray(ast.literal_eval(args['Std']), dtype=np.float32)
+                             if args['Std'] is not None else [1., 1., 1.])).reshape((3, 1, 1))
+
+        self.batch_size = int(args['BatchSize'])
+        self.batch_num = int(args['BatchNum'])
+        self.dataset = list(Path(self.data_dir).glob('*'))
+        random.shuffle(self.dataset)
+        self.dataset_iter = iter(self.dataset)
+
+    def __getitem__(self, index):
+        image_path = str(self.dataset[index].absolute())
+        data = self.process_image(image_path)
+        return data
+
+    def __len__(self):
+        return len(self.dataset)
+
+    def process_image(self, image_path):
+
+        img = self.cv2.imread(image_path)
+        if img.size == 0:
+            self.log.info('failed to read:', image_path)
+            return None
+        img = self.cv2.resize(img, self.resize_size)
+
+        img = img.astype('float32').transpose((2, 0, 1)) / 255
+        img -= self.mean
+        img /= self.std
+
+        return img
+
+
+class PaddleQuantizationProcess:
+    def __init__(self, log, model_reader, dataset, quant_params):
+        self.log = log
+        self.model_reader = model_reader
+        self.dataset = dataset
+        self.quant_params = quant_params
+
+    def transform_fn(self):
+        for data in self.dataset:
+            yield [data.astype(np.float32)]
+
+    def quantization_paddle(self):
+        place = paddle.CPUPlace()
+        exe = paddle.static.Executor(place)
+        data_loader = DataLoader(
+            self.dataset,
+            places=place,
+            feed_list=[self.quant_params.image],
+            drop_last=False,
+            return_list=False,
+            batch_size=self.dataset.batch_size,
+            shuffle=False)
+
+        quant_post_static(
+            executor=exe,
+            model_dir=self.model_reader.model_dir,
+            quantize_model_path=self.quant_params.save_dir,
+            data_loader=data_loader,
+            model_filename=self.model_reader.model_filename,
+            params_filename=self.model_reader.params_filename,
+            batch_size=self.dataset.batch_size,
+            batch_nums=self.dataset.batch_num,
+            algo='avg',
+            round_type='round',
+            hist_percent=0.9999,
+            is_full_quantize=False,
+            bias_correction=False,
+            onnx_format=False)
+
+
+class PaddleModelReader(ArgumentsParser):
+    def __init__(self, log):
+        super().__init__(log)
+
+    def _get_arguments(self):
+        self._log.info('Parsing model arguments.')
+        self.model_name = self.args['Name']
+        self.path_prefix = self.args['PathPrefix']
+        self.model_dir = self.args['ModelDir']
+        self.model_filename = self.args['ModelFileName']
+        self.params_filename = self.args['ParamsFileName']
+
+    def dict_for_iter_log(self):
+        return {
+            'Name': self.model_name,
+            'Model path prefix': self.path_prefix,
+        }
+
+
+class PaddleQuantParamReader(ArgumentsParser):
+    def __init__(self, log):
+        super().__init__(log)
+
+    def dict_for_iter_log(self):
+        return {
+            'InputShape': self.input_shape,
+            'InputName': self.input_name,
+            'SaveDir': self.save_dir,
+        }
+
+    def _get_arguments(self):
+        self.input_shape = ast.literal_eval(self.args['InputShape'])
+        self.image = paddle.static.data(name=self.args['InputName'], shape=[None] + self.input_shape, dtype='float32')
+        self.input_name = self.args['InputName']
+        self.save_dir = self.args['SaveDir']

src/quantization/paddlepaddle/quantization_paddlepaddle.py

@@ -0,0 +1,55 @@
+import paddle
+import argparse
+import sys
+import traceback
+from pathlib import Path
+from parameters import PaddleModelReader, PaddleDatasetReader, PaddleQuantizationProcess, PaddleQuantParamReader
+sys.path.append(str(Path(__file__).resolve().parents[3]))
+from src.utils.logger_conf import configure_logger  # noqa: E402
+from src.quantization.utils import ConfigParser  # noqa: E402
+
+
+log = configure_logger()
+
+
+def cli_argument_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-c', '--config',
+                        help='Path to the configuration file in the xml-format.',
+                        type=str,
+                        required=True,
+                        dest='config')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = cli_argument_parser()
+    try:
+        log.info(f'Parsing the configuration file {args.config}')
+        parser = ConfigParser(args.config)
+        paddle.enable_static()
+        config = parser.parse()
+        exit_code = 0
+        quant_params = PaddleQuantParamReader(log)
+        model_reader = PaddleModelReader(log)
+        for model_quant_config in config:
+            try:
+                data_reader = PaddleDatasetReader(model_quant_config[1]['Dataset'], log)
+                model_reader.add_arguments(model_quant_config[0]['Model'])
+                quant_params.add_arguments(model_quant_config[2]['QuantizationParameters'])
+                proc = PaddleQuantizationProcess(log, model_reader, data_reader, quant_params)
+                proc.quantization_paddle()
+
+            except Exception:
+                log.error(traceback.format_exc())
+                exit_code += 1
+        if exit_code:
+            sys.exit(1)
+    except Exception:
+        log.error(traceback.format_exc())
+        sys.exit(1)
+
+
+if __name__ == '__main__':
+    sys.exit(main() or 0)

