docs(ssd): 更新测试流程和使用示例

zjZSTU · zjZSTU · commit 9e3b366047bb · 2020-05-17T20:42:03.000+08:00
diff --git a/docs/index.md b/docs/index.md
@@ -30,4 +30,7 @@
         * 训练日志
     * 测试
         * 测试流程
-        * 测试日志
+        * 测试日志
+    * 使用
+        * 使用示例
+        * 使用日志
diff --git a/docs/ssd/demo_process.md b/docs/ssd/demo_process.md
@@ -0,0 +1,12 @@
+
+# 使用示例
+
+* `py/demo.py`
+
+使用示例和测试示例很相似，惟一区别在于其完成测试流程后进一步使用分类概率过滤，默认设置过滤阈值为`0.7`
+
+![](./imgs/003123.jpg)
+
+![](./imgs/004101.jpg)
+
+![](./imgs/008591.jpg)
diff --git a/docs/ssd/demo_res.md b/docs/ssd/demo_res.md
@@ -1,9 +1,9 @@
 
-# 测试日志
+# 使用日志
 
 测试检测结果，计算`FPS`
 
-## 测试日志
+## 使用日志
 
 ```
 $ python demo.py --config-file configs/vgg_ssd300_voc0712.yaml --images_dir demo --ckpt outputs/vgg_ssd300_voc0712/model_final.pth 
diff --git a/docs/ssd/imgs/003123.jpg b/docs/ssd/imgs/003123.jpg
diff --git a/docs/ssd/imgs/004101.jpg b/docs/ssd/imgs/004101.jpg
diff --git a/docs/ssd/imgs/008591.jpg b/docs/ssd/imgs/008591.jpg
diff --git a/docs/ssd/test_process.md b/docs/ssd/test_process.md
@@ -0,0 +1,24 @@
+
+# 测试流程
+
+* `py/ssd/test.py`
+
+## 测试参数
+
+* 数据集：`PASCAL VOC 07 test`
+* 批量大小：`10`
+* 目标检测：
+    * 中心坐标的标准方差（`CENTER_VARIANCE`）：`0.1`
+    * 长宽的标准方差（`SIZE_VARIANCE`）：`0.2`
+* `NMS`阈值（`NMS_THRESHOLD`）：`0.5`
+* 分类概率阈值（`CONFIDENCE_THRESHOLD`）：`0.01`
+* 保留预测框个数（`MAX_PER_IMAGE`）：`100`
+
+## 测试流程
+
+* 载入测试图像
+* 图像预处理：缩放到指定大小，减去图像均值，转换成`Tensor`格式
+* 模型计算：生成检测框偏移和检测框类别置信度
+* 结合先验框、检测框偏移以及目标检测参数得到检测框
+* 对逐个图像执行`NMS`操作以及分类阈值过滤
+* 保留前几个预测框
diff --git a/docs/ssd/test_res.md b/docs/ssd/test_res.md
@@ -0,0 +1,103 @@
+
+# 测试日志
+
+```
+$ python test.py --config-file configs/vgg_ssd300_voc0712.yaml 
+2020-05-17 20:17:43,458 SSD INFO: Using 1 GPUs
+2020-05-17 20:17:43,458 SSD INFO: Namespace(ckpt=None, config_file='configs/vgg_ssd300_voc0712.yaml', local_rank=0, opts=[], output_dir='eval_results')
+2020-05-17 20:17:43,458 SSD INFO: Loaded configuration file configs/vgg_ssd300_voc0712.yaml
+2020-05-17 20:17:43,458 SSD INFO: 
+MODEL:
+  NUM_CLASSES: 21
+INPUT:
+  IMAGE_SIZE: 300
+DATASETS:
+  TRAIN: ("voc_2007_trainval", "voc_2012_trainval")
+  TEST: ("voc_2007_test", )
+SOLVER:
+  MAX_ITER: 120000
+  LR_STEPS: [80000, 100000]
+  GAMMA: 0.1
+  BATCH_SIZE: 32
+  LR: 1e-3
+
+OUTPUT_DIR: 'outputs/vgg_ssd300_voc0712'
+2020-05-17 20:17:43,458 SSD INFO: Running with config:
+DATASETS:
+  TEST: ('voc_2007_test',)
+  TRAIN: ('voc_2007_trainval', 'voc_2012_trainval')
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+INPUT:
+  IMAGE_SIZE: 300
+  PIXEL_MEAN: [123, 117, 104]
+MODEL:
+  BACKBONE:
+    NAME: vgg
+    OUT_CHANNELS: (512, 1024, 512, 256, 256, 256)
+    PRETRAINED: True
+  BOX_HEAD:
+    NAME: SSDBoxHead
+    PREDICTOR: SSDBoxPredictor
