加载 ncnn 模型失败

Question

加载 ncnn 模型失败

popper0912 opened this issue 6 years ago · comments

popper0912 commented 6 years ago

使用 Darknet 训练模型后，转化为 ncnn 模型，最后使用 ncnn 加载 .param 文件，返回 -1 ，加载失败。不知道何原因？使用caffe 训练的模型，转化后，可以加载成功。

damone · Answer 1 · Mon Dec 17 2018 14:34:17 GMT+0800 (China Standard Time)

应该是转换参数有问题，找到出现错误的操作，检查一下转换后的参数吧，如果方便可以提供下网络模型文件

popper0912 · Answer 2 · Mon Dec 17 2018 15:22:09 GMT+0800 (China Standard Time)

应该是转换参数有问题，找到出现错误的操作，检查一下转换后的参数吧，如果方便可以提供下网络模型文件

我使用的是 yolov3-tiny 版的模型训练的自己的数据，darkNet cfg 文件如下：
[net]

Testing

batch=1
subdivisions=1

Training

#batch=64
#subdivisions=8
width=224
height=224
channels=1
momentum=0.9
decay=0.0005
angle=0
saturation = 1.0
exposure = 1.0
hue=0
flip=0

learning_rate=0.01
burn_in=1000
max_batches = 800000
policy=steps
steps=6000,20000,50000,100000
scales=.1,.1,.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=27
activation=linear

[yolo]
mask = 3,4,5
anchors = 21.2742,29.3429, 25.5435,43.5164, 28.0206,60.4466, 39.7142,54.9552, 33.7873,72.1623, 47.8516,66.9072, 45.8434,86.9286, 57.4758,106.0652, 110.5009,147.0255
classes=4
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=0

[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=27
activation=linear

[yolo]
mask = 0,1,2
anchors = 21.2742,29.3429, 25.5435,43.5164, 28.0206,60.4466, 39.7142,54.9552, 33.7873,72.1623, 47.8516,66.9072, 45.8434,86.9286, 57.4758,106.0652, 110.5009,147.0255
classes=4
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=0