tests/smoke_test/configs/quantization_models/resnet-50_PADDLEPADDLE.xml

@@ -0,0 +1,26 @@
+<?xml version="1.0" encoding="utf-8"?>
+<QuantizationConfigs>
+    <Config>
+        <Model>
+            <Name>resnet50-paddle</Name>
+            <PathPrefix>../models_dir/resnet50_paddle/inference</PathPrefix>
+            <ModelDir>../models_dir/resnet50_paddle</ModelDir>
+            <ModelFileName>resnet50.pdmodel</ModelFileName>
+            <ParamsFileName>resnet50.pdiparams</ParamsFileName>
+        </Model>
+        <Dataset>
+            <Name>test</Name>
+            <Path>../test_images/classification_images</Path>
+            <Mean>[123.675, 116.28, 103.53]</Mean>
+            <Std>[58.395, 57.12, 57.375]</Std>
+            <BatchSize>1</BatchSize>
+            <BatchNum>10</BatchNum>
+            <ResizeResolution>[224, 224]</ResizeResolution>
+        </Dataset>
+        <QuantizationParameters>
+            <InputShape>[3, 224, 224]</InputShape>
+            <InputName>inputs</InputName>
+            <SaveDir>res_dir</SaveDir>
+        </QuantizationParameters>
+    </Config>
+</QuantizationConfigs>

tests/smoke_test/quantization_smoke/conftest.py

@@ -6,6 +6,7 @@
 from tests.smoke_test.utils import execute_process
 from tests.smoke_test.conftest import (SCRIPT_DIR, OUTPUT_DIR, log,
                                        download_models, convert_models)
+from tests.smoke_test.benchmark_smoke.conftest import download_resnet50_paddle
 
 QUANTIZATION_CONFIG_DIR_PATH = Path(SCRIPT_DIR, 'configs', 'quantization_models')
 TVM_CONVERTER = Path.joinpath(SCRIPT_DIR.parents[1], 'src/model_converters/tvm_converter/tvm_converter.py')
@@ -43,6 +44,7 @@ def prepare_dl_models(request, overrided_models):
     models_per_mark = DL_MODELS
     enabled_models = overrided_models if overrided_models else models_per_mark
 
+    download_resnet50_paddle()
     download_models(models_list=enabled_models)
     convert_models(models_list=enabled_models)
     convert_models_to_tvm()

tests/smoke_test/quantization_smoke/test_quantization_smoke.py

@@ -6,6 +6,8 @@
 QUANTIZATION_TFLITE = Path.joinpath(SCRIPT_DIR.parents[1], 'src/quantization/tflite/quantization_tflite.py')
 QUANTIZATION_TVM = Path.joinpath(SCRIPT_DIR.parents[1], 'src/quantization/tvm/quantization_tvm.py')
 QUANTIZATION_NNCF = Path.joinpath(SCRIPT_DIR.parents[1], 'src/quantization/nncf/quantization_nncf.py')
+QUANTIZATION_PADDLE = Path.joinpath(SCRIPT_DIR.parents[1],
+                                    'src/quantization/paddlepaddle/quantization_paddlepaddle.py')
 
 TVM_CONVERTER = Path.joinpath(SCRIPT_DIR.parents[1], 'src/model_converters/tvm_converter/tvm_converter.py')
 
@@ -17,6 +19,8 @@ def test_smoke_dl_models(test_configuration):
         command_line = (f'python3 {QUANTIZATION_TFLITE} -c {test_configuration.config_path}')
     elif test_configuration.framework == 'TVM':
         command_line = (f'python3 {QUANTIZATION_TVM} -c {test_configuration.config_path}')
+    elif test_configuration.framework == 'PADDLEPADDLE':
+        command_line = (f'python3 {QUANTIZATION_PADDLE} -c {test_configuration.config_path}')
     else:
         raise Exception(f'Unsupported framework: {test_configuration.framework}')