update act paddle inference demo (#1432)

Co-authored-by: Nceci3 <ceci3@users.noreply.github.com>

example/auto_compression/detection/README.md

@@ -7,8 +7,7 @@
   - [3.1 环境准备](#31-准备环境)
   - [3.2 准备数据集](#32-准备数据集)
   - [3.3 准备预测模型](#33-准备预测模型)
-  - [3.4 测试模型精度](#34-测试模型精度)
-  - [3.5 自动压缩并产出模型](#35-自动压缩并产出模型)
+  - [3.4 自动压缩并产出模型](#34-自动压缩并产出模型)
 - [4.预测部署](#4预测部署)
 - [5.FAQ](5FAQ)
 
@@ -110,23 +109,52 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch --log_dir=log -
           --config_path=./configs/ppyoloe_l_qat_dis.yaml --save_dir='./output/'
 ```
 
-#### 3.5 测试模型精度
 
-使用eval.py脚本得到模型的mAP：
-```
-export CUDA_VISIBLE_DEVICES=0
-python eval.py --config_path=./configs/ppyoloe_l_qat_dis.yaml
-```
+## 4.预测部署
 
-**注意**：
-- 要测试的模型路径可以在配置文件中`model_dir`字段下进行修改。
+#### 4.1 Paddle Inference 验证性能
 
-## 4.预测部署
+量化模型在GPU上可以使用TensorRT进行加速，在CPU上可以使用MKLDNN进行加速。
+
+以下字段用于配置预测参数：
 
-- 如果模型包含NMS，可以参考[PaddleDetection部署教程](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.4/deploy)，GPU上量化模型开启TensorRT并设置trt_int8模式进行部署。
+| 参数名 | 含义 |
+|:------:|:------:|
+| model_path | inference 模型文件所在目录，该目录下需要有文件 model.pdmodel 和 model.pdiparams 两个文件 |
+| reader_config | eval时模型reader的配置文件路径 |
+| image_file | 如果只测试单张图片效果，直接根据image_file指定图片路径 |
+| device | 使用GPU或者CPU预测，可选CPU/GPU   |
+| use_trt | 是否使用 TesorRT 预测引擎   |
+| use_mkldnn | 是否启用```MKL-DNN```加速库，注意```use_mkldnn```与```use_gpu```同时为```True```时，将忽略```enable_mkldnn```，而使用```GPU```预测  |
+| cpu_threads | CPU预测时，使用CPU线程数量，默认10  |
+| precision | 预测精度，包括`fp32/fp16/int8`  |
 
-- 模型为PPYOLOE，同时不包含NMS，使用以下预测demo进行部署：
-  - Paddle-TensorRT C++部署
+
+- TensorRT预测：
+
+环境配置：如果使用 TesorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
+
+```shell
+python paddle_inference_eval.py \
+      --model_path=models/ppyoloe_crn_l_300e_coco_quant \
+      --reader_config=configs/yoloe_reader.yml \
+      --use_trt=True \
+      --precision=int8
+```
+
+- MKLDNN预测：
+
+```shell
+python paddle_inference_eval.py \
+      --model_path=models/ppyoloe_crn_l_300e_coco_quant \
+      --reader_config=configs/yoloe_reader.yml \
+      --device=CPU \
+      --use_mkldnn=True \
+      --cpu_threads=10 \
+      --precision=int8
+```
+
+- 模型为PPYOLOE，同时不包含NMS，可以使用C++预测demo进行测速：
 
   进入[cpp_infer](./cpp_infer_ppyoloe)文件夹内，请按照[C++ TensorRT Benchmark测试教程](./cpp_infer_ppyoloe/README.md)进行准备环境及编译，然后开始测试：
   ```shell
@@ -136,14 +164,6 @@ python eval.py --config_path=./configs/ppyoloe_l_qat_dis.yaml
   ./build/trt_run --model_file ppyoloe_s_quant/model.pdmodel --params_file ppyoloe_s_quant/model.pdiparams --run_mode=trt_int8
   ```
 
-  - Paddle-TensorRT Python部署:
-
-  首先安装带有TensorRT的[Paddle安装包](https://www.paddlepaddle.org.cn/inference/v2.3/user_guides/download_lib.html#python)。然后使用[paddle_trt_infer.py](./paddle_trt_infer.py)进行部署：
-  ```shell
-  python paddle_trt_infer.py --model_path=output --image_file=images/000000570688.jpg --benchmark=True --run_mode=trt_int8
-  ```
-
 ## 5.FAQ
 
-
 - 如果想对模型进行离线量化，可进入[Detection模型离线量化示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/post_training_quantization/detection)中进行实验。

example/auto_compression/detection/paddle_trt_infer.py → example/auto_compression/detection/paddle_inference_eval.py

@@ -13,16 +13,81 @@
 # limitations under the License.
 
 import os
-import cv2
-import numpy as np
 import argparse
 import time
+import sys
+import cv2
+import numpy as np
 
+import paddle
 from paddle.inference import Config
 from paddle.inference import create_predictor
+from ppdet.core.workspace import load_config, create
+from ppdet.metrics import COCOMetric
 
 from post_process import PPYOLOEPostProcess
 
+
+def argsparser():
+    """
+    argsparser func
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_path", type=str, help="inference model filepath")
+    parser.add_argument(
+        "--image_file",
+        type=str,
+        default=None,
+        help="image path, if set image_file, it will not eval coco.")
+    parser.add_argument(
+        "--reader_config",
+        type=str,
+        default=None,
+        help="path of datset and reader config.")
+    parser.add_argument(
+        "--benchmark",
+        type=bool,
+        default=False,
+        help="Whether run benchmark or not.")
+    parser.add_argument(
+        "--use_trt",
+        type=bool,
+        default=False,
+        help="Whether use TensorRT or not.")
+    parser.add_argument(
+        "--precision",
+        type=str,
+        default="paddle",
+        help="mode of running(fp32/fp16/int8)")
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="GPU",
+        help="Choose the device you want to run, it can be: CPU/GPU/XPU, default is GPU",
+    )
+    parser.add_argument(
+        "--use_dynamic_shape",
+        type=bool,
+        default=True,
+        help="Whether use dynamic shape or not.")
+    parser.add_argument(
+        "--use_mkldnn",
+        type=bool,
+        default=False,
+        help="Whether use mkldnn or not.")
+    parser.add_argument(
+        "--cpu_threads", type=int, default=10, help="Num of cpu threads.")
+    parser.add_argument("--img_shape", type=int, default=640, help="input_size")
+    parser.add_argument(
+        '--include_nms',
+        type=bool,
+        default=True,
+        help="Whether include nms or not.")
+
+    return parser
+
+
 CLASS_LABEL = [
     'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train',
     'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign',
@@ -67,6 +132,9 @@ def generate_scale(im, target_shape, keep_ratio=True):
 
 
 def image_preprocess(img_path, target_shape):
+    """
+    image_preprocess func
+    """
     img = cv2.imread(img_path)
     im_scale_y, im_scale_x = generate_scale(img, target_shape, keep_ratio=False)
     img = cv2.resize(
@@ -84,14 +152,17 @@ def image_preprocess(img_path, target_shape):
 
 
 def get_color_map_list(num_classes):
+    """
+    get_color_map_list func
+    """
     color_map = num_classes * [0, 0, 0]
     for i in range(0, num_classes):
         j = 0
         lab = i
         while lab:
-            color_map[i * 3] |= (((lab >> 0) & 1) << (7 - j))
-            color_map[i * 3 + 1] |= (((lab >> 1) & 1) << (7 - j))
-            color_map[i * 3 + 2] |= (((lab >> 2) & 1) << (7 - j))
+            color_map[i * 3] |= ((lab >> 0) & 1) << (7 - j)
+            color_map[i * 3 + 1] |= ((lab >> 1) & 1) << (7 - j)
+            color_map[i * 3 + 2] |= ((lab >> 2) & 1) << (7 - j)
             j += 1
             lab >>= 3
     color_map = [color_map[i:i + 3] for i in range(0, len(color_map), 3)]
@@ -99,6 +170,9 @@ def get_color_map_list(num_classes):
 
 
 def draw_box(image_file, results, class_label, threshold=0.5):
+    """
+    draw_box func
+    """
     srcimg = cv2.imread(image_file, 1)
     for i in range(len(results)):
         color_list = get_color_map_list(len(class_label))
@@ -114,115 +188,142 @@ def draw_box(image_file, results, class_label, threshold=0.5):
         color = tuple(clsid2color[classid])
 
         cv2.rectangle(srcimg, (xmin, ymin), (xmax, ymax), color, thickness=2)
-        print(class_label[classid] + ': ' + str(round(conf, 3)))
+        print(class_label[classid] + ": " + str(round(conf, 3)))
         cv2.putText(
             srcimg,
-            class_label[classid] + ':' + str(round(conf, 3)), (xmin, ymin - 10),
+            class_label[classid] + ":" + str(round(conf, 3)),
+            (xmin, ymin - 10),
             cv2.FONT_HERSHEY_SIMPLEX,
-            0.8, (0, 255, 0),
-            thickness=2)
+            0.8,
+            (0, 255, 0),
+            thickness=2, )
     return srcimg
 
 
-def load_predictor(model_dir,
-                   run_mode='paddle',
+def load_predictor(
+        model_dir,
+        precision="fp32",
+        use_trt=False,
+        use_mkldnn=False,
         batch_size=1,
-                   device='CPU',
+        device="CPU",
         min_subgraph_size=3,
         use_dynamic_shape=False,
-                   trt_min_shape=3,
+        trt_min_shape=1,
         trt_max_shape=1280,
         trt_opt_shape=640,
-                   trt_calib_mode=False,
-                   cpu_threads=1,
-                   enable_mkldnn=False,
-                   enable_mkldnn_bfloat16=False,
-                   delete_shuffle_pass=False):
+        cpu_threads=1, ):
     """set AnalysisConfig, generate AnalysisPredictor
     Args:
         model_dir (str): root path of __model__ and __params__
-        device (str): Choose the device you want to run, it can be: CPU/GPU/XPU, default is CPU
-        run_mode (str): mode of running(paddle/trt_fp32/trt_fp16/trt_int8)
+        precision (str): mode of running(fp32/fp16/int8)
+        use_trt (bool): whether use TensorRT or not.
+        use_mkldnn (bool): whether use MKLDNN or not in CPU.
+        device (str): Choose the device you want to run, it can be: CPU/GPU, default is CPU
         use_dynamic_shape (bool): use dynamic shape or not
         trt_min_shape (int): min shape for dynamic shape in trt
         trt_max_shape (int): max shape for dynamic shape in trt
         trt_opt_shape (int): opt shape for dynamic shape in trt
-        trt_calib_mode (bool): If the model is produced by TRT offline quantitative
-            calibration, trt_calib_mode need to set True
-        delete_shuffle_pass (bool): whether to remove shuffle_channel_detect_pass in TensorRT. 
-                                    Used by action model.
     Returns:
         predictor (PaddlePredictor): AnalysisPredictor
     Raises:
         ValueError: predict by TensorRT need device == 'GPU'.
     """
-    if device != 'GPU' and run_mode != 'paddle':
+    rerun_flag = False
+    if device != "GPU" and use_trt:
         raise ValueError(
-            "Predict by TensorRT mode: {}, expect device=='GPU', but device == {}"
-            .format(run_mode, device))
+            "Predict by TensorRT mode: {}, expect device=='GPU', but device == {}".
+            format(precision, device))
     config = Config(
-        os.path.join(model_dir, 'model.pdmodel'),
-        os.path.join(model_dir, 'model.pdiparams'))
-    if device == 'GPU':
+        os.path.join(model_dir, "model.pdmodel"),
+        os.path.join(model_dir, "model.pdiparams"))
+    if device == "GPU":
         # initial GPU memory(M), device ID
         config.enable_use_gpu(200, 0)
         # optimize graph and fuse op
         config.switch_ir_optim(True)
-    elif device == 'XPU':
-        config.enable_lite_engine()
-        config.enable_xpu(10 * 1024 * 1024)
     else:
         config.disable_gpu()
         config.set_cpu_math_library_num_threads(cpu_threads)
-        if enable_mkldnn:
-            try:
-                # cache 10 different shapes for mkldnn to avoid memory leak
-                config.set_mkldnn_cache_capacity(10)
+        config.switch_ir_optim()
+        if use_mkldnn:
             config.enable_mkldnn()
-                if enable_mkldnn_bfloat16:
-                    config.enable_mkldnn_bfloat16()
-            except Exception as e:
-                print(
-                    "The current environment does not support `mkldnn`, so disable mkldnn."
-                )
-                pass
+            if precision == "int8":
+                config.enable_mkldnn_int8(
+                    {"conv2d", "depthwise_conv2d", "transpose2", "pool2d"})
 
     precision_map = {
-        'trt_int8': Config.Precision.Int8,
-        'trt_fp32': Config.Precision.Float32,
-        'trt_fp16': Config.Precision.Half
+        "int8": Config.Precision.Int8,
+        "fp32": Config.Precision.Float32,
+        "fp16": Config.Precision.Half,
     }
-    if run_mode in precision_map.keys():
+    if precision in precision_map.keys() and use_trt:
         config.enable_tensorrt_engine(
             workspace_size=(1 << 25) * batch_size,
             max_batch_size=batch_size,
             min_subgraph_size=min_subgraph_size,
-            precision_mode=precision_map[run_mode],
-            use_static=False,
-            use_calib_mode=trt_calib_mode)
+            precision_mode=precision_map[precision],
+            use_static=True,
+            use_calib_mode=False, )
 
         if use_dynamic_shape:
-            min_input_shape = {
-                'image': [batch_size, 3, trt_min_shape, trt_min_shape]
-            }
-            max_input_shape = {
-                'image': [batch_size, 3, trt_max_shape, trt_max_shape]
-            }
-            opt_input_shape = {
-                'image': [batch_size, 3, trt_opt_shape, trt_opt_shape]
-            }
-            config.set_trt_dynamic_shape_info(min_input_shape, max_input_shape,
-                                              opt_input_shape)
-            print('trt set dynamic shape done!')
+            dynamic_shape_file = os.path.join(FLAGS.model_path,
+                                              "dynamic_shape.txt")
+            if os.path.exists(dynamic_shape_file):
+                config.enable_tuned_tensorrt_dynamic_shape(dynamic_shape_file,
+                                                           True)
+                print("trt set dynamic shape done!")
+            else:
+                config.collect_shape_range_info(dynamic_shape_file)
+                print("Start collect dynamic shape...")
+                rerun_flag = True
 
     # enable shared memory
     config.enable_memory_optim()
-    # disable feed, fetch OP, needed by zero_copy_run
-    config.switch_use_feed_fetch_ops(False)
-    if delete_shuffle_pass:
-        config.delete_pass("shuffle_channel_detect_pass")
     predictor = create_predictor(config)
-    return predictor
+    return predictor, rerun_flag
+
+
+def get_current_memory_mb():
+    """
+    It is used to Obtain the memory usage of the CPU and GPU during the running of the program.
+    And this function Current program is time-consuming.
+    """
+    try:
+        pkg.require('pynvml')
+    except:
+        from pip._internal import main
+        main(['install', 'pynvml'])
+    try:
+        pkg.require('psutil')
+    except:
+        from pip._internal import main
+        main(['install', 'psutil'])
+    try:
+        pkg.require('GPUtil')
+    except:
+        from pip._internal import main
+        main(['install', 'GPUtil'])
+    import pynvml
+    import psutil
+    import GPUtil
+
+    gpu_id = int(os.environ.get("CUDA_VISIBLE_DEVICES", 0))
+
+    pid = os.getpid()
+    p = psutil.Process(pid)
+    info = p.memory_full_info()
+    cpu_mem = info.uss / 1024.0 / 1024.0
+    gpu_mem = 0
+    gpu_percent = 0
+    gpus = GPUtil.getGPUs()
+    if gpu_id is not None and len(gpus) > 0:
+        gpu_percent = gpus[gpu_id].load
+        pynvml.nvmlInit()
+        handle = pynvml.nvmlDeviceGetHandleByIndex(0)
+        meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
+        gpu_mem = meminfo.used / 1024.0 / 1024.0
+    return round(cpu_mem, 4), round(gpu_mem, 4)
 
 
 def predict_image(predictor,
@@ -230,15 +331,17 @@ def predict_image(predictor,
                   image_shape=[640, 640],
                   warmup=1,
                   repeats=1,
-                  threshold=0.5,
-                  include_nms=True):
+                  threshold=0.5):
+    """
+    predict image main func
+    """
     img, scale_factor = image_preprocess(image_file, image_shape)
     inputs = {}
-    inputs['image'] = img