+  CENTER_VARIANCE: 0.1
+  DEVICE: cuda
+  META_ARCHITECTURE: SSDDetector
+  NEG_POS_RATIO: 3
+  NUM_CLASSES: 21
+  PRIORS:
+    ASPECT_RATIOS: [[2], [2, 3], [2, 3], [2, 3], [2], [2]]
+    BOXES_PER_LOCATION: [4, 6, 6, 6, 4, 4]
+    CLIP: True
+    FEATURE_MAPS: [38, 19, 10, 5, 3, 1]
+    MAX_SIZES: [60, 111, 162, 213, 264, 315]
+    MIN_SIZES: [30, 60, 111, 162, 213, 264]
+    STRIDES: [8, 16, 32, 64, 100, 300]
+  SIZE_VARIANCE: 0.2
+  THRESHOLD: 0.5
+OUTPUT_DIR: outputs/vgg_ssd300_voc0712
+SOLVER:
+  BATCH_SIZE: 32
+  GAMMA: 0.1
+  LR: 0.001
+  LR_STEPS: [80000, 100000]
+  MAX_ITER: 120000
+  MOMENTUM: 0.9
+  WARMUP_FACTOR: 0.3333333333333333
+  WARMUP_ITERS: 500
+  WEIGHT_DECAY: 0.0005
+TEST:
+  BATCH_SIZE: 10
+  CONFIDENCE_THRESHOLD: 0.01
+  MAX_PER_CLASS: -1
+  MAX_PER_IMAGE: 100
+  NMS_THRESHOLD: 0.45
+2020-05-17 20:17:44,736 SSD.inference INFO: Loading checkpoint from outputs/vgg_ssd300_voc0712/model_final.pth
+2020-05-17 20:17:44,834 SSD.inference INFO: Evaluating voc_2007_test dataset(4952 images):
+  0%|                                                                                                                                                                                      | 0/496 [00:00<?, ?it/s]/opt/conda/conda-bld/pytorch_1587428266983/work/torch/csrc/utils/python_arg_parser.cpp:756: UserWarning: This overload of nonzero is deprecated:
+	nonzero(Tensor input, *, Tensor out)
+Consider using one of the following signatures instead:
+	nonzero(Tensor input, *, bool as_tuple)
+100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 496/496 [00:41<00:00, 11.90it/s]
+2020-05-17 20:18:28,805 SSD.inference INFO: mAP: 0.7740
+aeroplane       : 0.8113
+bicycle         : 0.8322
+bird            : 0.7476
+boat            : 0.7160
+bottle          : 0.5331
+bus             : 0.8619
+car             : 0.8669
+cat             : 0.8781
+chair           : 0.6243
+cow             : 0.8297
+diningtable     : 0.7607
+dog             : 0.8393
+horse           : 0.8634
+motorbike       : 0.8420
+person          : 0.7972
+pottedplant     : 0.5075
+sheep           : 0.7754
+sofa            : 0.7775
+train           : 0.8548
+tvmonitor       : 0.7622
+```
diff --git a/docs/ssd/train_process.md b/docs/ssd/train_process.md
@@ -23,6 +23,9 @@
 * 模型
     * 基础网络：`VGG16`
     * 损失函数：`SmoothL1 Loss + Softmax Loss`
+* 目标检测：
+    * 中心坐标的标准方差（`CENTER_VARIANCE`）：`0.1`
+    * 长宽的标准方差（`SIZE_VARIANCE`）：`0.2`
 
 ## 训练流程
 
diff --git a/mkdocs.yml b/mkdocs.yml
@@ -82,4 +82,5 @@ nav:
                 - 测试流程: ssd/test_process.md
                 - 测试日志: ssd/test_res.md
             - 使用:
+                - 使用示例: ssd/demo_process.md
                 - 使用日志: ssd/demo_res.md
