add convert openvino example

Author: wz

.gitignore

@@ -15,6 +15,7 @@ flower_data
 *.xml
 *.bin
 *.mapping
+*.csv
 checkpoint
 data
 VOCdevkit

deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/README.md

@@ -0,0 +1,15 @@
+本项目展示如何将Pytorch中的ResNet34网络转成Openvino的IR格式，并进行量化处理，具体使用流程如下：
+1. 按照`requirements.txt`配置环境
+2. 下载事先训练好的ResNet34权重（之前在花分类数据集上训练得到的）放在当前文件夹下。百度云链接: https://pan.baidu.com/s/1x4WFX1HynYcXLium3UaaFQ  密码: qvi6
+3. 使用`convert_pytorch2onnx.py`将Resnet34转成ONNX格式
+4. 在命令行中使用以下指令将ONNX转成IR格式：
+```
+mo  --input_model resnet34.onnx \
+    --input_shape "[1,3,224,224]" \
+    --mean_values="[123.675,116.28,103.53]" \
+    --scale_values="[58.395,57.12,57.375]" \
+    --data_type FP32 \
+    --output_dir ir_output
+```
+5. 下载并解压花分类数据集，将`quantization_int8.py`中的`data_path`指向解压后的`flower_photos`
+6. 使用`quantization_int8.py`量化模型

deploying_service/deploying_pytorch/convert_openvino/compare_fps.py renamed to deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/compare_fps.py

@@ -1,9 +1,9 @@
-import cv2
 import time
 import numpy as np
 import torch
+import matplotlib.pyplot as plt
 from openvino.runtime import Core
-from model import mobilenet_v3_large
+from torchvision.models import resnet34
 
 
 def normalize(image: np.ndarray) -> np.ndarray:
@@ -39,6 +39,8 @@ def onnx_inference(onnx_path: str, image: np.ndarray, num_images: int = 20):
         f"seconds per image, FPS: {num_images / time_onnx:.2f}"
     )
 
+    return num_images / time_onnx
+
 
 def ir_inference(ir_path: str, image: np.ndarray, num_images: int = 20):
     # Load the network in Inference Engine
@@ -61,12 +63,13 @@ def ir_inference(ir_path: str, image: np.ndarray, num_images: int = 20):
         f"seconds per image, FPS: {num_images / time_ir:.2f}"
     )
 
+    return num_images / time_ir
+
 
-def pytorch_inference(weights_path: str, image: np.ndarray, num_images: int = 20):
+def pytorch_inference(image: np.ndarray, num_images: int = 20):
     image = torch.as_tensor(image, dtype=torch.float32)
 
-    model = mobilenet_v3_large(num_classes=5)
-    model.load_state_dict(torch.load(weights_path, map_location='cpu'))
+    model = resnet34(pretrained=False, num_classes=5)
     model.eval()
 
     with torch.no_grad():
@@ -81,29 +84,45 @@ def pytorch_inference(weights_path: str, image: np.ndarray, num_images: int = 20
         f"FPS: {num_images / time_torch:.2f}"
     )
 
+    return num_images / time_torch
+
+
+def plot_fps(v: dict):
+    x = list(v.keys())
+    y = list(v.values())
+
+    plt.bar(range(len(x)), y, align='center')
+    plt.xticks(range(len(x)), x)
+    for i, v in enumerate(y):
+        plt.text(x=i, y=v+0.5, s=f"{v:.2f}", ha='center')
+    plt.xlabel('model format')
+    plt.ylabel('fps')
+    plt.title('FPS comparison')
+    plt.show()
+    plt.savefig('fps_vs.jpg')
+
 
 def main():
     image_h = 224
     image_w = 224
-    image_filename = "test.jpg"
-    onnx_path = "mobilenet_v3.onnx"
-    ir_path = "ir_output/mobilenet_v3.xml"
-    pytorch_weights_path = "mbv3_flower.pth"
-
-    image = cv2.cvtColor(cv2.imread(image_filename), cv2.COLOR_BGR2RGB)
+    onnx_path = "resnet34.onnx"
+    ir_path = "ir_output/resnet34.xml"
 
-    resized_image = cv2.resize(image, (image_w, image_h))
-    normalized_image = normalize(resized_image)
+    image = np.random.randn(image_h, image_w, 3)
+    normalized_image = normalize(image)
 
     # Convert the resized images to network input shape
     # [h, w, c] -> [c, h, w] -> [1, c, h, w]
-    input_image = np.expand_dims(np.transpose(resized_image, (2, 0, 1)), 0)
+    input_image = np.expand_dims(np.transpose(image, (2, 0, 1)), 0)
     normalized_input_image = np.expand_dims(np.transpose(normalized_image, (2, 0, 1)), 0)
 