+    inputs["image"] = img
     if include_nms:
         inputs['scale_factor'] = scale_factor
     input_names = predictor.get_input_names()
-    for i in range(len(input_names)):
+    for i, _ in enumerate(input_names):
         input_tensor = predictor.get_input_handle(input_names[i])
         input_tensor.copy_from_cpu(inputs[input_names[i]])
 
@@ -246,16 +349,17 @@ def predict_image(predictor,
         predictor.run()
 
     np_boxes, np_boxes_num = None, None
-    predict_time = 0.
+    cpu_mems, gpu_mems = 0, 0
+    predict_time = 0.0
     time_min = float("inf")
-    time_max = float('-inf')
+    time_max = float("-inf")
     for i in range(repeats):
         start_time = time.time()
         predictor.run()
         output_names = predictor.get_output_names()
         boxes_tensor = predictor.get_output_handle(output_names[0])
         np_boxes = boxes_tensor.copy_to_cpu()
-        if include_nms:
+        if FLAGS.include_nms:
             boxes_num = predictor.get_output_handle(output_names[1])
             np_boxes_num = boxes_num.copy_to_cpu()
         end_time = time.time()
@@ -263,61 +367,132 @@ def predict_image(predictor,
         time_min = min(time_min, timed)
         time_max = max(time_max, timed)
         predict_time += timed
+        cpu_mem, gpu_mem = get_current_memory_mb()
+        cpu_mems += cpu_mem
+        gpu_mems += gpu_mem
 
     time_avg = predict_time / repeats
-    print('Inference time(ms): min={}, max={}, avg={}'.format(
+    print("[Benchmark]Avg cpu_mem:{} MB, avg gpu_mem: {} MB".format(
+        cpu_mems / repeats, gpu_mems / repeats))
+    print("[Benchmark]Inference time(ms): min={}, max={}, avg={}".format(
         round(time_min * 1000, 2),
         round(time_max * 1000, 1), round(time_avg * 1000, 1)))
-    if not include_nms:
+    if not FLAGS.include_nms:
         postprocess = PPYOLOEPostProcess(score_threshold=0.3, nms_threshold=0.6)
         res = postprocess(np_boxes, scale_factor)
     else:
         res = {'bbox': np_boxes, 'bbox_num': np_boxes_num}
     res_img = draw_box(
-        image_file, res['bbox'], CLASS_LABEL, threshold=threshold)
-    cv2.imwrite('result.jpg', res_img)
+        image_file, res["bbox"], CLASS_LABEL, threshold=threshold)
+    cv2.imwrite("result.jpg", res_img)
 
 
-if __name__ == '__main__':
+def eval(predictor, val_loader, metric, rerun_flag=False):
+    """
+    eval main func
+    """
+    cpu_mems, gpu_mems = 0, 0
+    predict_time = 0.0
+    time_min = float("inf")
+    time_max = float("-inf")
+    sample_nums = len(val_loader)
+    input_names = predictor.get_input_names()
+    output_names = predictor.get_output_names()
+    boxes_tensor = predictor.get_output_handle(output_names[0])
+    boxes_num = predictor.get_output_handle(output_names[1])
+    for batch_id, data in enumerate(val_loader):
+        data_all = {k: np.array(v) for k, v in data.items()}
+        for i, _ in enumerate(input_names):
+            input_tensor = predictor.get_input_handle(input_names[i])
+            input_tensor.copy_from_cpu(data_all[input_names[i]])
+        start_time = time.time()
+        predictor.run()
+        np_boxes = boxes_tensor.copy_to_cpu()
+        if FLAGS.include_nms:
+            np_boxes_num = boxes_num.copy_to_cpu()
+        if rerun_flag:
+            return
+        end_time = time.time()
+        timed = end_time - start_time
+        time_min = min(time_min, timed)
+        time_max = max(time_max, timed)
+        predict_time += timed
+        cpu_mem, gpu_mem = get_current_memory_mb()
+        cpu_mems += cpu_mem
+        gpu_mems += gpu_mem
+        if not FLAGS.include_nms:
+            postprocess = PPYOLOEPostProcess(
+                score_threshold=0.3, nms_threshold=0.6)
+            res = postprocess(np_boxes, data_all['scale_factor'])
+        else:
+            res = {'bbox': np_boxes, 'bbox_num': np_boxes_num}
+        metric.update(data_all, res)
+        if batch_id % 100 == 0:
+            print("Eval iter:", batch_id)
+            sys.stdout.flush()
+    metric.accumulate()
+    metric.log()
+    map_res = metric.get_results()
+    metric.reset()
+    time_avg = predict_time / sample_nums
+    print("[Benchmark]Avg cpu_mem:{} MB, avg gpu_mem: {} MB".format(
+        cpu_mems / sample_nums, gpu_mems / sample_nums))
+    print("[Benchmark]Inference time(ms): min={}, max={}, avg={}".format(
+        round(time_min * 1000, 2),
+        round(time_max * 1000, 1), round(time_avg * 1000, 1)))
+    print("[Benchmark] COCO mAP: {}".format(map_res["bbox"][0]))
+    sys.stdout.flush()
 
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        '--image_file', type=str, default=None, help="image path")
-    parser.add_argument(
-        '--model_path', type=str, help="inference model filepath")
-    parser.add_argument(
-        '--benchmark',
-        type=bool,
-        default=False,
-        help="Whether run benchmark or not.")
-    parser.add_argument(
-        '--run_mode',
-        type=str,
-        default='paddle',
-        help="mode of running(paddle/trt_fp32/trt_fp16/trt_int8)")
-    parser.add_argument(
-        '--device',
-        type=str,
-        default='GPU',
-        help="Choose the device you want to run, it can be: CPU/GPU/XPU, default is GPU"
-    )
-    parser.add_argument('--img_shape', type=int, default=640, help="input_size")
-    parser.add_argument(
-        '--include_nms',
-        type=bool,
-        default=True,
-        help="Whether include nms or not.")
-    args = parser.parse_args()
 
-    predictor = load_predictor(
-        args.model_path, run_mode=args.run_mode, device=args.device)
+def main():
+    """
+    main func
+    """
+    predictor, rerun_flag = load_predictor(
+        FLAGS.model_path,
+        device=FLAGS.device,
+        use_trt=FLAGS.use_trt,
+        use_mkldnn=FLAGS.use_mkldnn,
+        precision=FLAGS.precision,
+        use_dynamic_shape=FLAGS.use_dynamic_shape,
+        cpu_threads=FLAGS.cpu_threads)
+
+    if FLAGS.image_file:
         warmup, repeats = 1, 1
-    if args.benchmark:
+        if FLAGS.benchmark:
             warmup, repeats = 50, 100
         predict_image(
             predictor,
-        args.image_file,
-        image_shape=[args.img_shape, args.img_shape],
+            FLAGS.image_file,
+            image_shape=[FLAGS.img_shape, FLAGS.img_shape],
             warmup=warmup,
-        repeats=repeats,
-        include_nms=args.include_nms)
+            repeats=repeats)
+    else:
+        reader_cfg = load_config(FLAGS.reader_config)
+
+        dataset = reader_cfg["EvalDataset"]
+        global val_loader
+        val_loader = create("EvalReader")(reader_cfg["EvalDataset"],
+                                          reader_cfg["worker_num"],
+                                          return_list=True)
+        clsid2catid = {v: k for k, v in dataset.catid2clsid.items()}
+        anno_file = dataset.get_anno()
+        metric = COCOMetric(
+            anno_file=anno_file, clsid2catid=clsid2catid, IouType="bbox")
+        eval(predictor, val_loader, metric, rerun_flag=rerun_flag)
+
+    if rerun_flag:
+        print(
+            "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
+        )
+
+
+if __name__ == "__main__":
+    paddle.enable_static()
+    parser = argsparser()
+    FLAGS = parser.parse_args()
+
+    # DataLoader need run on cpu
+    paddle.set_device("cpu")
+
+    main()

example/auto_compression/image_classification/README.md

@@ -113,48 +113,56 @@ python -m paddle.distributed.launch run.py --save_dir='./save_quant_mobilev1/' -
 
 注意 ```learning rate``` 与 ```batch size``` 呈线性关系，这里单卡 ```batch size``` 为32，对应的 ```learning rate``` 为0.015，那么如果 ```batch size``` 减小4倍改为8，```learning rate``` 也需除以4；多卡时 ```batch size``` 为32，```learning rate``` 需乘上卡数。所以改变 ```batch size``` 或改变训练卡数都需要对应修改 ```learning rate```。
 
-**验证精度**
-
-根据训练log可以看到模型验证的精度，若需再次验证精度，修改配置文件```./configs/MobileNetV1/qat_dis.yaml```中所需验证模型的文件夹路径及模型和参数名称```model_dir, model_filename, params_filename```，然后使用以下命令进行验证：
-
-```shell
-export CUDA_VISIBLE_DEVICES=0
-python eval.py --config_path='./configs/MobileNetV1/qat_dis.yaml'
-```
-
 
 ## 4.预测部署
-#### 4.1 Python预测推理
 
-环境配置：若使用 TesorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
+#### 4.1 Paddle Inference 验证性能
+
+量化模型在GPU上可以使用TensorRT进行加速，在CPU上可以使用MKLDNN进行加速。
 
 以下字段用于配置预测参数：
-- ```inference_model_dir```：inference 模型文件所在目录，该目录下需要有文件 .pdmodel 和 .pdiparams 两个文件
-- ```model_filename```：inference_model_dir文件夹下的模型文件名称
-- ```params_filename```：inference_model_dir文件夹下的参数文件名称
-
-- ```batch_size```：预测一个batch的大小
-- ```image_size```：输入图像的大小
-- ```use_tensorrt```：是否使用 TesorRT 预测引擎
-- ```use_gpu```：是否使用 GPU 预测
-- ```enable_mkldnn```：是否启用```MKL-DNN```加速库，注意```enable_mkldnn```与```use_gpu```同时为```True```时，将忽略```enable_mkldnn```，而使用```GPU```预测
-- ```use_fp16```：是否启用```FP16```
-- ```use_int8```：是否启用```INT8```
+
+| 参数名 | 含义 |
+|:------:|:------:|
+| model_path | inference 模型文件所在目录，该目录下需要有文件 .pdmodel 和 .pdiparams 两个文件 |
+| model_filename | inference_model_dir文件夹下的模型文件名称 |
+| params_filename | inference_model_dir文件夹下的参数文件名称 |
+| data_path | 数据集路径  |
+| batch_size | 预测一个batch的大小   |
+| image_size | 输入图像的大小   |
+| use_gpu | 是否使用 GPU 预测   |
+| use_trt | 是否使用 TesorRT 预测引擎   |
+| use_mkldnn | 是否启用```MKL-DNN```加速库，注意```use_mkldnn```与```use_gpu```同时为```True```时，将忽略```use_mkldnn```，而使用```GPU```预测  |
+| cpu_num_threads | CPU预测时，使用CPU线程数量，默认10  |
+| use_fp16 | 使用TensorRT时，是否启用```FP16```  |
+| use_int8 | 是否启用```INT8``` |
 
 注意：
 - 请注意模型的输入数据尺寸，如InceptionV3输入尺寸为299，部分模型需要修改参数：```image_size```
-- 如果希望提升评测模型速度，使用 ```GPU``` 评测时，建议开启 ```TensorRT``` 加速预测，使用 ```CPU``` 评测时，建议开启 ```MKL-DNN``` 加速预测
 
 
-准备好inference模型后，使用以下命令进行预测：
+- TensorRT预测：
+
+环境配置：如果使用 TesorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
+
+```shell
+python paddle_inference_eval.py \
+      --model_path=models/ResNet50_vd_QAT \
+      --use_trt=True \
+      --use_int8=True \
+      --use_gpu=True \
+      --data_path=./dataset/ILSVRC2012/
+```
+
+- MKLDNN预测：
+
 ```shell
-python infer.py --model_dir='MobileNetV1_infer' \
---model_filename='inference.pdmodel' \
---model_filename='inference.pdiparams' \
---eval=True \
---use_gpu=True \
---enable_mkldnn=True \
---use_int8=True
+python paddle_inference_eval.py \
+      --model_path=models/ResNet50_vd_QAT \
+      --data_path=./dataset/ILSVRC2012/ \
+      --cpu_num_threads=10 \
+      --use_mkldnn=True \
+      --use_int8=True
 ```
 
 #### 4.2 PaddleLite端侧部署

example/auto_compression/image_classification/infer.py → example/auto_compression/image_classification/paddle_inference_eval.py

@@ -13,76 +13,72 @@
 # limitations under the License.
 
 import os
-import numpy as np
-import cv2
 import time
 import sys
 import argparse
+import numpy as np
+import cv2
 import yaml
-from tqdm import tqdm
 
-from utils import preprocess, postprocess
 import paddle
 from paddle.inference import create_predictor
-from paddleslim.common import load_config
 from paddle.io import DataLoader
-from imagenet_reader import ImageNetDataset, process_image
+from imagenet_reader import ImageNetDataset
 
 
 def argsparser():
+    """
+    argsparser func
+    """
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument(
-        '--model_dir',
+        "--model_path",
         type=str,
-        default='./MobileNetV1_infer',
-        help='model directory')
+        default="./MobileNetV1_infer",
+        help="model directory")
     parser.add_argument(
-        '--model_filename',
+        "--model_filename",
         type=str,
-        default='inference.pdmodel',
-        help='model file name')
+        default="inference.pdmodel",
+        help="model file name")
     parser.add_argument(
-        '--params_filename',
+        "--params_filename",
         type=str,
-        default='inference.pdiparams',
-        help='params file name')
-    parser.add_argument('--batch_size', type=int, default=1)
-    parser.add_argument('--img_size', type=int, default=224)
-    parser.add_argument('--resize_size', type=int, default=256)
+        default="inference.pdiparams",
+        help="params file name")
+    parser.add_argument("--batch_size", type=int, default=1)
+    parser.add_argument("--img_size", type=int, default=224)
+    parser.add_argument("--resize_size", type=int, default=256)
     parser.add_argument(
-        '--eval', type=bool, default=False, help='Whether to evaluate')
-    parser.add_argument('--data_path', type=str, default='./ILSVRC2012/')
+        "--data_path", type=str, default="./dataset/ILSVRC2012/")
     parser.add_argument(
-        '--use_gpu', type=bool, default=False, help='Whether to use gpu')
+        "--use_gpu", type=bool, default=False, help="Whether to use gpu")
     parser.add_argument(
-        '--enable_mkldnn',
-        type=bool,
-        default=False,
-        help='Whether to use mkldnn')
+        "--use_trt", type=bool, default=False, help="Whether to use tensorrt")
     parser.add_argument(
-        '--cpu_num_threads', type=int, default=10, help='Number of cpu threads')
+        "--use_mkldnn", type=bool, default=False, help="Whether to use mkldnn")
     parser.add_argument(
-        '--use_fp16', type=bool, default=False, help='Whether to use fp16')
+        "--cpu_num_threads", type=int, default=10, help="Number of cpu threads")
     parser.add_argument(
-        '--use_int8', type=bool, default=False, help='Whether to use int8')
+        "--use_fp16", type=bool, default=False, help="Whether to use fp16")
     parser.add_argument(
-        '--use_tensorrt',
-        type=bool,
-        default=True,
-        help='Whether to use tensorrt')
+        "--use_int8", type=bool, default=False, help="Whether to use int8")
+    parser.add_argument("--gpu_mem", type=int, default=8000, help="GPU memory")
+    parser.add_argument("--ir_optim", type=bool, default=True)
     parser.add_argument(
-        '--enable_profile',
+        "--use_dynamic_shape",
         type=bool,
-        default=False,
-        help='Whether to enable profile')
-    parser.add_argument('--gpu_mem', type=int, default=8000, help='GPU memory')
-    parser.add_argument('--ir_optim', type=bool, default=True)
+        default=True,
+        help="Whether use dynamic shape or not.")
     return parser
 
 
 def eval_reader(data_dir, batch_size, crop_size, resize_size):
+    """
+    eval reader func
+    """
     val_reader = ImageNetDataset(
-        mode='val',
+        mode="val",
         data_dir=data_dir,
         crop_size=crop_size,
         resize_size=resize_size)
@@ -96,14 +92,17 @@ def eval_reader(data_dir, batch_size, crop_size, resize_size):
 
 
 class Predictor(object):
-    def __init__(self, args):
+    """
+    Paddle Inference Predictor class
+    """
 
+    def __init__(self):
         # HALF precission predict only work when using tensorrt
         if args.use_fp16 is True:
-            assert args.use_tensorrt is True
-        self.args = args
+            assert args.use_trt is True
 
-        self.paddle_predictor = self.create_paddle_predictor()
+        self.rerun_flag = False
+        self.paddle_predictor = self._create_paddle_predictor()
         input_names = self.paddle_predictor.get_input_names()
         self.input_tensor = self.paddle_predictor.get_input_handle(input_names[
             0])
@@ -112,96 +111,94 @@ class Predictor(object):
