optim loss

README.md

@@ -16,7 +16,7 @@ Network|mAPval 0.5|mAPval 0.5:0.95|Resolution|Run Time(4xCore)|Run Time(1xCore)|
 [nanodet_m](https://github.com/RangiLyu/nanodet)|-|20.6%|320X320|49.24ms|160.35ms|0.95M
 [yolo-fastestv1.1](https://github.com/dog-qiuqiu/Yolo-Fastest/tree/master/ModelZoo/yolo-fastest-1.1_coco)|24.40%|-|320X320|26.60ms|75.74ms|0.35M
 [yolo-fastestv2](https://github.com/dog-qiuqiu/Yolo-FastestV2/tree/main/modelzoo)|24.10%|-|352X352|23.8ms|68.9ms|0.25M
-FastestDet|25.0%|12.3%|352X352|23.51ms|70.62ms|0.24M
+FastestDet|25.3%|13.0%|352X352|23.51ms|70.62ms|0.24M
 * ***Test platform Radxa Rock3A RK3568 ARM Cortex-A55 CPU，Based on [NCNN](https://github.com/Tencent/ncnn)***
 * ***CPU lock frequency 2.0GHz***
 # Improvement
@@ -40,7 +40,7 @@ Intel|i7-8700(X86-cpu)|Linux(amd64)|ncnn|4.51ms|4.33ms
 ## Test
 * Picture test
   ```
-  python3 test.py --yaml configs/coco.yaml --weight weights/coco_ap05_0.250_280epoch.pth --img data/3.jpg
+  python3 test.py --yaml configs/coco.yaml --weight weights/weight_AP05:0.253207_280-epoch.pth --img data/3.jpg
   ```
 <div align=center>
 <img src="https://github.com/dog-qiuqiu/FastestDet/blob/main/result.png"> />
@@ -148,7 +148,7 @@ TRAIN:
 ### Evaluation
 * Calculate map evaluation
   ```
-  python3 eval.py --yaml configs/coco.yaml --weight weights/coco_ap05_0.250_280epoch.pth
+  python3 eval.py --yaml configs/coco.yaml --weight weights/weight_AP05:0.253207_280-epoch.pth
   ```
 * COCO2017 evaluation
   ```
@@ -161,30 +161,30 @@ TRAIN:
   DONE (t=30.85s).
   Accumulating evaluation results...
   DONE (t=4.97s).
-  Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.123
-  Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.250
-  Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.109
-  Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.017
-  Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.115
-  Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.238
-  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.139
-  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.199
-  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.205
-  Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.035
-  Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.218
-  Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.374
+  Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.130
+  Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.253
+  Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.119
+  Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.021
+  Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.129
+  Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.237
+  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.142
+  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.208
+  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.214
+  Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.043
+  Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.236
+  Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.372
 
   ```
 # Deploy
 ## Export onnx
 * You can export .onnx by adding the --onnx option when executing test.py
   ```
-  python3 test.py --yaml configs/coco.yaml --weight weights/coco_ap05_0.250_280epoch.pth --img data/3.jpg --onnx
+  python3 test.py --yaml configs/coco.yaml --weight weights/weight_AP05:0.253207_280-epoch.pth --img data/3.jpg --onnx
   ```
 ## Export torchscript
 * You can export .pt by adding the --torchscript option when executing test.py
   ```
-  python3 test.py --yaml configs/coco.yaml --weight weights/coco_ap05_0.250_280epoch.pth --img data/3.jpg --torchscript
+  python3 test.py --yaml configs/coco.yaml --weight weights/weight_AP05:0.253207_280-epoch.pth --img data/3.jpg --torchscript
   ```
 ## NCNN
 * Need to compile ncnn and opencv in advance and modify the path in build.sh

example/onnx-runtime/runtime.py

@@ -20,7 +20,7 @@ def preprocess(src_img, size):
     return output.astype('float32')
 
 # nms算法
-def nms(dets, thresh=0.35):
+def nms(dets, thresh=0.45):
     # dets:N*M,N是bbox的个数，M的前4位是对应的（x1,y1,x2,y2），第5位是对应的分数
     # #thresh:0.3,0.5....
     x1 = dets[:, 0]
@@ -89,7 +89,7 @@ def detection(session, img, input_width, input_height, thresh):
 
             # 解析检测框置信度
             obj_score, cls_score = data[0], data[5:].max()
-            score = obj_score * cls_score
+            score = (obj_score ** 0.6) * (cls_score ** 0.4)
 