-    onnx_inference(onnx_path, normalized_input_image, num_images=50)
-    ir_inference(ir_path, input_image, num_images=50)
-    pytorch_inference(pytorch_weights_path, normalized_input_image, num_images=50)
+    onnx_fps = onnx_inference(onnx_path, normalized_input_image, num_images=50)
+    ir_fps = ir_inference(ir_path, input_image, num_images=50)
+    pytorch_fps = pytorch_inference(normalized_input_image, num_images=50)
+    plot_fps({"pytorch": round(pytorch_fps, 2),
+              "onnx": round(onnx_fps, 2),
+              "ir": round(ir_fps, 2)})
 
 
 if __name__ == '__main__':
-    main()
+    main()

deploying_service/deploying_pytorch/convert_openvino/compare_onnx_and_ir.py renamed to deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/compare_onnx_and_ir.py

@@ -1,4 +1,3 @@
-import cv2
 import numpy as np
 from openvino.runtime import Core
 
@@ -48,18 +47,15 @@ def ir_inference(ir_path: str, image: np.ndarray):
 def main():
     image_h = 224
     image_w = 224
-    image_filename = "test.jpg"
-    onnx_path = "mobilenet_v3.onnx"
-    ir_path = "ir_output/mobilenet_v3.xml"
+    onnx_path = "resnet34.onnx"
+    ir_path = "ir_output/resnet34.xml"
 
-    image = cv2.cvtColor(cv2.imread(image_filename), cv2.COLOR_BGR2RGB)
-
-    resized_image = cv2.resize(image, (image_w, image_h))
-    normalized_image = normalize(resized_image)
+    image = np.random.randn(image_h, image_w, 3)
+    normalized_image = normalize(image)
 
     # Convert the resized images to network input shape
     # [h, w, c] -> [c, h, w] -> [1, c, h, w]
-    input_image = np.expand_dims(np.transpose(resized_image, (2, 0, 1)), 0)
+    input_image = np.expand_dims(np.transpose(image, (2, 0, 1)), 0)
     normalized_input_image = np.expand_dims(np.transpose(normalized_image, (2, 0, 1)), 0)
 
     onnx_res = onnx_inference(onnx_path, normalized_input_image)

deploying_service/deploying_pytorch/convert_openvino/convert_pytorch2onnx.py renamed to deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/convert_pytorch2onnx.py

@@ -3,7 +3,7 @@
 import onnx
 import onnxruntime
 import numpy as np
-from model import mobilenet_v3_large
+from torchvision.models import resnet34
 
 device = torch.device("cpu")
 
@@ -13,15 +13,15 @@ def to_numpy(tensor):
 
 
 def main():
-    weights_path = "mbv3_flower.pth"
-    onnx_file_name = "mobilenet_v3.onnx"
+    weights_path = "resNet34(flower).pth"
+    onnx_file_name = "resnet34.onnx"
     batch_size = 1
     img_h = 224
     img_w = 224
     img_channel = 3
 
     # create model and load pretrain weights
-    model = mobilenet_v3_large(num_classes=5)
+    model = resnet34(pretrained=False, num_classes=5)
     model.load_state_dict(torch.load(weights_path, map_location='cpu'))
 
     model.eval()

deploying_service/deploying_pytorch/convert_openvino/model.py renamed to deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/model.py

@@ -3,20 +3,20 @@
 from compression.graph import load_model, save_model
 from compression.graph.model_utils import compress_model_weights
 from compression.pipeline.initializer import create_pipeline
-from utils import MyDataSet, Accuracy, read_split_data
+from utils import MyDataLoader, Accuracy, read_split_data
 
 
 def main():
     data_path = "/data/flower_photos"
-    ir_model_xml = "ir_output/mobilenet_v3.xml"
-    ir_model_bin = "ir_output/mobilenet_v3.bin"
+    ir_model_xml = "ir_output/resnet34.xml"
+    ir_model_bin = "ir_output/resnet34.bin"
     save_dir = "quant_ir_output"
-    model_name = "quantized_mobilenet_v3"
+    model_name = "quantized_resnet34"
     img_w = 224
     img_h = 224
 
     model_config = Dict({
-        'model_name': 'mobilenet_v3',
+        'model_name': 'resnet34',
         'model': ir_model_xml,
         'weights': ir_model_bin
     })
@@ -45,7 +45,7 @@ def main():
 
     # Step 2: Initialize the data loader.
     _, _, val_images_path, val_images_label = read_split_data(data_path, val_rate=0.2)
-    data_loader = MyDataSet(dataset_config, val_images_path, val_images_label, img_w, img_h)
+    data_loader = MyDataLoader(dataset_config, val_images_path, val_images_label, img_w, img_h)
 
     # Step 3 (Optional. Required for AccuracyAwareQuantization): Initialize the metric.
     metric = Accuracy(top_k=1)

deploying_service/deploying_pytorch/convert_openvino/quantization_int8.py renamed to deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/quantization_int8.py