         self.output_tensor = self.paddle_predictor.get_output_handle(
             output_names[0])
 
-    def create_paddle_predictor(self):
-        inference_model_dir = self.args.model_dir
-        model_file = os.path.join(inference_model_dir, self.args.model_filename)
-        params_file = os.path.join(inference_model_dir,
-                                   self.args.params_filename)
+    def _create_paddle_predictor(self):
+        inference_model_dir = args.model_path
+        model_file = os.path.join(inference_model_dir, args.model_filename)
+        params_file = os.path.join(inference_model_dir, args.params_filename)
         config = paddle.inference.Config(model_file, params_file)
         precision = paddle.inference.Config.Precision.Float32
-        if self.args.use_int8:
+        if args.use_int8:
             precision = paddle.inference.Config.Precision.Int8
-        elif self.args.use_fp16:
+        elif args.use_fp16:
             precision = paddle.inference.Config.Precision.Half
 
-        if self.args.use_gpu:
-            config.enable_use_gpu(self.args.gpu_mem, 0)
+        if args.use_gpu:
+            config.enable_use_gpu(args.gpu_mem, 0)
         else:
             config.disable_gpu()
-            if self.args.enable_mkldnn:
-                # cache 10 different shapes for mkldnn to avoid memory leak
-                config.set_mkldnn_cache_capacity(10)
+            config.set_cpu_math_library_num_threads(args.cpu_num_threads)
+            config.switch_ir_optim()
+            if args.use_mkldnn:
                 config.enable_mkldnn()
-        config.set_cpu_math_library_num_threads(self.args.cpu_num_threads)
+                if args.use_int8:
+                    config.enable_mkldnn_int8(
+                        {"conv2d", "depthwise_conv2d", "transpose2", "pool2d"})
 
-        if self.args.enable_profile:
-            config.enable_profile()
-        config.switch_ir_optim(self.args.ir_optim)  # default true
-        if self.args.use_tensorrt:
+        config.switch_ir_optim(args.ir_optim)  # default true
+        if args.use_trt:
             config.enable_tensorrt_engine(
                 precision_mode=precision,
-                max_batch_size=self.args.batch_size,
+                max_batch_size=args.batch_size,
                 workspace_size=1 << 30,
                 min_subgraph_size=30,
-                use_calib_mode=False)
+                use_static=True,
+                use_calib_mode=False, )
+
+            if args.use_dynamic_shape:
+                dynamic_shape_file = os.path.join(inference_model_dir,
+                                                  "dynamic_shape.txt")
+                if os.path.exists(dynamic_shape_file):
+                    config.enable_tuned_tensorrt_dynamic_shape(
+                        dynamic_shape_file, True)
+                    print("trt set dynamic shape done!")
+                else:
+                    config.collect_shape_range_info(dynamic_shape_file)
+                    print("Start collect dynamic shape...")
+                    self.rerun_flag = True
 
         config.enable_memory_optim()
-        # use zero copy
-        config.switch_use_feed_fetch_ops(False)
         predictor = create_predictor(config)
 
         return predictor
 
-    def predict(self):
-        test_num = 1000
-        test_time = 0.0
-        for i in range(0, test_num + 10):
-            inputs = np.random.rand(self.args.batch_size, 3, self.args.img_size,
-                                    self.args.img_size).astype(np.float32)
-            start_time = time.time()
-            self.input_tensor.copy_from_cpu(inputs)
-            self.paddle_predictor.run()
-            batch_output = self.output_tensor.copy_to_cpu().flatten()
-            if i >= 10:
-                test_time += time.time() - start_time
-            time.sleep(0.01)  # sleep for T4 GPU
-
-        fp_message = "FP16" if self.args.use_fp16 else "FP32"
-        fp_message = "INT8" if self.args.use_int8 else fp_message
-        trt_msg = "using tensorrt" if self.args.use_tensorrt else "not using tensorrt"
-        print("{0}\t{1}\tbatch size: {2}\ttime(ms): {3}".format(
-            trt_msg, fp_message, args.batch_size, 1000 * test_time / test_num))
-
     def eval(self):
-        if os.path.exists(self.args.data_path):
+        """
+        eval func
+        """
+        if os.path.exists(args.data_path):
             val_loader = eval_reader(
-                self.args.data_path,
-                batch_size=self.args.batch_size,
-                crop_size=self.args.img_size,
-                resize_size=self.args.resize_size)
+                args.data_path,
+                batch_size=args.batch_size,
+                crop_size=args.img_size,
+                resize_size=args.resize_size)
         else:
-            image = np.ones((1, 3, self.args.img_size,
-                             self.args.img_size)).astype(np.float32)
+            image = np.ones(
+                (1, 3, args.img_size, args.img_size)).astype(np.float32)
             label = None
             val_loader = [[image, label]]
         results = []
-        with tqdm(
-                total=len(val_loader),
-                bar_format='Evaluation stage, Run batch:|{bar}| {n_fmt}/{total_fmt}',
-                ncols=80) as t:
-            for batch_id, (image, label) in enumerate(val_loader):
         input_names = self.paddle_predictor.get_input_names()
-                input_tensor = self.paddle_predictor.get_input_handle(
-                    input_names[0])
+        input_tensor = self.paddle_predictor.get_input_handle(input_names[0])
         output_names = self.paddle_predictor.get_output_names()
-                output_tensor = self.paddle_predictor.get_output_handle(
-                    output_names[0])
-
+        output_tensor = self.paddle_predictor.get_output_handle(output_names[0])
+        predict_time = 0.0
+        time_min = float("inf")
+        time_max = float("-inf")
+        sample_nums = len(val_loader)
+        for batch_id, (image, label) in enumerate(val_loader):
             image = np.array(image)
 
             input_tensor.copy_from_cpu(image)
+            start_time = time.time()
             self.paddle_predictor.run()
             batch_output = output_tensor.copy_to_cpu()
+            end_time = time.time()
+            timed = end_time - start_time
+            time_min = min(time_min, timed)
+            time_max = max(time_max, timed)
+            predict_time += timed
+            if self.rerun_flag:
+                return
             sort_array = batch_output.argsort(axis=1)
             top_1_pred = sort_array[:, -1:][:, ::-1]
             if label is None:
@@ -211,22 +208,43 @@ class Predictor(object):
             top_1 = np.mean(label == top_1_pred)
             top_5_pred = sort_array[:, -5:][:, ::-1]
             acc_num = 0
-                for i in range(len(label)):
+            for i, _ in enumerate(label):
                 if label[i][0] in top_5_pred[i]:
                     acc_num += 1
             top_5 = float(acc_num) / len(label)
             results.append([top_1, top_5])
+            if batch_id % 100 == 0:
+                print("Eval iter:", batch_id)
+                sys.stdout.flush()
 
         result = np.mean(np.array(results), axis=0)
-            t.update()
-        print('Evaluation result: {}'.format(result[0]))
+        fp_message = "FP16" if args.use_fp16 else "FP32"
+        fp_message = "INT8" if args.use_int8 else fp_message
+        print_msg = "Paddle"
+        if args.use_trt:
+            print_msg = "using TensorRT"
+        elif args.use_mkldnn:
+            print_msg = "using MKLDNN"
+        time_avg = predict_time / sample_nums
+        print(
+            "[Benchmark]{}\t{}\tbatch size: {}.Inference time(ms): min={}, max={}, avg={}".
+            format(
+                print_msg,
+                fp_message,
+                args.batch_size,
+                round(time_min * 1000, 2),
+                round(time_max * 1000, 1),
+                round(time_avg * 1000, 1), ))
+        print("[Benchmark] Evaluation acc result: {}".format(result[0]))
+        sys.stdout.flush()
 
 
 if __name__ == "__main__":
     parser = argsparser()
-    global args
     args = parser.parse_args()
-    predictor = Predictor(args)
-    predictor.predict()
-    if args.eval:
+    predictor = Predictor()
     predictor.eval()
+    if predictor.rerun_flag:
+        print(
+            "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
+        )

example/auto_compression/nlp/README.md

@@ -194,25 +194,41 @@ Quantization:
 
 ## 5. 预测部署
 
-- Python部署:
+量化模型在GPU上可以使用TensorRT进行加速，在CPU上可以使用MKLDNN进行加速。
 
-首先安装带有TensorRT的[Paddle安装包](https://www.paddlepaddle.org.cn/inference/v2.3/user_guides/download_lib.html#python)。
 
-然后使用[infer.py](./infer.py)进行部署：
+- TensorRT预测：
 
-本示例将以ERNIE 3.0-Medium模型、afqmc数据集的为例，介绍如何利用Paddle—TensorRT测试压缩后模型的精度和速度。
+环境配置：如果使用 TesorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
 
-精度测试方法：
+首先下载量化好的模型：
 ```shell
-python infer.py --task_name='afqmc' --model_path='./save_ernie3.0_afqmc/' --device='gpu' --use_trt --int8
+wget https://bj.bcebos.com/v1/paddle-slim-models/act/save_ppminilm_afqmc_new_calib.tar
+tar -xf save_ppminilm_afqmc_new_calib.tar
 ```
 
-速度测试方法
 ```shell
-python infer.py --task_name='afqmc' --model_path='./save_ernie3.0_afqmc/' --device='gpu' --use_trt --int8 --perf
+python paddle_inference_eval.py \
+      --model_path=save_ernie3_afqmc_new_cablib \
+      --model_filename=infer.pdmodel \
+      --params_filename=infer.pdiparams \
+      --task_name='afqmc' \
+      --use_trt \
+      --precision=int8
 ```
 
-- [PP-MiniLM Paddle Inference Python部署](https://github.com/PaddlePaddle/PaddleNLP/tree/develop/examples/model_compression/pp-minilm)
-- [ERNIE-3.0 Paddle Inference Python部署](https://github.com/PaddlePaddle/PaddleNLP/tree/develop/model_zoo/ernie-3.0)
+- MKLDNN预测：
+
+```shell
+python paddle_inference_eval.py \
+      --model_path=save_ernie3_afqmc_new_cablib \
+      --model_filename=infer.pdmodel \
+      --params_filename=infer.pdiparams \
+      --task_name='afqmc' \
+      --device=cpu \
+      --use_mkldnn=True \
+      --cpu_threads=10 \
+      --precision=int8
+```
 
 ## 6. FAQ

example/auto_compression/nlp/infer.py → example/auto_compression/nlp/paddle_inference_eval.py

@@ -45,96 +45,42 @@ METRIC_CLASSES = {
 }
 
 
-def convert_example(example, dataset, tokenizer, label_list,
-                    max_seq_length=512):
-    assert dataset in ['glue', 'clue'
-                       ], "This demo only supports for dataset glue or clue"
-    """Convert a glue example into necessary features."""
-    if dataset == 'glue':
-        # `label_list == None` is for regression task
-        label_dtype = "int64" if label_list else "float32"
-        # Get the label
-        label = example['labels']
-        label = np.array([label], dtype=label_dtype)
-        # Convert raw text to feature
-        example = tokenizer(example['sentence'], max_seq_len=max_seq_length)
-
-        return example['input_ids'], example['token_type_ids'], label
-
-    else:  #if dataset == 'clue':
-        # `label_list == None` is for regression task
-        label_dtype = "int64" if label_list else "float32"
-        # Get the label
-        example['label'] = np.array(
-            example["label"], dtype="int64").reshape((-1, 1))
-        label = example['label']
-        # Convert raw text to feature
-        if 'keyword' in example:  # CSL
-            sentence1 = " ".join(example['keyword'])
-            example = {
-                'sentence1': sentence1,
-                'sentence2': example['abst'],
-                'label': example['label']
-            }
-        elif 'target' in example:  # wsc
-            text, query, pronoun, query_idx, pronoun_idx = example[
-                'text'], example['target']['span1_text'], example['target'][
-                    'span2_text'], example['target']['span1_index'], example[
-                        'target']['span2_index']
-            text_list = list(text)
-            assert text[pronoun_idx:(pronoun_idx + len(
-                pronoun))] == pronoun, "pronoun: {}".format(pronoun)
-            assert text[query_idx:(query_idx + len(query)
-                                   )] == query, "query: {}".format(query)
-            if pronoun_idx > query_idx:
-                text_list.insert(query_idx, "_")
-                text_list.insert(query_idx + len(query) + 1, "_")
-                text_list.insert(pronoun_idx + 2, "[")
-                text_list.insert(pronoun_idx + len(pronoun) + 2 + 1, "]")
-            else:
-                text_list.insert(pronoun_idx, "[")
-                text_list.insert(pronoun_idx + len(pronoun) + 1, "]")
-                text_list.insert(query_idx + 2, "_")
-                text_list.insert(query_idx + len(query) + 2 + 1, "_")
-            text = "".join(text_list)
-            example['sentence'] = text
-        if tokenizer is None:
-            return example
-        if 'sentence' in example:
-            example = tokenizer(example['sentence'], max_seq_len=max_seq_length)
-        elif 'sentence1' in example:
-            example = tokenizer(
-                example['sentence1'],
-                text_pair=example['sentence2'],
-                max_seq_len=max_seq_length)
-        return example['input_ids'], example['token_type_ids'], label
-
-
 def parse_args():
+    """
+    parse_args func
+    """
     parser = argparse.ArgumentParser()
-
-    # Required parameters
+    parser.add_argument(
+        "--model_path",
+        default="./afqmc",
+        type=str,
+        required=True,
+        help="The path prefix of inference model to be used.", )
+    parser.add_argument(
+        "--model_filename",
+        type=str,
+        default="inference.pdmodel",
+        help="model file name")
+    parser.add_argument(
+        "--params_filename",
+        type=str,
+        default="inference.pdiparams",
+        help="params file name")
     parser.add_argument(
         "--task_name",
-        default='afqmc',
+        default="afqmc",
         type=str,
         help="The name of the task to perform predict, selected in the list: " +
         ", ".join(METRIC_CLASSES.keys()), )
     parser.add_argument(
         "--dataset",
-        default='clue',
+        default="clue",
         type=str,
         help="The dataset of model.", )
-    parser.add_argument(
-        "--model_path",
-        default='./quant_models/model',
-        type=str,
-        required=True,
-        help="The path prefix of inference model to be used.", )
     parser.add_argument(
         "--device",
         default="gpu",
-        choices=["gpu", "cpu", "xpu"],
+        choices=["gpu", "cpu"],
         help="Device selected for inference.", )
     parser.add_argument(
         "--batch_size",
@@ -154,25 +100,101 @@ def parse_args():
         help="Warmup steps for performance test.", )
     parser.add_argument(
         "--use_trt",
-        action='store_true',
+        action="store_true",
         help="Whether to use inference engin TensorRT.", )
     parser.add_argument(
-        "--perf",
-        action='store_true',
-        help="Whether to test performance.", )