===========================================================
转化后的ncnn 模型 param 参数如下:
7767517
50 52
Input data 0 1 data 0=224 1=224 2=3
Convolution conv_0 1 1 data conv_0 0=16 1=3 2=1 3=1 4=1 5=0 6=432
BatchNorm conv_0_batch_norm 1 1 conv_0 conv_0_batch_norm 0=16 1=0.00001
ReLU conv_0_activation 1 1 conv_0_batch_norm conv_0_activation 0=0.1
Pooling maxpool_1 1 1 conv_0_activation maxpool_1 0=0 1=2 2=2 3=0 5=1 13=0 14=1 15=1
Convolution conv_2 1 1 maxpool_1 conv_2 0=32 1=3 2=1 3=1 4=1 5=0 6=4608
BatchNorm conv_2_batch_norm 1 1 conv_2 conv_2_batch_norm 0=32 1=0.00001
ReLU conv_2_activation 1 1 conv_2_batch_norm conv_2_activation 0=0.1
Pooling maxpool_3 1 1 conv_2_activation maxpool_3 0=0 1=2 2=2 3=0 5=1 13=0 14=1 15=1
Convolution conv_4 1 1 maxpool_3 conv_4 0=64 1=3 2=1 3=1 4=1 5=0 6=18432
BatchNorm conv_4_batch_norm 1 1 conv_4 conv_4_batch_norm 0=64 1=0.00001
ReLU conv_4_activation 1 1 conv_4_batch_norm conv_4_activation 0=0.1
Pooling maxpool_5 1 1 conv_4_activation maxpool_5 0=0 1=2 2=2 3=0 5=1 13=0 14=1 15=1
Convolution conv_6 1 1 maxpool_5 conv_6 0=128 1=3 2=1 3=1 4=1 5=0 6=73728
BatchNorm conv_6_batch_norm 1 1 conv_6 conv_6_batch_norm 0=128 1=0.00001
ReLU conv_6_activation 1 1 conv_6_batch_norm conv_6_activation 0=0.1
Pooling maxpool_7 1 1 conv_6_activation maxpool_7 0=0 1=2 2=2 3=0 5=1 13=0 14=1 15=1
Convolution conv_8 1 1 maxpool_7 conv_8 0=256 1=3 2=1 3=1 4=1 5=0 6=294912
BatchNorm conv_8_batch_norm 1 1 conv_8 conv_8_batch_norm 0=256 1=0.00001
ReLU conv_8_activation 1 1 conv_8_batch_norm conv_8_activation 0=0.1
Split conv_8_activation_split 1 2 conv_8_activation conv_8_activation_split_0 conv_8_activation_split_1
Pooling maxpool_9 1 1 conv_8_activation_split_0 maxpool_9 0=0 1=2 2=2 3=0 5=1 13=0 14=1 15=1
Convolution conv_10 1 1 maxpool_9 conv_10 0=512 1=3 2=1 3=1 4=1 5=0 6=1179648
BatchNorm conv_10_batch_norm 1 1 conv_10 conv_10_batch_norm 0=512 1=0.00001
ReLU conv_10_activation 1 1 conv_10_batch_norm conv_10_activation 0=0.1
Pooling maxpool_11 1 1 conv_10_activation maxpool_11 0=0 1=2 2=1 3=0 5=1 13=0 14=1 15=1
Convolution conv_12 1 1 maxpool_11 conv_12 0=1024 1=3 2=1 3=1 4=1 5=0 6=4718592
BatchNorm conv_12_batch_norm 1 1 conv_12 conv_12_batch_norm 0=1024 1=0.00001
ReLU conv_12_activation 1 1 conv_12_batch_norm conv_12_activation 0=0.1
Convolution conv_13 1 1 conv_12_activation conv_13 0=256 1=1 2=1 3=1 4=0 5=0 6=262144
BatchNorm conv_13_batch_norm 1 1 conv_13 conv_13_batch_norm 0=256 1=0.00001
ReLU conv_13_activation 1 1 conv_13_batch_norm conv_13_activation 0=0.1
Split conv_13_activation_split 1 2 conv_13_activation conv_13_activation_split_0 conv_13_activation_split_1
Convolution conv_14 1 1 conv_13_activation_split_0 conv_14 0=512 1=3 2=1 3=1 4=1 5=0 6=1179648
BatchNorm conv_14_batch_norm 1 1 conv_14 conv_14_batch_norm 0=512 1=0.00001
ReLU conv_14_activation 1 1 conv_14_batch_norm conv_14_activation 0=0.1
Convolution conv_15 1 1 conv_14_activation conv_15 0=27 1=1 2=1 3=1 4=0 5=1 6=13824
DarknetActivation conv_15_activation 1 1 conv_15 conv_15_activation 0=3
Concat route_17 1 1 conv_13_activation_split_1 route_17 0=0
Convolution conv_18 1 1 route_17 conv_18 0=128 1=1 2=1 3=1 4=0 5=0 6=32768
BatchNorm conv_18_batch_norm 1 1 conv_18 conv_18_batch_norm 0=128 1=0.00001
ReLU conv_18_activation 1 1 conv_18_batch_norm conv_18_activation 0=0.1
Interp upsample_19 1 1 conv_18_activation upsample_19 0=1 1=2.f 2=2.f
Concat route_20 2 1 upsample_19 conv_8_activation_split_1 route_20 0=0
Convolution conv_21 1 1 route_20 conv_21 0=256 1=3 2=1 3=1 4=1 5=0 6=884736
BatchNorm conv_21_batch_norm 1 1 conv_21 conv_21_batch_norm 0=256 1=0.00001
ReLU conv_21_activation 1 1 conv_21_batch_norm conv_21_activation 0=0.1
Convolution conv_22 1 1 conv_21_activation conv_22 0=27 1=1 2=1 3=1 4=0 5=1 6=6912
DarknetActivation conv_22_activation 1 1 conv_22 conv_22_activation 0=3
Yolov3Detection yolo_23 2 1 conv_15_activation conv_22_activation yolo_23 0=4 1=6 2=0 3=0.25f 4=0.45f 5=224 6=224 -23307=12,39.714199,54.955200,33.787300,72.162300,47.851601,66.907204,21.274200,29.342899,25.543501,43.516399,28.020599,60.446602

我尝试注释掉 DarknetActivation 这一层加载便成功了不知道如何解决，请教一下

damone · Answer 3 · Mon Dec 17 2018 15:37:03 GMT+0800 (China Standard Time)

DarknetActivation 是自定义的网络层，你需要将其注册到ncnn network中，具体的你可以参考example/yolo.cpp

  CustomizedNet yolo;
  register_darknet_layer(yolo);
  yolo.load_param(param);

popper0912 · Answer 4 · Mon Dec 17 2018 16:04:15 GMT+0800 (China Standard Time)

DarknetActivation 是自定义的网络层，你需要将其注册到ncnn network中，具体的你可以参考example/yolo.cpp
  CustomizedNet yolo;
  register_darknet_layer(yolo);
  yolo.load_param(param);

谢谢