YOLOv3-tiny 和 DarkNet - 2 类，但仅分类 1

Question

我想用 yolov3-tiny 创建一个检测汽车和摩托车的网络，所以我创建了自己的数据集，其中包含 100 个汽车图像和 100 个摩托车图像。在注释文件中，标签是正确的（0 表示汽车，1 表示摩托车）。由于我有 2 个类，我已将 yolov3-tiny.cfg 修改为将类设置为 2，将过滤器设置为 (classes+5)*3 = 21。我还将批处理修改为 24，细分为 8。这是完整的cfg 文件（./custom/car_moto-yolov3-tiny.cfg）：

[net]
# Testing
batch=24
subdivisions=8
# Training
# batch=64
# subdivisions=2
width=416
height=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1

learning_rate=0.001
burn_in=1000
max_batches = 5000
policy=steps
steps=400000,450000
scales=.1,.1

[convolutional]
batch_normalize=1
filters=16
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=1

[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky

###########

[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

[convolutional]
size=1
stride=1
pad=1
filters=21
activation=linear



[yolo]
mask = 3,4,5
anchors = 10,14,  23,27,  37,58,  81,82,  135,169,  344,319
classes=2
num=6
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

[route]
layers = -4

[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

[upsample]
stride=2

[route]
layers = -1, 8

[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

[convolutional]
size=1
stride=1
pad=1
filters=21
activation=linear

[yolo]
mask = 0,1,2
anchors = 10,14,  23,27,  37,58,  81,82,  135,169,  344,319
classes=2
num=6
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

然后我创建了 .names 文件 (./data/custom_cfg.names)：

cars
motorcylcles

创建了 train.txt、test.txt 和 obj.data (./custom/obj.data)：

classes= 2
train  = data/train.txt
valid  = data/test.txt
names = data/custom_cfg.names
backup = yolov3-tiny_2classes

最后，我通过执行命令使用darknet53.conv.74模型开始了训练

darknet detector train custom/obj.data custom/car_moto-yolov3-tiny.cfg darknet53.conv.74.

这是训练过程的示例输出：

5998: 0.619981, 0.483621 avg loss, 0.001000 rate, 0.499771 seconds, 143952 images, 0.011460 hours left
Loaded: 0.000028 seconds
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.828822), count: 3, class_loss = 0.617392, iou_loss = 0.131867, total_loss = 0.749259 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000005, iou_loss = 0.000000, total_loss = 0.000005 
 total_bbox = 172619, rewritten_bbox = 0.000000 % 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.762853), count: 3, class_loss = 1.025908, iou_loss = 0.332650, total_loss = 1.358558 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000002, iou_loss = 0.000000, total_loss = 0.000002 
 total_bbox = 172622, rewritten_bbox = 0.000000 % 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.841586), count: 3, class_loss = 0.431826, iou_loss = 0.156079, total_loss = 0.587906 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000002, iou_loss = 0.000000, total_loss = 0.000002 
 total_bbox = 172625, rewritten_bbox = 0.000000 % 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.743486), count: 3, class_loss = 0.947532, iou_loss = 0.385455, total_loss = 1.332986
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000001, iou_loss = 0.000000, total_loss = 0.000001
 total_bbox = 172628, rewritten_bbox = 0.000000 % 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.798283), count: 3, class_loss = 0.519976, iou_loss = 0.205342, total_loss = 0.725317 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000001, iou_loss = 0.000000, total_loss = 0.000001 
 total_bbox = 172631, rewritten_bbox = 0.000000 % 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.843649), count: 3, class_loss = 0.362844, iou_loss = 0.099908, total_loss = 0.462752 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000002, iou_loss = 0.000000, total_loss = 0.000002 
 total_bbox = 172634, rewritten_bbox = 0.000000 % 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.757608), count: 4, class_loss = 0.757355, iou_loss = 0.273935, total_loss = 1.031290 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000002, iou_loss = 0.000000, total_loss = 0.000002 
 total_bbox = 172638, rewritten_bbox = 0.000000 % 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 16 Avg (IOU: 0.775621), count: 3, class_loss = 0.883560, iou_loss = 0.192446, total_loss = 1.076006 
v3 (mse loss, Normalizer: (iou: 0.75, obj: 1.00, cls: 1.00) Region 23 Avg (IOU: 0.000000), count: 1, class_loss = 0.000000, iou_loss = 0.000000, total_loss = 0.000000 
 total_bbox = 172641, rewritten_bbox = 0.000000 % 

就是这样。我总是得到 Region 23 Avg (IOU: 0.000000)、class_loss = 0.000000、iou_loss = 0.000000、total_loss = 0.000000 或非常接近的值。然后，当我让网络运行时，它只分类了一类，即“汽车”。甚至摩托车也被归类为汽车。 有人可以帮忙吗？

以下是一些输入图像示例：

Answer 1

只需要调整.cfg文件中的一些参数就可以了。