+        "--precision",
+        type=str,
+        default="fp32",
+        choices=["fp32", "fp16", "int8"],
+        help="The precision of inference. It can be 'fp32', 'fp16' or 'int8'. Default is 'fp16'.",
+    )
     parser.add_argument(
-        "--int8",
-        action='store_true',
-        help="Whether to use int8 inference.", )
+        "--use_mkldnn",
+        type=bool,
+        default=False,
+        help="Whether use mkldnn or not.")
     parser.add_argument(
-        "--fp16",
-        action='store_true',
-        help="Whether to use float16 inference.", )
+        "--cpu_threads", type=int, default=1, help="Num of cpu threads.")
     args = parser.parse_args()
     return args
 
 
+def _convert_example(example,
+                     dataset,
+                     tokenizer,
+                     label_list,
+                     max_seq_length=512):
+    assert dataset in ["glue", "clue"
+                       ], "This demo only supports for dataset glue or clue"
+    """Convert a glue example into necessary features."""
+    if dataset == "glue":
+        # `label_list == None` is for regression task
+        label_dtype = "int64" if label_list else "float32"
+        # Get the label
+        label = example["labels"]
+        label = np.array([label], dtype=label_dtype)
+        # Convert raw text to feature
+        example = tokenizer(example["sentence"], max_seq_len=max_seq_length)
+
+        return example["input_ids"], example["token_type_ids"], label
+
+    else:  # if dataset == 'clue':
+        # `label_list == None` is for regression task
+        label_dtype = "int64" if label_list else "float32"
+        # Get the label
+        example["label"] = np.array(
+            example["label"], dtype="int64").reshape((-1, 1))
+        label = example["label"]
+        # Convert raw text to feature
+        if "keyword" in example:  # CSL
+            sentence1 = " ".join(example["keyword"])
+            example = {
+                "sentence1": sentence1,
+                "sentence2": example["abst"],
+                "label": example["label"]
+            }
+        elif "target" in example:  # wsc
+            text, query, pronoun, query_idx, pronoun_idx = (
+                example["text"],
+                example["target"]["span1_text"],
+                example["target"]["span2_text"],
+                example["target"]["span1_index"],
+                example["target"]["span2_index"], )
+            text_list = list(text)
+            assert text[pronoun_idx:(pronoun_idx + len(
+                pronoun))] == pronoun, "pronoun: {}".format(pronoun)
+            assert text[query_idx:(query_idx + len(query)
+                                   )] == query, "query: {}".format(query)
+            if pronoun_idx > query_idx:
+                text_list.insert(query_idx, "_")
+                text_list.insert(query_idx + len(query) + 1, "_")
+                text_list.insert(pronoun_idx + 2, "[")
+                text_list.insert(pronoun_idx + len(pronoun) + 2 + 1, "]")
+            else:
+                text_list.insert(pronoun_idx, "[")
+                text_list.insert(pronoun_idx + len(pronoun) + 1, "]")
+                text_list.insert(query_idx + 2, "_")
+                text_list.insert(query_idx + len(query) + 2 + 1, "_")
+            text = "".join(text_list)
+            example["sentence"] = text
+        if tokenizer is None:
+            return example
+        if "sentence" in example:
+            example = tokenizer(example["sentence"], max_seq_len=max_seq_length)
+        elif "sentence1" in example:
+            example = tokenizer(
+                example["sentence1"],
+                text_pair=example["sentence2"],
+                max_seq_len=max_seq_length)
+        return example["input_ids"], example["token_type_ids"], label
+
+
 class Predictor(object):
+    """
+    Inference Predictor class
+    """
+
     def __init__(self, predictor, input_handles, output_handles):
         self.predictor = predictor
         self.input_handles = input_handles
@@ -180,60 +202,50 @@ class Predictor(object):
 
     @classmethod
     def create_predictor(cls, args):
-        config = paddle.inference.Config(args.model_path + "infer.pdmodel",
-                                         args.model_path + "infer.pdiparams")
+        """
+        create_predictor func
+        """
+        cls.rerun_flag = False
+        config = paddle.inference.Config(
+            os.path.join(args.model_path, args.model_filename),
+            os.path.join(args.model_path, args.params_filename))
         if args.device == "gpu":
             # set GPU configs accordingly
             config.enable_use_gpu(100, 0)
             cls.device = paddle.set_device("gpu")
-        elif args.device == "cpu":
-            # set CPU configs accordingly,
-            # such as enable_mkldnn, set_cpu_math_library_num_threads
-            config.disable_gpu()
-            cls.device = paddle.set_device("cpu")
-        elif args.device == "xpu":
-            # set XPU configs accordingly
-            config.enable_xpu(100)
-        if args.use_trt:
-            if args.int8:
-                config.enable_tensorrt_engine(
-                    workspace_size=1 << 30,
-                    precision_mode=inference.PrecisionType.Int8,
-                    max_batch_size=args.batch_size,
-                    min_subgraph_size=5,
-                    use_static=False,
-                    use_calib_mode=False)
-            elif args.fp16:
-                config.enable_tensorrt_engine(
-                    workspace_size=1 << 30,
-                    precision_mode=inference.PrecisionType.Half,
-                    max_batch_size=args.batch_size,
-                    min_subgraph_size=5,
-                    use_static=False,
-                    use_calib_mode=False)
         else:
+            config.disable_gpu()
+            config.set_cpu_math_library_num_threads(args.cpu_threads)
+            config.switch_ir_optim()
+            if args.use_mkldnn:
+                config.enable_mkldnn()
+                if args.precision == "int8":
+                    config.enable_mkldnn_int8()
+
+        precision_map = {
+            "int8": inference.PrecisionType.Int8,
+            "fp32": inference.PrecisionType.Float32,
+            "fp16": inference.PrecisionType.Half,
+        }
+        if args.precision in precision_map.keys() and args.use_trt:
             config.enable_tensorrt_engine(
                 workspace_size=1 << 30,
-                    precision_mode=inference.PrecisionType.Float32,
                 max_batch_size=args.batch_size,
                 min_subgraph_size=5,
-                    use_static=False,
-                    use_calib_mode=False)
-            print("Enable TensorRT is: {}".format(
-                config.tensorrt_engine_enabled()))
+                precision_mode=precision_map[args.precision],
+                use_static=True,
+                use_calib_mode=False, )
 
             dynamic_shape_file = os.path.join(args.model_path,
-                                              'dynamic_shape.txt')
+                                              "dynamic_shape.txt")
             if os.path.exists(dynamic_shape_file):
                 config.enable_tuned_tensorrt_dynamic_shape(dynamic_shape_file,
                                                            True)
-                print('trt set dynamic shape done!')
+                print("trt set dynamic shape done!")
             else:
                 config.collect_shape_range_info(dynamic_shape_file)
-                print(
-                    'Start collect dynamic shape... Please eval again to get real result in TensorRT'
-                )
-                sys.exit()
+                print("Start collect dynamic shape...")
+                cls.rerun_flag = True
 
         predictor = paddle.inference.create_predictor(config)
 
@@ -249,6 +261,9 @@ class Predictor(object):
         return cls(predictor, input_handles, output_handles)
 
     def predict_batch(self, data):
+        """
+        predict from batch func
+        """
         for input_field, input_handle in zip(data, self.input_handles):
             input_handle.copy_from_cpu(input_field)
         self.predictor.run()
@@ -257,11 +272,11 @@ class Predictor(object):
         ]
         return output
 
-    def convert_predict_batch(self, args, data, tokenizer, batchify_fn,
+    def _convert_predict_batch(self, args, data, tokenizer, batchify_fn,
                                label_list):
         examples = []
         for example in data:
-            example = convert_example(
+            example = _convert_example(
                 example,
                 args.dataset,
                 tokenizer,
@@ -272,64 +287,82 @@ class Predictor(object):
         return examples
 
     def predict(self, dataset, tokenizer, batchify_fn, args):
+        """
+        predict func
+        """
         batches = [
             dataset[idx:idx + args.batch_size]
             for idx in range(0, len(dataset), args.batch_size)
         ]
-        if args.perf:
+
         for i, batch in enumerate(batches):
-                examples = self.convert_predict_batch(
+            examples = self._convert_predict_batch(
                 args, batch, tokenizer, batchify_fn, dataset.label_list)
             input_ids, segment_ids, label = batchify_fn(examples)
             output = self.predict_batch([input_ids, segment_ids])
             if i > args.perf_warmup_steps:
                 break
-            start_time = time.time()
-            for i, batch in enumerate(batches):
-                examples = self.convert_predict_batch(
-                    args, batch, tokenizer, batchify_fn, dataset.label_list)
-                input_ids, segment_ids, _ = batchify_fn(examples)
-                output = self.predict_batch([input_ids, segment_ids])
-
-            end_time = time.time()
-            sequences_num = i * args.batch_size
-            print("task name: %s, time: %s qps/s, " %
-                  (args.task_name, sequences_num / (end_time - start_time)))
+            if self.rerun_flag:
+                return
 
-        else:
         metric = METRIC_CLASSES[args.task_name]()
         metric.reset()
+        predict_time = 0.0
         for i, batch in enumerate(batches):
-                examples = self.convert_predict_batch(
+            examples = self._convert_predict_batch(
                 args, batch, tokenizer, batchify_fn, dataset.label_list)
             input_ids, segment_ids, label = batchify_fn(examples)
+            start_time = time.time()
             output = self.predict_batch([input_ids, segment_ids])
+            end_time = time.time()
+            predict_time += end_time - start_time
             correct = metric.compute(
                 paddle.to_tensor(output),
                 paddle.to_tensor(np.array(label).flatten()))
             metric.update(correct)
 
+        sequences_num = i * args.batch_size
+        print(
+            "[benchmark]task name: {}, batch size: {} Inference time per batch: {}ms, qps: {}.".
+            format(
+                args.task_name,
+                args.batch_size,
+                round(predict_time * 1000 / i, 2),
+                round(sequences_num / predict_time, 2), ))
         res = metric.accumulate()
-            print("task name: %s, acc: %s, \n" % (args.task_name, res), end='')
+        print(
+            "[benchmark]task name: %s, acc: %s. \n" % (args.task_name, res),
+            end="")
+        sys.stdout.flush()
 
 
 def main():
+    """
+    main func
+    """
     paddle.seed(42)
     args = parse_args()
     args.task_name = args.task_name.lower()
+    if args.use_mkldnn:
+        paddle.set_device("cpu")
 
     predictor = Predictor.create_predictor(args)
 
-    dev_ds = load_dataset('clue', args.task_name, splits='dev')
+    dev_ds = load_dataset("clue", args.task_name, splits="dev")
     tokenizer = AutoTokenizer.from_pretrained(args.model_path)
     batchify_fn = lambda samples, fn=Tuple(
         Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input
         Pad(axis=0, pad_val=tokenizer.pad_token_id),  # segment
-        Stack(dtype="int64" if dev_ds.label_list else "float32")  # label
+        Stack(dtype="int64" if dev_ds.label_list else "float32"),  # label
     ): fn(samples)
 
-    outputs = predictor.predict(dev_ds, tokenizer, batchify_fn, args)
+    predictor.predict(dev_ds, tokenizer, batchify_fn, args)
+    if predictor.rerun_flag:
+        print(
+            "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
+        )
 
 
 if __name__ == "__main__":
+    paddle.set_device("cpu")
     main()

example/auto_compression/pytorch_huggingface/README.md

@@ -14,12 +14,9 @@
 ## 1. 简介
 飞桨模型转换工具[X2Paddle](https://github.com/PaddlePaddle/X2Paddle)支持将```Caffe/TensorFlow/ONNX/PyTorch```的模型一键转为飞桨（PaddlePaddle）的预测模型。借助X2Paddle的能力，PaddleSlim的自动压缩功能可方便地用于各种框架的推理模型。
 
-
 本示例将以[Pytorch](https://github.com/pytorch/pytorch)框架的自然语言处理模型为例，介绍如何自动压缩其他框架中的自然语言处理模型。本示例会利用[huggingface](https://github.com/huggingface/transformers)开源transformers库，将Pytorch框架模型转换为Paddle框架模型，再使用ACT自动压缩功能进行自动压缩。本示例使用的自动压缩策略为剪枝蒸馏和量化训练。
 
 
-
-
 ## 2. Benchmark
 [BERT](https://arxiv.org/abs/1810.04805) （```Bidirectional Encoder Representations from Transformers```）以Transformer 编码器为网络基本组件，使用掩码语言模型（```Masked Language Model```）和邻接句子预测（```Next Sentence Prediction```）两个任务在大规模无标注文本语料上进行预训练（pre-train），得到融合了双向内容的通用语义表示模型。以预训练产生的通用语义表示模型为基础，结合任务适配的简单输出层，微调（fine-tune）后即可应用到下游的NLP任务，效果通常也较直接在下游的任务上训练的模型更优。此前BERT即在[GLUE](https://gluebenchmark.com/tasks)评测任务上取得了SOTA的结果。
 
@@ -192,41 +189,38 @@ python run.py --config_path=./configs/cola.yaml  --eval True
 
 ## 4. 预测部署
 
-环境配置：若使用 Paddle TensorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
+量化模型在GPU上可以使用TensorRT进行加速，在CPU上可以使用MKLDNN进行加速。
+
 
-启动配置：
+- TensorRT预测：
 
-除需传入```task_name```任务名称，```model_name_or_path```模型名称，```model_path```保存inference模型的路径等基本参数外，还需根据预测环境传入预测参数：
-- ```device```：默认为gpu，可选为gpu, cpu, xpu
-- ```use_trt```：是否使用 TesorRT 预测引擎
-- ```int8```：是否启用```INT8```
-- ```fp16```：是否启用```FP16```
+环境配置：如果使用 TesorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
 
-准备好inference模型后，可以使用```infer.py```进行预测，如使用 TesorRT 预测引擎测试 FP32 模型：
+首先下载量化好的模型：
 ```shell
-python -u ./infer.py \
-    --task_name cola \
-    --model_name_or_path bert-base-cased \
-    --model_path ./x2paddle_cola/model \
-    --batch_size 1 \
-    --max_seq_length 128 \
-    --device gpu \
-    --use_trt
+wget https://bj.bcebos.com/v1/paddle-slim-models/act/x2paddle_cola_new_calib.tar
+tar -xf x2paddle_cola_new_calib.tar
 ```
 
-如使用 TesorRT 预测引擎测试 INT8 模型：
 ```shell
-python -u ./infer.py \
-    --task_name cola \
-    --model_name_or_path bert-base-cased \
-    --model_path ./output/cola/model \
-    --batch_size 1 \
-    --max_seq_length 128 \
-    --device gpu \
+python paddle_inference_eval.py \
+      --model_path=x2paddle_cola_new_calib \
       --use_trt \
-    --int8
+      --precision=int8 \
+      --batch_size=1
 ```
 
+- MKLDNN预测：
+
+```shell
+python paddle_inference_eval.py \
+      --model_path=x2paddle_cola_new_calib \
+      --device=cpu \
+      --use_mkldnn=True \
+      --cpu_threads=10 \
+      --batch_size=1 \
+      --precision=int8
