fix quant online bug (#268)

contrib/HumanSeg/models/humanseg.py

@@ -25,6 +25,7 @@ import time
 import tqdm
 import cv2
 import yaml
+import shutil
 import paddleslim as slim
 
 import utils
@@ -243,30 +244,11 @@ class SegModel(object):
 
         if self.status == 'Normal':
             fluid.save(self.train_prog, osp.join(save_dir, 'model'))
+            model_info['status'] = 'Normal'
         elif self.status == 'Quant':
-            float_prog, _ = slim.quant.convert(
-                self.test_prog, self.exe.place, save_int8=True)
-            test_input_names = [
-                var.name for var in list(self.test_inputs.values())
-            ]
-            test_outputs = list(self.test_outputs.values())
-            fluid.io.save_inference_model(
-                dirname=save_dir,
-                executor=self.exe,
-                params_filename='__params__',
-                feeded_var_names=test_input_names,
-                target_vars=test_outputs,
-                main_program=float_prog)
+            fluid.save(self.test_prog, osp.join(save_dir, 'model'))
+            model_info['status'] = 'QuantOnline'
 
-            model_info['_ModelInputsOutputs'] = dict()
-            model_info['_ModelInputsOutputs']['test_inputs'] = [
-                [k, v.name] for k, v in self.test_inputs.items()
-            ]
-            model_info['_ModelInputsOutputs']['test_outputs'] = [
-                [k, v.name] for k, v in self.test_outputs.items()
-            ]
-
-        model_info['status'] = self.status
         with open(
                 osp.join(save_dir, 'model.yml'), encoding='utf-8',
                 mode='w') as f:
@@ -308,40 +290,57 @@ class SegModel(object):
         logging.info("Model for inference deploy saved in {}.".format(save_dir))
 
     def export_quant_model(self,
-                           dataset,
-                           save_dir,
+                           dataset=None,
+                           save_dir=None,
                            batch_size=1,
                            batch_nums=10,
-                           cache_dir="./.temp"):
-        self.arrange_transform(transforms=dataset.transforms, mode='quant')
-        dataset.num_samples = batch_size * batch_nums
-        try:
-            from utils import HumanSegPostTrainingQuantization
-        except:
-            raise Exception(
-                "Model Quantization is not available, try to upgrade your paddlepaddle>=1.7.0"
-            )
-        is_use_cache_file = True
-        if cache_dir is None:
-            is_use_cache_file = False
-        post_training_quantization = HumanSegPostTrainingQuantization(
-            executor=self.exe,
-            dataset=dataset,
-            program=self.test_prog,
-            inputs=self.test_inputs,
-            outputs=self.test_outputs,
-            batch_size=batch_size,
-            batch_nums=batch_nums,
-            scope=None,
-            algo='KL',
-            quantizable_op_type=["conv2d", "depthwise_conv2d", "mul"],
-            is_full_quantize=False,
-            is_use_cache_file=is_use_cache_file,
-            cache_dir=cache_dir)
-        post_training_quantization.quantize()
-        post_training_quantization.save_quantized_model(save_dir)
-        if cache_dir is not None:
-            os.system('rm -r' + cache_dir)
+                           cache_dir=".temp",
+                           quant_type="offline"):
+        if quant_type == "offline":
+            self.arrange_transform(transforms=dataset.transforms, mode='quant')
+            dataset.num_samples = batch_size * batch_nums
+            try:
+                from utils import HumanSegPostTrainingQuantization
+            except:
+                raise Exception(
+                    "Model Quantization is not available, try to upgrade your paddlepaddle>=1.8.1"
+                )
+            is_use_cache_file = True
+            if cache_dir is None:
+                is_use_cache_file = False
+            post_training_quantization = HumanSegPostTrainingQuantization(
+                executor=self.exe,
+                dataset=dataset,
+                program=self.test_prog,
+                inputs=self.test_inputs,
+                outputs=self.test_outputs,
+                batch_size=batch_size,
+                batch_nums=batch_nums,
+                scope=None,
+                algo='KL',
+                quantizable_op_type=["conv2d", "depthwise_conv2d", "mul"],
+                is_full_quantize=False,
+                is_use_cache_file=is_use_cache_file,
+                cache_dir=cache_dir)
+            post_training_quantization.quantize()
+            post_training_quantization.save_quantized_model(save_dir)
+            if cache_dir is not None:
+                shutil.rmtree(cache_dir)
+        else:
+            float_prog, _ = slim.quant.convert(
+                self.test_prog, self.exe.place, save_int8=True)
+            test_input_names = [
+                var.name for var in list(self.test_inputs.values())
+            ]
+            test_outputs = list(self.test_outputs.values())
+            fluid.io.save_inference_model(
+                dirname=save_dir,
+                executor=self.exe,
+                params_filename='__params__',
+                feeded_var_names=test_input_names,
+                target_vars=test_outputs,
+                main_program=float_prog)
+
         model_info = self.get_model_info()
         model_info['status'] = 'Quant'
 