@@ -3,4 +3,5 @@ torchvision==0.12.0
 onnx==1.9.0
 onnxruntime==1.8.0
 protobuf==3.19.4
-openvino-dev==2022.1.0
+openvino-dev==2022.1.0
+matplotlib

deploying_service/deploying_pytorch/convert_openvino/requirements.txt renamed to deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/requirements.txt

@@ -2,9 +2,10 @@
 import json
 import random
 
-import cv2
+from PIL import Image
 import numpy as np
 from compression.api import DataLoader, Metric
+from torchvision.transforms import transforms
 
 
 def read_split_data(root: str, val_rate: float = 0.2):
@@ -57,7 +58,6 @@ def read_split_data(root: str, val_rate: float = 0.2):
 
 # Custom implementation of classification accuracy metric.
 class Accuracy(Metric):
-
     # Required methods
     def __init__(self, top_k=1):
         super().__init__()
@@ -104,13 +104,17 @@ def get_attributes(self):
                              'type': 'accuracy'}}
 
 
-class MyDataSet(DataLoader):
+class MyDataLoader(DataLoader):
     def __init__(self, cfg, images_path: list, images_label: list, img_w: int = 224, img_h: int = 224):
         super().__init__(cfg)
         self.images_path = images_path
         self.images_label = images_label
         self.image_w = img_w
         self.image_h = img_h
+        self.transforms = transforms.Compose([
+            transforms.Resize(min(img_h, img_w)),
+            transforms.CenterCrop((img_h, img_w))
+        ])
 
     def __len__(self):
         return len(self.images_label)
@@ -123,11 +127,11 @@ def __getitem__(self, index):
         if index >= len(self):
             raise IndexError
 
-        image = cv2.cvtColor(cv2.imread(self.images_path[index]), cv2.COLOR_BGR2RGB)
-        resized_image = cv2.resize(image, (self.image_w, self.image_h))
+        img = Image.open(self.images_path[index])
+        img = self.transforms(img)
 
         # Convert the resized images to network input shape
         # [h, w, c] -> [c, h, w] -> [1, c, h, w]
-        img = np.expand_dims(np.transpose(resized_image, (2, 0, 1)), 0)
+        img = np.expand_dims(np.transpose(np.array(img), (2, 0, 1)), 0)
 
         return (index, self.images_label[index]), img

deploying_service/deploying_pytorch/convert_openvino/utils.py renamed to deploying_service/deploying_pytorch/convert_openvino/convert_resnet34/utils.py

@@ -0,0 +1,61 @@
+OpenVINO量化YOLOv5
+
+1. 按照`requirements.txt`配置环境
+2. 将YOLOv5转为ONNX
+YOLOv5官方有提供导出ONNX以及OpenVINO的方法，但我这里仅导出成ONNX，这里以YOLOv5s为例
+```
+python export.py --weights yolov5s.pt --include onnx
+```
+
+3. ONNX转换为IR
+使用OpenVINO的`mo`工具将ONNX转为OpenVINO的IR格式
+```
+mo  --input_model yolov5s.onnx \
+    --input_shape "[1,3,640,640]" \
+    --scale 255 \
+    --data_type FP32 \
+    --output_dir ir_output
+```
+
+4. 量化模型
+使用`quantization_int8.py`进行模型的量化，量化过程中需要使用到COCO2017数据集，需要将`data_path`指向coco2017目录
+```
+├── coco2017: 数据集根目录
+     ├── train2017: 所有训练图像文件夹(118287张)
+     ├── val2017: 所有验证图像文件夹(5000张)
+     └── annotations: 对应标注文件夹
+              ├── instances_train2017.json: 对应目标检测、分割任务的训练集标注文件
+              ├── instances_val2017.json: 对应目标检测、分割任务的验证集标注文件
+              ├── captions_train2017.json: 对应图像描述的训练集标注文件
+              ├── captions_val2017.json: 对应图像描述的验证集标注文件
+              ├── person_keypoints_train2017.json: 对应人体关键点检测的训练集标注文件
+              └── person_keypoints_val2017.json: 对应人体关键点检测的验证集标注文件夹
+```
+
+5. benchmark
+直接利用`benchmark_app`工具测试量化前后的`Throughput`，这里以`CPU: Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz`设备为例
+```
+benchmark_app -m ir_output/yolov5s.xml -d CPU -api sync
+```
+output：
+```
+Latency:
+    Median:     59.56 ms
+    AVG:        63.30 ms
+    MIN:        57.88 ms
+    MAX:        99.89 ms
+Throughput: 16.79 FPS
+```
+
+```
+benchmark_app -m quant_ir_output/quantized_yolov5s.xml -d CPU -api sync
+```
+output:
+```
+Latency:
+    Median:     42.97 ms
+    AVG:        46.56 ms
+    MIN:        41.18 ms
+    MAX:        95.75 ms
+Throughput: 23.27 FPS
+```