+```
 
 
 

example/auto_compression/pytorch_huggingface/infer.py → example/auto_compression/pytorch_huggingface/paddle_inference_eval.py

@@ -22,9 +22,9 @@ import numpy as np
 
 import paddle
 from paddle import inference
+from paddle.metric import Metric, Accuracy, Precision, Recall
 from paddlenlp.datasets import load_dataset
 from paddlenlp.data import Stack, Tuple, Pad
-from paddle.metric import Metric, Accuracy, Precision, Recall
 from paddlenlp.metrics import AccuracyAndF1, Mcc, PearsonAndSpearman
 from paddlenlp.transformers import BertForSequenceClassification, BertTokenizer
 
@@ -53,35 +53,46 @@ task_to_keys = {
 
 
 def parse_args():
+    """
+    parse_args func
+    """
     parser = argparse.ArgumentParser()
-
-    # Required parameters
+    parser.add_argument(
+        "--model_path",
+        default="./x2paddle_cola",
+        type=str,
+        required=True,
+        help="The path prefix of inference model to be used.", )
+    parser.add_argument(
+        "--model_filename",
+        type=str,
+        default="model.pdmodel",
+        help="model file name")
+    parser.add_argument(
+        "--params_filename",
+        type=str,
+        default="model.pdiparams",
+        help="params file name")
     parser.add_argument(
         "--task_name",
-        default='cola',
+        default="cola",
         type=str,
         help="The name of the task to perform predict, selected in the list: " +
         ", ".join(METRIC_CLASSES.keys()), )
     parser.add_argument(
         "--model_type",
-        default='bert-base-cased',
+        default="bert-base-cased",
         type=str,
         help="Model type selected in bert.")
     parser.add_argument(
         "--model_name_or_path",
-        default='bert-base-cased',
+        default="bert-base-cased",
         type=str,
         help="The directory or name of model.", )
-    parser.add_argument(
-        "--model_path",
-        default='./quant_models/model',
-        type=str,
-        required=True,
-        help="The path prefix of inference model to be used.", )
     parser.add_argument(
         "--device",
         default="gpu",
-        choices=["gpu", "cpu", "xpu"],
+        choices=["gpu", "cpu"],
         help="Device selected for inference.", )
     parser.add_argument(
         "--batch_size",
@@ -101,42 +112,45 @@ def parse_args():
         help="Warmup steps for performance test.", )
     parser.add_argument(
         "--use_trt",
-        action='store_true',
+        action="store_true",
         help="Whether to use inference engin TensorRT.", )
     parser.add_argument(
-        "--perf",
-        action='store_true',
-        help="Whether to test performance.", )
+        "--precision",
+        type=str,
+        default="fp32",
+        choices=["fp32", "fp16", "int8"],
+        help="The precision of inference. It can be 'fp32', 'fp16' or 'int8'. Default is 'fp16'.",
+    )
     parser.add_argument(
-        "--int8",
-        action='store_true',
-        help="Whether to use int8 inference.", )
+        "--use_mkldnn",
+        type=bool,
+        default=False,
+        help="Whether use mkldnn or not.")
     parser.add_argument(
-        "--fp16",
-        action='store_true',
-        help="Whether to use float16 inference.", )
+        "--cpu_threads", type=int, default=1, help="Num of cpu threads.")
     args = parser.parse_args()
     return args
 
 
-def convert_example(example,
+def _convert_example(
+        example,
         tokenizer,
         label_list,
         max_seq_length=512,
         task_name=None,
         is_test=False,
-                    padding='max_length',
-                    return_attention_mask=True):
+        padding="max_length",
+        return_attention_mask=True, ):
     if not is_test:
         # `label_list == None` is for regression task
         label_dtype = "int64" if label_list else "float32"
         # Get the label
-        label = example['labels']
+        label = example["labels"]
         label = np.array([label], dtype=label_dtype)
     # Convert raw text to feature
     sentence1_key, sentence2_key = task_to_keys[task_name]
-    texts = ((example[sentence1_key], ) if sentence2_key is None else
-             (example[sentence1_key], example[sentence2_key]))
+    texts = (example[sentence1_key], ) if sentence2_key is None else (
+        example[sentence1_key], example[sentence2_key])
     example = tokenizer(
         *texts,
         max_seq_len=max_seq_length,
@@ -144,19 +158,23 @@ def convert_example(example,
         return_attention_mask=return_attention_mask)
     if not is_test:
         if return_attention_mask:
-            return example['input_ids'], example['attention_mask'], example[
-                'token_type_ids'], label
+            return example["input_ids"], example["attention_mask"], example[
+                "token_type_ids"], label
         else:
-            return example['input_ids'], example['token_type_ids'], label
+            return example["input_ids"], example["token_type_ids"], label
     else:
         if return_attention_mask:
-            return example['input_ids'], example['attention_mask'], example[
-                'token_type_ids']
+            return example["input_ids"], example["attention_mask"], example[
+                "token_type_ids"]
         else:
-            return example['input_ids'], example['token_type_ids']
+            return example["input_ids"], example["token_type_ids"]
 
 
 class Predictor(object):
+    """
+    Inference Predictor class
+    """
+
     def __init__(self, predictor, input_handles, output_handles):
         self.predictor = predictor
         self.input_handles = input_handles
@@ -164,60 +182,51 @@ class Predictor(object):
 
     @classmethod
     def create_predictor(cls, args):
-        config = paddle.inference.Config(args.model_path + ".pdmodel",
-                                         args.model_path + ".pdiparams")
+        """
+        create_predictor func
+        """
+        cls.rerun_flag = False
+        config = paddle.inference.Config(
+            os.path.join(args.model_path, args.model_filename),
+            os.path.join(args.model_path, args.params_filename))
         if args.device == "gpu":
             # set GPU configs accordingly
             config.enable_use_gpu(100, 0)
             cls.device = paddle.set_device("gpu")
-        elif args.device == "cpu":
-            # set CPU configs accordingly,
-            # such as enable_mkldnn, set_cpu_math_library_num_threads
-            config.disable_gpu()
-            cls.device = paddle.set_device("cpu")
-        elif args.device == "xpu":
-            # set XPU configs accordingly
-            config.enable_xpu(100)
-        if args.use_trt:
-            if args.int8:
-                config.enable_tensorrt_engine(
-                    workspace_size=1 << 30,
-                    precision_mode=inference.PrecisionType.Int8,
-                    max_batch_size=args.batch_size,
-                    min_subgraph_size=5,
-                    use_static=False,
-                    use_calib_mode=False)
-            elif args.fp16:
-                config.enable_tensorrt_engine(
-                    workspace_size=1 << 30,
-                    precision_mode=inference.PrecisionType.Half,
-                    max_batch_size=args.batch_size,
-                    min_subgraph_size=5,
-                    use_static=False,
-                    use_calib_mode=False)
         else:
+            config.disable_gpu()
+            config.set_cpu_math_library_num_threads(args.cpu_threads)
+            config.switch_ir_optim()
+            if args.use_mkldnn:
+                config.enable_mkldnn()
+                if args.precision == "int8":
+                    config.enable_mkldnn_int8(
+                        {"fc", "reshape2", "transpose2", "slice"})
+
+        precision_map = {
+            "int8": inference.PrecisionType.Int8,
+            "fp32": inference.PrecisionType.Float32,
+            "fp16": inference.PrecisionType.Half,
+        }
+        if args.precision in precision_map.keys() and args.use_trt:
             config.enable_tensorrt_engine(
                 workspace_size=1 << 30,
-                    precision_mode=inference.PrecisionType.Float32,
                 max_batch_size=args.batch_size,
                 min_subgraph_size=5,
-                    use_static=False,
-                    use_calib_mode=False)
-            print("Enable TensorRT is: {}".format(
-                config.tensorrt_engine_enabled()))
+                precision_mode=precision_map[args.precision],
+                use_static=True,
+                use_calib_mode=False, )
 
-            model_dir = os.path.dirname(args.model_path)
-            dynamic_shape_file = os.path.join(model_dir, 'dynamic_shape.txt')
+            dynamic_shape_file = os.path.join(args.model_path,
+                                              "dynamic_shape.txt")
             if os.path.exists(dynamic_shape_file):
                 config.enable_tuned_tensorrt_dynamic_shape(dynamic_shape_file,
                                                            True)
-                print('trt set dynamic shape done!')
+                print("trt set dynamic shape done!")
             else:
                 config.collect_shape_range_info(dynamic_shape_file)
-                print(
-                    'Start collect dynamic shape... Please eval again to get real result in TensorRT'
-                )
-                sys.exit()
+                print("Start collect dynamic shape...")
+                cls.rerun_flag = True
 
         predictor = paddle.inference.create_predictor(config)
 
@@ -233,6 +242,9 @@ class Predictor(object):
         return cls(predictor, input_handles, output_handles)
 
     def predict(self, dataset, collate_fn, args):
+        """
+        predict func
+        """
         batch_sampler = paddle.io.BatchSampler(
             dataset, batch_size=args.batch_size, shuffle=False)
         data_loader = paddle.io.DataLoader(
@@ -241,94 +253,92 @@ class Predictor(object):
             collate_fn=collate_fn,
             num_workers=0,
             return_list=True)
-        end_time = 0
-        if args.perf:
+
         for i, data in enumerate(data_loader):
             for input_field, input_handle in zip(data, self.input_handles):
-                    input_handle.copy_from_cpu(input_field.numpy(
-                    ) if isinstance(input_field, paddle.Tensor) else
-                                               input_field)
-
+                input_handle.copy_from_cpu(input_field.numpy() if isinstance(
+                    input_field, paddle.Tensor) else input_field)
             self.predictor.run()
-
             output = [
                 output_handle.copy_to_cpu()
                 for output_handle in self.output_handles
             ]
-
             if i > args.perf_warmup_steps:
                 break
+            if self.rerun_flag:
+                return
 
-            time1 = time.time()
-            for i, data in enumerate(data_loader):
-                for input_field, input_handle in zip(data, self.input_handles):
-                    input_handle.copy_from_cpu(input_field.numpy(
-                    ) if isinstance(input_field, paddle.Tensor) else
-                                               input_field)
-                self.predictor.run()
-                output = [
-                    output_handle.copy_to_cpu()
-                    for output_handle in self.output_handles
-                ]
-
-            sequences_num = i * args.batch_size
-            print("task name: %s, time: %s qps/s, " %
-                  (args.task_name, sequences_num / (time.time() - time1)))
-
-        else:
         metric = METRIC_CLASSES[args.task_name]()
         metric.reset()
+        predict_time = 0.0
         for i, data in enumerate(data_loader):
             for input_field, input_handle in zip(data, self.input_handles):
-                    input_handle.copy_from_cpu(input_field.numpy(
-                    ) if isinstance(input_field, paddle.Tensor) else
-                                               input_field)
+                input_handle.copy_from_cpu(input_field.numpy() if isinstance(
+                    input_field, paddle.Tensor) else input_field)
+            start_time = time.time()
             self.predictor.run()
             output = [
                 output_handle.copy_to_cpu()
                 for output_handle in self.output_handles
             ]
-
+            end_time = time.time()
+            predict_time += end_time - start_time
             label = data[-1]
             correct = metric.compute(
                 paddle.to_tensor(output[0]),
                 paddle.to_tensor(np.array(label).flatten()))
-                print(correct)
             metric.update(correct)
 
+        sequences_num = i * args.batch_size
+        print(
+            "[benchmark]task name: {}, batch size: {} Inference time per batch: {}ms, qps: {}.".
+            format(
+                args.task_name,
+                args.batch_size,
+                round(predict_time * 1000 / i, 2),
+                round(sequences_num / predict_time, 2), ))
         res = metric.accumulate()
-            print("task name: %s, acc: %s, \n" % (args.task_name, res), end='')
+        print(
+            "[benchmark]task name: %s, acc: %s. \n" % (args.task_name, res),
+            end="")
+        sys.stdout.flush()
 
 
 def main():
+    """
+    main func
+    """
     paddle.seed(42)
     args = parse_args()
+    if args.use_mkldnn:
+        paddle.set_device("cpu")
 
     predictor = Predictor.create_predictor(args)
 
     args.task_name = args.task_name.lower()
     args.model_type = args.model_type.lower()
 
-    dev_ds = load_dataset('glue', args.task_name, splits='dev')
+    dev_ds = load_dataset("glue", args.task_name, splits="dev")
     tokenizer = BertTokenizer.from_pretrained(args.model_name_or_path)
 
     trans_func = partial(
-        convert_example,
+        _convert_example,
         tokenizer=tokenizer,
         label_list=dev_ds.label_list,
         max_seq_length=args.max_seq_length,
         task_name=args.task_name,
-        return_attention_mask=True)
+        return_attention_mask=True, )
 
     dev_ds = dev_ds.map(trans_func)
     batchify_fn = lambda samples, fn=Tuple(
         Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input
         Pad(axis=0, pad_val=0),
         Pad(axis=0, pad_val=tokenizer.pad_token_id),  # segment
-        Stack(dtype="int64" if dev_ds.label_list else "float32")  # label
+        Stack(dtype="int64" if dev_ds.label_list else "float32"),  # label
     ): fn(samples)
     predictor.predict(dev_ds, batchify_fn, args)
 
 
 if __name__ == "__main__":
+    paddle.set_device("cpu")
     main()

example/auto_compression/pytorch_yolo_series/README.md

@@ -8,7 +8,6 @@
   - [3.2 准备数据集](#32-准备数据集)
   - [3.3 准备预测模型](#33-准备预测模型)
   - [3.4 自动压缩并产出模型](#34-自动压缩并产出模型)
-  - [3.5 测试模型精度](#35-测试模型精度)
 - [4.预测部署](#4预测部署)
 - [5.FAQ](5FAQ)
 
@@ -149,14 +148,6 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch --log_dir=log -
           --config_path=./configs/yolov7_tiny_qat_dis.yaml --save_dir='./output/'
 ```
 