@@ -593,6 +592,16 @@ class SegModel(object):
                         'Current evaluated best model in eval_dataset is epoch_{}, miou={}'
                         .format(best_model_epoch, best_miou))
 
+        if quant:
+            if osp.exists(osp.join(save_dir, "best_model")):
+                fluid.load(
+                    program=self.test_prog,
+                    model_path=osp.join(save_dir, "best_model"),
+                    executor=self.exe)
+            self.export_quant_model(
+                save_dir=osp.join(save_dir, "best_model_export"),
+                quant_type="online")
+
     def evaluate(self, eval_dataset, batch_size=1, epoch_id=None):
         """评估。
 

contrib/HumanSeg/models/load_model.py

@@ -34,7 +34,7 @@ def load_model(model_dir):
         raise Exception("There's no attribute {} in models".format(
             info['Model']))
     model = getattr(models, info['Model'])(**info['_init_params'])
-    if status == "Normal":
+    if status in ["Normal", "QuantOnline"]:
         startup_prog = fluid.Program()
         model.test_prog = fluid.Program()
         with fluid.program_guard(model.test_prog, startup_prog):
@@ -42,11 +42,16 @@ def load_model(model_dir):
                 model.test_inputs, model.test_outputs = model.build_net(
                     mode='test')
         model.test_prog = model.test_prog.clone(for_test=True)
+        if status == "QuantOnline":
+            print('test quant online')
+            import paddleslim as slim
+            model.test_prog = slim.quant.quant_aware(
+                model.test_prog, model.exe.place, for_test=True)
         model.exe.run(startup_prog)
-        import pickle
-        with open(osp.join(model_dir, 'model.pdparams'), 'rb') as f:
-            load_dict = pickle.load(f)
-        fluid.io.set_program_state(model.test_prog, load_dict)
+        fluid.load(model.test_prog, osp.join(model_dir, 'model'))
+        if status == "QuantOnline":
+            model.test_prog = slim.quant.convert(model.test_prog,
+                                                 model.exe.place)
 
     elif status in ['Infer', 'Quant']:
         [prog, input_names, outputs] = fluid.io.load_inference_model(

deploy/python/README.md

@@ -44,7 +44,7 @@ yum install -y libXext libSM libXrender
 ### 5.1 准备模型
 请使用[模型导出工具](../../docs/model_export.md) 导出您的模型, 或点击下载我们的[人像分割样例模型](https://bj.bcebos.com/paddleseg/inference/human_freeze_model.zip)用于测试。
 
-模型导出的目录通常包括三个文件: 
+模型导出的目录通常包括三个文件:
 ```
 ├── model #  模型文件
 ├── params # 参数文件
@@ -79,7 +79,7 @@ DEPLOY:
 ### 5.2 执行预测程序
 在终端输入以下命令进行预测:
 ```bash
-python infer.py --conf=/path/to/deploy.yaml --input_dir/path/to/images_directory --use_pr=False
+python infer.py --conf=/path/to/deploy.yaml --input_dir/path/to/images_directory
 ```
 参数说明如下:
 
@@ -87,9 +87,6 @@ python infer.py --conf=/path/to/deploy.yaml --input_dir/path/to/images_directory
 |-------|-------|----------|
 | conf | Yes|模型配置的Yaml文件路径 |
 | input_dir |Yes| 需要预测的图片目录 |
-| use_pr |No|是否使用优化模型，默认为False|
-
-* 优化模型：使用`PaddleSeg 0.3.0`版导出的为优化模型, 此前版本导出的模型即为未优化版本。优化模型把图像的预处理以及后处理部分融入到模型网络中使用`GPU` 完成，相比原来`CPU` 中的处理提升了计算性能。
 
 **注意**: 如果硬件支持且安装的是从源码编译集成`TensorRT`的`PaddlePaddle`, 则可以使用参数`--trt_mode=fp16` 表示开启`FP16` 精度优化, 使用`trt_mode=fp32` 表示使用`FP32` 精度。
 

pdseg/export_model.py

@@ -53,7 +53,7 @@ def parse_args():
 def export_inference_config():
     deploy_cfg = '''DEPLOY:
         USE_GPU : 1
-        USE_PR : 1
+        USE_PR : 0
         MODEL_PATH : "%s"
         MODEL_FILENAME : "%s"
         PARAMS_FILENAME : "%s"