             # 阈值筛选
             if score > thresh:
@@ -114,14 +114,14 @@ def detection(session, img, input_width, input_height, thresh):
 
 if __name__ == '__main__':
     # 读取图片
-    img = cv2.imread("1.jpg")
+    img = cv2.imread("3.jpg")
     # 模型输入的宽高
     input_width, input_height = 352, 352
     # 加载模型
     session = onnxruntime.InferenceSession('FastestDet.onnx')
     # 目标检测
     start = time.perf_counter()
-    bboxes = detection(session, img, input_width, input_height, 0.8)
+    bboxes = detection(session, img, input_width, input_height, 0.65)
     end = time.perf_counter()
     time = (end - start) * 1000.
     print("forward time:%fms"%time)

module/loss.py

@@ -94,7 +94,8 @@ class DetectorLoss(nn.Module):
 
         # 定义obj和cls的损失函数
         BCEcls = nn.NLLLoss() 
-        BCEobj = nn.BCELoss(reduction='none')
+        # smmoth L1相比于bce效果最好
+        BCEobj = nn.SmoothL1Loss(reduction='none')
         
         # 构建ground truth
         gt_box, gt_cls, ps_index = self.build_target(preds, targets)
@@ -134,7 +135,8 @@ class DetectorLoss(nn.Module):
             ps = torch.log(pcls[b, gy, gx])
             cls_loss = BCEcls(ps, gt_cls[0][f])
 
-            tobj[b, gy, gx] = 1.0
+            # iou aware
+            tobj[b, gy, gx] = iou.float()
             # 统计每个图片正样本的数量
             n = torch.bincount(b)
             factor[b, gy, gx] =  (1. / (n[b] / (H * W))) * 0.25

test.py

@@ -44,7 +44,7 @@ if __name__ == '__main__':
     # 模型加载
     print("load weight from:%s"%opt.weight)
     model = Detector(cfg.category_num, True).to(device)
-    model.load_state_dict(torch.load(opt.weight))
+    model.load_state_dict(torch.load(opt.weight, map_location=device))
     #sets the module in eval node
     model.eval()
     

utils/datasets.py

@@ -5,7 +5,7 @@ import numpy as np
 import torch
 import random
 
-def random_scale(image, boxes):
+def random_crop(image, boxes):
     height, width, _ = image.shape
     # random crop imgage
     cw, ch = random.randint(int(width * 0.75), width), random.randint(int(height * 0.75), height)
@@ -29,6 +29,30 @@ def random_scale(image, boxes):
 
     return roi, output
 
+def random_narrow(image, boxes):
+    height, width, _ = image.shape
+    # random narrow
+    cw, ch = random.randint(width, int(width * 1.25)), random.randint(height, int(height * 1.25))
+    cx, cy = random.randint(0, cw - width), random.randint(0, ch - height)
+
+    background = np.ones((ch, cw, 3), np.uint8) * 128
+    background[cy:cy + height, cx:cx + width] = image
+
+    output = []
+    for box in boxes:
+        index, category = box[0], box[1]
+        bx, by = box[2] * width, box[3] * height
+        bw, bh = box[4] * width, box[5] * height
+
+        bx, by = (bx + cx)/cw, (by + cy)/ch
+        bw, bh = bw/cw, bh/ch
+
+        output.append([index, category, bx, by, bw, bh])
+
+    output = np.array(output, dtype=float)
+
+    return background, output
+
 def collate_fn(batch):
     img, label = zip(*batch)
     for i, l in enumerate(label):
@@ -84,7 +108,10 @@ class TensorDataset():
 
         # 是否进行数据增强
         if self.aug:
-            img, label = random_scale(img, label)
+            if random.randint(1, 10) % 2 == 0:
+                img, label = random_narrow(img, label)
+            else:
+                img, label = random_crop(img, label)
 
         img = cv2.resize(img, (self.img_width, self.img_height), interpolation = cv2.INTER_LINEAR) 
 

utils/tool.py

@@ -72,7 +72,7 @@ def handle_preds(preds, device, conf_thresh=0.25, nms_thresh=0.45):
     pcls = pred[:, :, :, 5:]
 
     # 检测框置信度
-    bboxes[..., 4] = pobj.squeeze(-1) * pcls.max(dim=-1)[0]
+    bboxes[..., 4] = (pobj.squeeze(-1) ** 0.6) * (pcls.max(dim=-1)[0] ** 0.4)
     bboxes[..., 5] = pcls.argmax(dim=-1)
 
     # 检测框的坐标