-#### 3.5 测试模型精度
-
-修改[yolov7_qat_dis.yaml](./configs/yolov7_qat_dis.yaml)中`model_dir`字段为模型存储路径，然后使用eval.py脚本得到模型的mAP：
-```
-export CUDA_VISIBLE_DEVICES=0
-python eval.py --config_path=./configs/yolov7_tiny_qat_dis.yaml
-```
-
 
 ## 4.预测部署
 
@@ -164,31 +155,60 @@ python eval.py --config_path=./configs/yolov7_tiny_qat_dis.yaml
 ```shell
 ├── model.pdiparams         # Paddle预测模型权重
 ├── model.pdmodel           # Paddle预测模型文件
-├── calibration_table.txt   # Paddle量化后校准表
 ├── ONNX
 │   ├── quant_model.onnx      # 量化后转出的ONNX模型
 │   ├── calibration.cache     # TensorRT可以直接加载的校准表
 ```
 
-#### 导出至ONNX使用TensorRT部署
+#### Paddle Inference部署测试
 
-加载`quant_model.onnx`和`calibration.cache`，可以直接使用TensorRT测试脚本进行验证，详细代码可参考[TensorRT部署](./TensorRT)
+量化模型在GPU上可以使用TensorRT进行加速，在CPU上可以使用MKLDNN进行加速。
+
+以下字段用于配置预测参数：
+
+| 参数名 | 含义 |
+|:------:|:------:|
+| model_path | inference 模型文件所在目录，该目录下需要有文件 model.pdmodel 和 model.pdiparams 两个文件 |
+| dataset_dir | eval时数据验证集路径， 默认`dataset/coco` |
+| image_file | 如果只测试单张图片效果，直接根据image_file指定图片路径 |
+| device | 使用GPU或者CPU预测，可选CPU/GPU   |
+| use_trt | 是否使用 TesorRT 预测引擎   |
+| use_mkldnn | 是否启用```MKL-DNN```加速库，注意```use_mkldnn```与```use_gpu```同时为```True```时，将忽略```enable_mkldnn```，而使用```GPU```预测  |
+| cpu_threads | CPU预测时，使用CPU线程数量，默认10  |
+| precision | 预测精度，包括`fp32/fp16/int8`  |
+
+ TensorRT Python部署:
+
+首先安装带有TensorRT的[Paddle安装包](https://www.paddlepaddle.org.cn/inference/v2.3/user_guides/download_lib.html#python)。
+
+然后使用[paddle_inference_eval.py](./paddle_inference_eval.py)进行部署：
 
-- python测试：
 ```shell
-cd TensorRT
-python trt_eval.py --onnx_model_file=output/ONNX/quant_model.onnx \
-                   --calibration_file=output/ONNX/calibration.cache \
-                   --image_file=../images/000000570688.jpg \
-                   --precision_mode=int8
+python paddle_inference_eval.py \
+      --model_path=output \
+      --reader_config=configs/yoloe_reader.yml \
+      --use_trt=True \
+      --precision=int8
 ```
 
-- 速度测试
+- MKLDNN预测：
+
 ```shell
-trtexec --onnx=output/ONNX/quant_model.onnx --avgRuns=1000 --workspace=1024 --calib=output/ONNX/calibration.cache --int8
+python paddle_inference_eval.py \
+      --model_path=output \
+      --reader_config=configs/yoloe_reader.yml \
+      --device=CPU \
+      --use_mkldnn=True \
+      --cpu_threads=10 \
+      --precision=int8
+```
+
+- 测试单张图片
+
+```shell
+python paddle_inference_eval.py --model_path=output --image_file=images/000000570688.jpg --use_trt=True --precision=int8
 ```
 
-#### Paddle-TensorRT部署
 - C++部署
 
 进入[cpp_infer](./cpp_infer)文件夹内，请按照[C++ TensorRT Benchmark测试教程](./cpp_infer/README.md)进行准备环境及编译，然后开始测试：
@@ -199,13 +219,22 @@ bash compile.sh
 ./build/trt_run --model_file yolov7_quant/model.pdmodel --params_file yolov7_quant/model.pdiparams --run_mode=trt_int8
 ```
 
-- Python部署:
+#### 导出至ONNX使用TensorRT部署
 
-首先安装带有TensorRT的[Paddle安装包](https://www.paddlepaddle.org.cn/inference/v2.3/user_guides/download_lib.html#python)。
+加载`quant_model.onnx`和`calibration.cache`，可以直接使用TensorRT测试脚本进行验证，详细代码可参考[TensorRT部署](./TensorRT)
+
+- python测试：
+```shell
+cd TensorRT
+python trt_eval.py --onnx_model_file=output/ONNX/quant_model.onnx \
+                   --calibration_file=output/ONNX/calibration.cache \
+                   --image_file=../images/000000570688.jpg \
+                   --precision_mode=int8
+```
 
-然后使用[paddle_trt_infer.py](./paddle_trt_infer.py)进行部署：
+- 速度测试
 ```shell
-python paddle_trt_infer.py --model_path=output --image_file=images/000000570688.jpg --benchmark=True --run_mode=trt_int8
+trtexec --onnx=output/ONNX/quant_model.onnx --avgRuns=1000 --workspace=1024 --calib=output/ONNX/calibration.cache --int8
 ```
 
 ## 5.FAQ

example/auto_compression/pytorch_yolo_series/paddle_inference_eval.py

+# Copyright (c) 2022 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 time
+import os
+import sys
+import argparse
+import cv2
+import numpy as np
+from tqdm import tqdm
+import pkg_resources as pkg
+
+import paddle
+from paddle.inference import Config
+from paddle.inference import create_predictor
+from dataset import COCOValDataset
+from post_process import YOLOPostProcess, coco_metric
+
+
+def argsparser():
+    """
+    argsparser func
+    """
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--model_path", type=str, help="inference model filepath")
+    parser.add_argument(
+        "--image_file",
+        type=str,
+        default=None,
+        help="image path, if set image_file, it will not eval coco.")
+    parser.add_argument(
+        "--dataset_dir",
+        type=str,
+        default="dataset/coco",
+        help="COCO dataset dir.")
+    parser.add_argument(
+        "--val_image_dir",
+        type=str,
+        default="val2017",
+        help="COCO dataset val image dir.")
+    parser.add_argument(
+        "--val_anno_path",
+        type=str,
+        default="annotations/instances_val2017.json",
+        help="COCO dataset anno path.")
+    parser.add_argument(
+        "--benchmark",
+        type=bool,
+        default=False,
+        help="Whether run benchmark or not.")
+    parser.add_argument(
+        "--use_dynamic_shape",
+        type=bool,
+        default=True,
+        help="Whether use dynamic shape or not.")
+    parser.add_argument(
+        "--use_trt",
+        type=bool,
+        default=False,
+        help="Whether use TensorRT or not.")
+    parser.add_argument(
+        "--precision",
+        type=str,
+        default="paddle",
+        help="mode of running(fp32/fp16/int8)")
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="GPU",
+        help="Choose the device you want to run, it can be: CPU/GPU/XPU, default is GPU",
+    )
+    parser.add_argument(
+        "--arch", type=str, default="YOLOv5", help="architectures name.")
+    parser.add_argument("--img_shape", type=int, default=640, help="input_size")
+    parser.add_argument(
+        "--batch_size", type=int, default=1, help="Batch size of model input.")
+    parser.add_argument(
+        "--use_mkldnn",
+        type=bool,
+        default=False,
+        help="Whether use mkldnn or not.")
+    parser.add_argument(
+        "--cpu_threads", type=int, default=1, help="Num of cpu threads.")
+    return parser
+
+
+CLASS_LABEL = [
+    'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train',
+    'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign',
+    'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
+    'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag',
+    'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite',
+    'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
+    'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon',
+    'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot',
+    'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant',
+    'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote',
+    'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',
+    'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
+    'hair drier', 'toothbrush'
+]
+
+
+def preprocess(image, input_size, mean=None, std=None, swap=(2, 0, 1)):
+    """
+    image preprocess func
+    """
+    if len(image.shape) == 3:
+        padded_img = np.ones((input_size[0], input_size[1], 3)) * 114.0
+    else:
+        padded_img = np.ones(input_size) * 114.0
+    img = np.array(image)
+    r = min(input_size[0] / img.shape[0], input_size[1] / img.shape[1])
+    resized_img = cv2.resize(
+        img,
+        (int(img.shape[1] * r), int(img.shape[0] * r)),
+        interpolation=cv2.INTER_LINEAR, ).astype(np.float32)
+    padded_img[:int(img.shape[0] * r), :int(img.shape[1] * r)] = resized_img
+
+    padded_img = padded_img[:, :, ::-1]
+    padded_img /= 255.0
+    if mean is not None:
+        padded_img -= mean
+    if std is not None:
+        padded_img /= std
+    padded_img = padded_img.transpose(swap)
+    padded_img = np.ascontiguousarray(padded_img, dtype=np.float32)
+    return padded_img, r
+
+
+def get_color_map_list(num_classes):
+    """
+    get_color_map_list func
+    """
+    color_map = num_classes * [0, 0, 0]
+    for i in range(0, num_classes):
+        j = 0
+        lab = i
+        while lab:
+            color_map[i * 3] |= ((lab >> 0) & 1) << (7 - j)
+            color_map[i * 3 + 1] |= ((lab >> 1) & 1) << (7 - j)
+            color_map[i * 3 + 2] |= ((lab >> 2) & 1) << (7 - j)
+            j += 1
+            lab >>= 3
+    color_map = [color_map[i:i + 3] for i in range(0, len(color_map), 3)]
+    return color_map
+
+
+def draw_box(img, boxes, scores, cls_ids, conf=0.5, class_names=None):
+    """
+    draw_box func
+    """
+    color_list = get_color_map_list(len(class_names))
+    for i, _ in enumerate(boxes):
+        box = boxes[i]
+        cls_id = int(cls_ids[i])
+        color = tuple(color_list[cls_id])
+        score = scores[i]
+        if score < conf:
+            continue
+        x0 = int(box[0])
+        y0 = int(box[1])
+        x1 = int(box[2])
+        y1 = int(box[3])
+
+        text = "{}:{:.1f}%".format(class_names[cls_id], score * 100)
+        font = cv2.FONT_HERSHEY_SIMPLEX
+
+        txt_size = cv2.getTextSize(text, font, 0.4, 1)[0]
+        cv2.rectangle(img, (x0, y0), (x1, y1), color, 2)
+        cv2.rectangle(img, (x0, y0 + 1), (
+            x0 + txt_size[0] + 1, y0 + int(1.5 * txt_size[1])), color, -1)
+        cv2.putText(
+            img,
+            text, (x0, y0 + txt_size[1]),
+            font,
+            0.8, (0, 255, 0),
+            thickness=2)
+
+    return img
+
+
+def get_current_memory_mb():
+    """
+    It is used to Obtain the memory usage of the CPU and GPU during the running of the program.
+    And this function Current program is time-consuming.
+    """
+    import pynvml
+    import psutil
+    import GPUtil
+
+    gpu_id = int(os.environ.get("CUDA_VISIBLE_DEVICES", 0))
+
+    pid = os.getpid()
+    p = psutil.Process(pid)
+    info = p.memory_full_info()
+    cpu_mem = info.uss / 1024.0 / 1024.0
+    gpu_mem = 0
+    gpu_percent = 0
+    gpus = GPUtil.getGPUs()
+    if gpu_id is not None and len(gpus) > 0:
+        gpu_percent = gpus[gpu_id].load
+        pynvml.nvmlInit()
+        handle = pynvml.nvmlDeviceGetHandleByIndex(0)
+        meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
+        gpu_mem = meminfo.used / 1024.0 / 1024.0
+    return round(cpu_mem, 4), round(gpu_mem, 4)
+
+
+def load_predictor(
+        model_dir,
+        precision="fp32",
+        use_trt=False,
+        use_mkldnn=False,
+        batch_size=1,
+        device="CPU",
+        min_subgraph_size=3,
+        use_dynamic_shape=False,
+        trt_min_shape=1,
+        trt_max_shape=1280,
+        trt_opt_shape=640,
+        cpu_threads=1, ):
+    """set AnalysisConfig, generate AnalysisPredictor
+    Args:
+        model_dir (str): root path of __model__ and __params__
+        precision (str): mode of running(fp32/fp16/int8)
+        use_trt (bool): whether use TensorRT or not.
+        use_mkldnn (bool): whether use MKLDNN or not in CPU.
+        device (str): Choose the device you want to run, it can be: CPU/GPU, default is CPU
+        use_dynamic_shape (bool): use dynamic shape or not
+        trt_min_shape (int): min shape for dynamic shape in trt
+        trt_max_shape (int): max shape for dynamic shape in trt
+        trt_opt_shape (int): opt shape for dynamic shape in trt
+    Returns:
+        predictor (PaddlePredictor): AnalysisPredictor
+    Raises:
+        ValueError: predict by TensorRT need device == 'GPU'.
+    """
+    rerun_flag = False
+    if device != "GPU" and use_trt:
+        raise ValueError(
+            "Predict by TensorRT mode: {}, expect device=='GPU', but device == {}".
+            format(precision, device))
+    config = Config(
+        os.path.join(model_dir, "model.pdmodel"),
+        os.path.join(model_dir, "model.pdiparams"))
+    if device == "GPU":
+        # initial GPU memory(M), device ID
+        config.enable_use_gpu(200, 0)
+        # optimize graph and fuse op
+        config.switch_ir_optim(True)
+    else:
+        config.disable_gpu()
+        config.set_cpu_math_library_num_threads(cpu_threads)
+        config.switch_ir_optim()
+        if use_mkldnn:
+            config.enable_mkldnn()
+            if precision == "int8":
+                config.enable_mkldnn_int8({"conv2d", "transpose2", "pool2d"})
+
+    precision_map = {
+        "int8": Config.Precision.Int8,
+        "fp32": Config.Precision.Float32,
+        "fp16": Config.Precision.Half,
+    }
+    if precision in precision_map.keys() and use_trt:
+        config.enable_tensorrt_engine(
+            workspace_size=(1 << 25) * batch_size,
+            max_batch_size=batch_size,
+            min_subgraph_size=min_subgraph_size,
+            precision_mode=precision_map[precision],
+            use_static=True,
+            use_calib_mode=False, )
+
+        if use_dynamic_shape:
+            dynamic_shape_file = os.path.join(FLAGS.model_path,
+                                              "dynamic_shape.txt")
+            if os.path.exists(dynamic_shape_file):
+                config.enable_tuned_tensorrt_dynamic_shape(dynamic_shape_file,
+                                                           True)
+                print("trt set dynamic shape done!")
+            else:
+                config.collect_shape_range_info(dynamic_shape_file)
+                print("Start collect dynamic shape...")
+                rerun_flag = True
+
+    # enable shared memory
+    config.enable_memory_optim()
+    predictor = create_predictor(config)
+    return predictor, rerun_flag
+
+
+def eval(predictor, val_loader, anno_file, rerun_flag=False):
+    """
+    eval main func
+    """
+    bboxes_list, bbox_nums_list, image_id_list = [], [], []
+    cpu_mems, gpu_mems = 0, 0
+    sample_nums = len(val_loader)
+    predict_time = 0.0
+    time_min = float("inf")
+    time_max = float("-inf")
+    input_names = predictor.get_input_names()
+    output_names = predictor.get_output_names()
+    boxes_tensor = predictor.get_output_handle(output_names[0])
+    for batch_id, data in enumerate(val_loader):
+        data_all = {k: np.array(v) for k, v in data.items()}
+        inputs = {}
+        if FLAGS.arch == "YOLOv6":
+            inputs["x2paddle_image_arrays"] = data_all["image"]
+        else:
+            inputs["x2paddle_images"] = data_all["image"]
+        for i, _ in enumerate(input_names):
+            input_tensor = predictor.get_input_handle(input_names[i])
+            input_tensor.copy_from_cpu(inputs[input_names[i]])
+        start_time = time.time()
+        predictor.run()
+        outs = boxes_tensor.copy_to_cpu()
+        end_time = time.time()
+        timed = end_time - start_time
+        time_min = min(time_min, timed)
+        time_max = max(time_max, timed)
+        predict_time += timed
+        if rerun_flag:
+            return
+        postprocess = YOLOPostProcess(
+            score_threshold=0.001, nms_threshold=0.65, multi_label=True)
+        res = postprocess(np.array(outs), data_all["scale_factor"])
+        bboxes_list.append(res["bbox"])
+        bbox_nums_list.append(res["bbox_num"])
+        image_id_list.append(np.array(data_all["im_id"]))
+        cpu_mem, gpu_mem = get_current_memory_mb()
+        cpu_mems += cpu_mem
+        gpu_mems += gpu_mem
+        if batch_id % 100 == 0:
+            print("Eval iter:", batch_id)
+            sys.stdout.flush()
+    print("[Benchmark]Avg cpu_mem:{} MB, avg gpu_mem: {} MB".format(
+        cpu_mems / sample_nums, gpu_mems / sample_nums))
+    time_avg = predict_time / sample_nums
+    print("[Benchmark]Inference time(ms): min={}, max={}, avg={}".format(
+        round(time_min * 1000, 2),
+        round(time_max * 1000, 1), round(time_avg * 1000, 1)))
+
+    map_res = coco_metric(anno_file, bboxes_list, bbox_nums_list, image_id_list)
+    print("[Benchmark] COCO mAP: {}".format(map_res[0]))
+    sys.stdout.flush()
+
+
+def infer(predictor):
+    """
+    infer image main func
+    """
+    warmup, repeats = 1, 1
+    if FLAGS.benchmark:
+        warmup, repeats = 50, 100
+    origin_img = cv2.imread(FLAGS.image_file)
+    input_image, scale_factor = preprocess(origin_img,
+                                           [FLAGS.img_shape, FLAGS.img_shape])
+    input_image = np.expand_dims(input_image, axis=0)
+    scale_factor = np.array([[scale_factor, scale_factor]])
+    inputs = {}
+    if FLAGS.arch == "YOLOv6":
+        inputs["x2paddle_image_arrays"] = input_image
+    else:
+        inputs["x2paddle_images"] = input_image
+    input_names = predictor.get_input_names()
+    for i, _ in enumerate(input_names):
+        input_tensor = predictor.get_input_handle(input_names[i])
+        input_tensor.copy_from_cpu(inputs[input_names[i]])
+
+    for i in range(warmup):
+        predictor.run()
+
+    np_boxes = None
+    predict_time = 0.0
+    time_min = float("inf")
+    time_max = float("-inf")
+    cpu_mems, gpu_mems = 0, 0
+    for i in range(repeats):
+        start_time = time.time()
+        predictor.run()
+        output_names = predictor.get_output_names()
+        boxes_tensor = predictor.get_output_handle(output_names[0])
+        np_boxes = boxes_tensor.copy_to_cpu()
+        end_time = time.time()
+        timed = end_time - start_time
+        time_min = min(time_min, timed)
+        time_max = max(time_max, timed)
+        predict_time += timed
+        cpu_mem, gpu_mem = get_current_memory_mb()
+        cpu_mems += cpu_mem
+        gpu_mems += gpu_mem
+    print("[Benchmark]Avg cpu_mem:{} MB, avg gpu_mem: {} MB".format(
+        cpu_mems / repeats, gpu_mems / repeats))
+
+    time_avg = predict_time / repeats
+    print("[Benchmark]Inference time(ms): min={}, max={}, avg={}".format(
+        round(time_min * 1000, 2),
+        round(time_max * 1000, 1), round(time_avg * 1000, 1)))
+    postprocess = YOLOPostProcess(
+        score_threshold=0.001, nms_threshold=0.65, multi_label=True)
+    res = postprocess(np_boxes, scale_factor)
+    # Draw rectangles and labels on the original image
+    dets = res["bbox"]
+    if dets is not None:
+        final_boxes, final_scores, final_class = dets[:, 2:], dets[:,
+                                                                   1], dets[:,
+                                                                            0]
+        res_img = draw_box(
+            origin_img,
+            final_boxes,
+            final_scores,
+            final_class,
+            conf=0.5,
+            class_names=CLASS_LABEL)
+        cv2.imwrite("output.jpg", res_img)
+        print("The prediction results are saved in output.jpg.")
+
+
+def main():
+    """
+    main func
+    """
+    predictor, rerun_flag = load_predictor(
+        FLAGS.model_path,
+        device=FLAGS.device,
+        use_trt=FLAGS.use_trt,
+        use_mkldnn=FLAGS.use_mkldnn,
+        precision=FLAGS.precision,
+        use_dynamic_shape=FLAGS.use_dynamic_shape,
+        cpu_threads=FLAGS.cpu_threads, )
+
+    if FLAGS.image_file:
+        infer(predictor)
+    else:
+        dataset = COCOValDataset(
+            dataset_dir=FLAGS.dataset_dir,
+            image_dir=FLAGS.val_image_dir,
+            anno_path=FLAGS.val_anno_path)
+        anno_file = dataset.ann_file
+        val_loader = paddle.io.DataLoader(
+            dataset, batch_size=FLAGS.batch_size, drop_last=True)
+        eval(predictor, val_loader, anno_file, rerun_flag=rerun_flag)
+
+    if rerun_flag:
+        print(
+            "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
+        )
+
+
+if __name__ == "__main__":
+    paddle.enable_static()
+    parser = argsparser()
+    FLAGS = parser.parse_args()
+
+    # DataLoader need run on cpu
+    paddle.set_device("cpu")
+
+    main()

example/auto_compression/semantic_segmentation/README.md

@@ -8,8 +8,7 @@
   - [3.2 准备数据集](#32-准备数据集)
   - [3.3 准备预测模型](#33-准备预测模型)
   - [3.4 自动压缩并产出模型](#34-自动压缩并产出模型)
-- [4.评估精度](#4评估精度)
-- [5.预测部署](#5预测部署)
+- [4.预测部署](#4预测部署)
 - [5.FAQ](5FAQ)
 
 ## 1.简介
@@ -156,104 +155,68 @@ python -m paddle.distributed.launch run.py --config_path='./configs/pp_humanseg/
 压缩完成后会在`save_dir`中产出压缩好的预测模型，可直接预测部署。
 
 
-## 4.评估精度
+## 4.预测部署
 
-本小节以人像分割模型和小数据集为例, 介绍如何在测试集上评估压缩后的模型.
+#### 4.1 Paddle Inference 验证性能
 
-下载经过量化训练压缩后的推理模型：
-```
-wget https://bj.bcebos.com/v1/paddle-slim-models/act/PaddleSeg/qat/pp_humanseg_qat.zip
-unzip pp_humanseg_qat.zip
-```
+量化模型在GPU上可以使用TensorRT进行加速，在CPU上可以使用MKLDNN进行加速。
 
-通过以下命令下载人像分割示例数据:
+以下字段用于配置预测参数：
 
-```shell
-cd ./data
-python download_data.py mini_humanseg
-cd -
-
-```
+| 参数名 | 含义 |
+|:------:|:------:|
+| model_path | inference 模型文件所在目录，该目录下需要有文件 .pdmodel 和 .pdiparams 两个文件 |
+| model_filename | inference_model_dir文件夹下的模型文件名称 |
+| params_filename | inference_model_dir文件夹下的参数文件名称 |
+| dataset | 选择数据集的类型，可选：`human`, `cityscape`。  |
+| dataset_config | 数据集配置的config  |
+| image_file | 待测试单张图片的路径，如果设置image_file，则dataset_config将无效。   |
+| device | 预测时的设备，可选：`CPU`, `GPU`。  |
+| use_trt | 是否使用 TesorRT 预测引擎，在device为```GPU```时生效。   |
+| use_mkldnn | 是否启用```MKL-DNN```加速库，注意```use_mkldnn```，在device为```CPU```时生效。  |
+| cpu_threads | CPU预测时，使用CPU线程数量，默认10  |
+| precision | 预测时精度，可选：`fp32`, `fp16`, `int8`。 |
 
-执行以下命令评估模型在测试集上的精度：
-
-```
-python eval.py \
---model_dir ./pp_humanseg_qat \
---model_filename model.pdmodel \
---params_filename model.pdiparams \
---dataset_config configs/dataset/humanseg_dataset.yaml
-
-```
 
-## 5.预测部署
+- TensorRT预测：
 
-本小节以人像分割为例, 介绍如何使用Paddle Inference推理库执行压缩后的模型.
+环境配置：如果使用 TesorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
 
-### 5.1 安装推理库
+准备好预测模型，并且修改dataset_config中数据集路径为正确的路径后，启动测试：
 
-请参考该链接安装Python版本的PaddleInference推理库: [推理库安装教程](https://www.paddlepaddle.org.cn/inference/user_guides/download_lib.html#python)
-
-### 5.2 准备模型和数据
-
-从 [2.Benchmark](#2Benchmark) 的表格中获得压缩前后的推理模型的下载链接，执行以下命令下载并解压推理模型：
-
-下载Float32数值类型的模型：
-```
-wget https://paddleseg.bj.bcebos.com/dygraph/ppseg/ppseg_lite_portrait_398x224_with_softmax.tar.gz
-tar -xzf ppseg_lite_portrait_398x224_with_softmax.tar.gz
-mv ppseg_lite_portrait_398x224_with_softmax pp_humanseg_fp32
-```
-
-下载经过量化训练压缩后的推理模型：
-```
-wget https://bj.bcebos.com/v1/paddle-slim-models/act/PaddleSeg/qat/pp_humanseg_qat.zip
-unzip pp_humanseg_qat.zip
+```shell
+python paddle_inference_eval.py \
+      --model_path=pp_liteseg_qat \
+      --dataset='cityscape' \
+      --dataset_config=configs/dataset/cityscapes_1024x512_scale1.0.yml \
+      --use_trt=True \
+      --precision=int8
 ```
 
-准备好需要处理的图片，这里直接使用人像示例图片 `./data/human_demo.jpg`。
-
-### 5.3 执行推理
-
-执行以下命令，直接使用飞桨框架的原生推理（仅支持Float32, 无需依赖TensorRT）:
+- MKLDNN预测：
 
-```
-export CUDA_VISIBLE_DEVICES=0
-python infer.py \
---image_file "./data/human_demo.jpg" \
---model_path "./pp_humanseg_fp32/model.pdmodel" \
---params_path "./pp_humanseg_fp32/model.pdiparams" \
---save_file "./humanseg_result_fp32.png" \
---dataset "human" \
---benchmark True \
---precision "fp32"
+```shell
+python paddle_inference_eval.py \
+      --model_path=pp_liteseg_qat \
+      --dataset='cityscape' \
+      --dataset_config=configs/dataset/cityscapes_1024x512_scale1.0.yml \
+      --device=CPU \
+      --use_mkldnn=True \
+      --precision=int8 \
+      --cpu_threads=10
 ```
 
-执行以下命令，使用Int8推理:
+#### 4.2 Paddle Inference 测试单张图片
 
-```
-export CUDA_VISIBLE_DEVICES=0
-python infer.py \
---image_file "./data/human_demo.jpg" \
---model_path "./pp_humanseg_qat/model.pdmodel" \
---params_path "./pp_humanseg_qat/model.pdiparams" \
---save_file "./humanseg_result_qat.png" \
---dataset "human" \
---benchmark True \
---use_trt True \
---precision "int8"
-```
+利用人像分割测试单张图片：
 
-执行以下命令，使用Paddle Inference在相应数据集上测试精度:
-
-```
-export CUDA_VISIBLE_DEVICES=0
-python infer.py \
---model_path "./pp_humanseg_qat/model.pdmodel" \
---params_path "./pp_humanseg_qat/model.pdiparams" \
---dataset_config configs/dataset/humanseg_dataset.yaml \
---use_trt True \
---precision "int8"
+```shell
+python paddle_inference_eval.py \
+      --model_path=pp_humanseg_qat \
+      --dataset='human' \
+       --image_file=./data/human_demo.jpg \
+      --use_trt=True \
+      --precision=int8
 ```
 
 <table><tbody>
@@ -287,17 +250,11 @@ Int8推理结果
 
 </tbody></table>
 
-执行以下命令查看更多关于 `infer.py` 使用说明：
-
-```
-python infer.py --help
-```
-
 
-### 5.4 更多部署教程
+### 4.3 更多部署教程
 
 - [Paddle Inference Python部署](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/deployment/inference/python_inference.md)
 - [Paddle Inference C++部署](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/deployment/inference/cpp_inference.md)
 - [Paddle Lite部署](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/deployment/lite/lite.md)
 
-## 6.FAQ
+## 5.FAQ

example/auto_compression/semantic_segmentation/infer.py → example/auto_compression/semantic_segmentation/paddle_inference_eval.py

@@ -12,11 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import cv2
-import numpy as np
 import argparse
 import time
-from tqdm import tqdm
+import os
+import sys
+import cv2
+import numpy as np
 import paddle
 import paddleseg.transforms as T
 from paddleseg.cvlibs import Config as PaddleSegDataConfig
@@ -38,31 +39,35 @@ def _transforms(dataset):
     return transforms
 
 
-def auto_tune_trt(args, data):
-    auto_tuned_shape_file = "./auto_tuning_shape"
-    pred_cfg = PredictConfig(args.model_path, args.params_path)
-    pred_cfg.enable_use_gpu(100, 0)
-    pred_cfg.collect_shape_range_info("./auto_tuning_shape")
-    predictor = create_predictor(pred_cfg)
-    input_names = predictor.get_input_names()
-    input_handle = predictor.get_input_handle(input_names[0])
-    input_handle.reshape(data.shape)
-    input_handle.copy_from_cpu(data)
-    predictor.run()
-    return auto_tuned_shape_file
-
-
-def load_predictor(args, data):
-    pred_cfg = PredictConfig(args.model_path, args.params_path)
-    pred_cfg.disable_glog_info()
+def load_predictor(args):
+    """
+    load predictor func
+    """
+    rerun_flag = False
+    model_file = os.path.join(args.model_path, args.model_filename)
+    params_file = os.path.join(args.model_path, args.params_filename)
+    pred_cfg = PredictConfig(model_file, params_file)
     pred_cfg.enable_memory_optim()
     pred_cfg.switch_ir_optim(True)
     if args.device == "GPU":
         pred_cfg.enable_use_gpu(100, 0)
+    else:
+        pred_cfg.disable_gpu()
+        pred_cfg.set_cpu_math_library_num_threads(args.cpu_threads)
+        if args.use_mkldnn:
+            pred_cfg.enable_mkldnn()
+            if args.precision == "int8":
+                pred_cfg.enable_mkldnn_int8({
+                    "conv2d", "depthwise_conv2d", "pool2d", "elementwise_mul"
+                })
 
     if args.use_trt:
         # To collect the dynamic shapes of inputs for TensorRT engine
-        auto_tuned_shape_file = auto_tune_trt(args, data)
+        dynamic_shape_file = os.path.join(args.model_path, "dynamic_shape.txt")
+        if os.path.exists(dynamic_shape_file):
+            pred_cfg.enable_tuned_tensorrt_dynamic_shape(dynamic_shape_file,
+                                                         True)
+            print("trt set dynamic shape done!")
             precision_map = {
                 "fp16": PrecisionType.Half,
                 "fp32": PrecisionType.Float32,
@@ -73,27 +78,33 @@ def load_predictor(args, data):
                 max_batch_size=1,
                 min_subgraph_size=4,
                 precision_mode=precision_map[args.precision],
-            use_static=False,
-            use_calib_mode=False)
-        allow_build_at_runtime = True
-        pred_cfg.enable_tuned_tensorrt_dynamic_shape(auto_tuned_shape_file,
-                                                     allow_build_at_runtime)
+                use_static=True,
+                use_calib_mode=False, )
+        else:
+            pred_cfg.disable_gpu()
+            pred_cfg.set_cpu_math_library_num_threads(10)
+            pred_cfg.collect_shape_range_info(dynamic_shape_file)
+            print("Start collect dynamic shape...")
+            rerun_flag = True
+
     predictor = create_predictor(pred_cfg)
-    return predictor
+    return predictor, rerun_flag
 
 
 def predict_image(args):
-
+    """
+    predict image func
+    """
     transforms = _transforms(args.dataset)
     transform = T.Compose(transforms)
 
     # Step1: Load image and preprocess
-    im = cv2.imread(args.image_file).astype('float32')
+    im = cv2.imread(args.image_file).astype("float32")
     data, _ = transform(im)
     data = np.array(data)[np.newaxis, :]
 
     # Step2: Prepare prdictor
-    predictor = load_predictor(args, data)
+    predictor, rerun_flag = load_predictor(args)
 
     # Step3: Inference
     input_names = predictor.get_input_names()
@@ -114,14 +125,21 @@ def predict_image(args):
     for i in range(repeats):
         predictor.run()
         results = output_handle.copy_to_cpu()
+        if rerun_flag:
+            print(
+                "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
+            )
+            return
     total_time = time.time() - start_time
     avg_time = float(total_time) / repeats
-    print(f"Average inference time: \033[91m{round(avg_time*1000, 2)}ms\033[0m")
+    print(
+        f"[Benchmark]Average inference time: \033[91m{round(avg_time*1000, 2)}ms\033[0m"
+    )
 
     # Step4: Post process
     if args.dataset == "human":
         results = reverse_transform(
-            paddle.to_tensor(results), im.shape, transforms, mode='bilinear')
+            paddle.to_tensor(results), im.shape, transforms, mode="bilinear")
         results = np.argmax(results, axis=1)
     result = get_pseudo_color_map(results[0])
 
@@ -132,8 +150,11 @@ def predict_image(args):
 
 
 def eval(args):
+    """
+    eval mIoU func
+    """
     # DataLoader need run on cpu
-    paddle.set_device('cpu')
+    paddle.set_device("cpu")
     data_cfg = PaddleSegDataConfig(args.dataset_config)
     eval_dataset = data_cfg.val_dataset
 
@@ -142,48 +163,56 @@ def eval(args):
     loader = paddle.io.DataLoader(
         eval_dataset,
         batch_sampler=batch_sampler,
-        num_workers=1,
+        num_workers=0,
         return_list=True)
 
-    total_iters = len(loader)
+    predictor, rerun_flag = load_predictor(args)
+
     intersect_area_all = 0
     pred_area_all = 0
     label_area_all = 0
-
-    print("Start evaluating (total_samples: {}, total_iters: {})...".format(
-        len(eval_dataset), total_iters))
-
-    init_predictor = False
-    for (image, label) in tqdm(loader):
-        label = np.array(label).astype('int64')
-        ori_shape = np.array(label).shape[-2:]
-        data = np.array(image)
-
-        if not init_predictor:
-            predictor = load_predictor(args, data)
-            init_predictor = True
-
     input_names = predictor.get_input_names()
     input_handle = predictor.get_input_handle(input_names[0])
-        input_handle.reshape(data.shape)
-        input_handle.copy_from_cpu(data)
-
-        predictor.run()
-
     output_names = predictor.get_output_names()
     output_handle = predictor.get_output_handle(output_names[0])
+    total_samples = len(eval_dataset)
+    sample_nums = len(loader)
+    batch_size = int(total_samples / sample_nums)
+    predict_time = 0.0
+    time_min = float("inf")
+    time_max = float("-inf")
+    print("Start evaluating (total_samples: {}, total_iters: {}).".format(
+        total_samples, sample_nums))
+    for batch_id, data in enumerate(loader):
+        image = np.array(data[0])
+        label = np.array(data[1]).astype("int64")
+        ori_shape = np.array(label).shape[-2:]
+        input_handle.reshape(image.shape)
+        input_handle.copy_from_cpu(image)
+        start_time = time.time()
+        predictor.run()
         results = output_handle.copy_to_cpu()
+        end_time = time.time()
+        timed = end_time - start_time
+        time_min = min(time_min, timed)
+        time_max = max(time_max, timed)
+        predict_time += timed
+        if rerun_flag:
+            print(
+                "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
+            )
+            return
 
         logit = reverse_transform(
             paddle.to_tensor(results),
             ori_shape,
             eval_dataset.transforms.transforms,
-            mode='bilinear')
+            mode="bilinear")
         pred = paddle.to_tensor(logit)
         if len(
                 pred.shape
         ) == 4:  # for humanseg model whose prediction is distribution but not class id
-            pred = paddle.argmax(pred, axis=1, keepdim=True, dtype='int32')
+            pred = paddle.argmax(pred, axis=1, keepdim=True, dtype="int32")
 
         intersect_area, pred_area, label_area = metrics.calculate_area(
             pred,
@@ -193,71 +222,95 @@ def eval(args):
         intersect_area_all = intersect_area_all + intersect_area
         pred_area_all = pred_area_all + pred_area
         label_area_all = label_area_all + label_area
+        if batch_id % 100 == 0:
+            print("Eval iter:", batch_id)
+            sys.stdout.flush()
 
-    class_iou, miou = metrics.mean_iou(intersect_area_all, pred_area_all,
+    _, miou = metrics.mean_iou(intersect_area_all, pred_area_all,
                                label_area_all)
-    class_acc, acc = metrics.accuracy(intersect_area_all, pred_area_all)
+    _, acc = metrics.accuracy(intersect_area_all, pred_area_all)
     kappa = metrics.kappa(intersect_area_all, pred_area_all, label_area_all)
-    class_dice, mdice = metrics.dice(intersect_area_all, pred_area_all,
-                                     label_area_all)
-
-    infor = "[EVAL] #Images: {} mIoU: {:.4f} Acc: {:.4f} Kappa: {:.4f} Dice: {:.4f}".format(
-        len(eval_dataset), miou, acc, kappa, mdice)
+    _, mdice = metrics.dice(intersect_area_all, pred_area_all, label_area_all)
+
+    time_avg = predict_time / sample_nums
+    print(
+        "[Benchmark]Batch size: {}, Inference time(ms): min={}, max={}, avg={}".
+        format(batch_size,
+               round(time_min * 1000, 2),
+               round(time_max * 1000, 1), round(time_avg * 1000, 1)))
+    infor = "[Benchmark] #Images: {} mIoU: {:.4f} Acc: {:.4f} Kappa: {:.4f} Dice: {:.4f}".format(
+        total_samples, miou, acc, kappa, mdice)
     print(infor)
+    sys.stdout.flush()
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
 
     parser = argparse.ArgumentParser()
     parser.add_argument(
-        '--image_file',
+        "--model_path", type=str, help="inference model filepath")
+    parser.add_argument(
+        "--model_filename",
+        type=str,
+        default="model.pdmodel",
+        help="model file name")
+    parser.add_argument(
+        "--params_filename",
+        type=str,
+        default="model.pdiparams",
+        help="params file name")
+    parser.add_argument(
+        "--image_file",
         type=str,
         default=None,
         help="Image path to be processed.")
     parser.add_argument(
-        '--save_file',
+        "--save_file",
         type=str,
         default=None,
         help="The path to save the processed image.")
     parser.add_argument(
-        '--model_path', type=str, help="Inference model filepath.")
-    parser.add_argument(
-        '--params_path', type=str, help="Inference parameters filepath.")
-    parser.add_argument(
-        '--dataset',
+        "--dataset",
         type=str,
         default="human",
         choices=["human", "cityscape"],
-        help="The type of given image which can be 'human' or 'cityscape'.")
+        help="The type of given image which can be 'human' or 'cityscape'.", )
     parser.add_argument(
-        '--dataset_config',
+        "--dataset_config",
         type=str,
         default=None,
         help="path of dataset config.")
     parser.add_argument(
-        '--benchmark',
+        "--benchmark",
         type=bool,
         default=False,
         help="Whether to run benchmark or not.")
     parser.add_argument(
-        '--use_trt',
+        "--use_trt",
         type=bool,
         default=False,
         help="Whether to use tensorrt engine or not.")
     parser.add_argument(
-        '--device',
+        "--device",
         type=str,
-        default='GPU',
+        default="GPU",
         choices=["CPU", "GPU"],
-        help="Choose the device you want to run, it can be: CPU/GPU, default is GPU"
+        help="Choose the device you want to run, it can be: CPU/GPU, default is GPU",
     )
     parser.add_argument(
-        '--precision',
+        "--precision",
         type=str,
-        default='fp32',
+        default="fp32",
         choices=["fp32", "fp16", "int8"],
-        help="The precision of inference. It can be 'fp32', 'fp16' or 'int8'. Default is 'fp16'."
+        help="The precision of inference. It can be 'fp32', 'fp16' or 'int8'. Default is 'fp16'.",
     )
+    parser.add_argument(
+        "--use_mkldnn",
+        type=bool,
+        default=False,
+        help="Whether use mkldnn or not.")
+    parser.add_argument(
+        "--cpu_threads", type=int, default=1, help="Num of cpu threads.")
     args = parser.parse_args()
     if args.image_file:
         predict_image(args)

paddleslim/common/load_model.py

@@ -230,7 +230,6 @@ def export_onnx(model_dir,
         opset_version=opset_version,
         enable_onnx_checker=True,
         deploy_backend=deploy_backend,
-        scale_file=os.path.join(model_dir, 'calibration_table.txt'),
         calibration_file=os.path.join(
             save_file_path.rstrip(os.path.split(save_file_path)[-1]),
             'calibration.cache'))