Darknet: need instruction on writing a self-defined network using darknet

[shartoo]

Hi,there.I'm trying to implement a MOT network using darknet but got some error without any clue ,could anybody tell me what's going wrong?

network structure

The network structure of Simple Online and Realtime Tracking with a Deep Association Metric is below

my cfg file

[net]
batch=64
subdivisions=64
height=256
width=128
channels=3
hue=.1
saturation=.75
exposure=.75

policy=poly
power=4
max_batches = 100000
momentum=0.9
decay=0.0005

learning_rate=0.001
burn_in=10000

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

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

[maxpool]
size=3
stride=2


# Residual Block stride=1,filters=32
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=elu

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

[shortcut]
activation=elu
from=-3

# Residual Block stride=1,filters=32
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=elu

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

[shortcut]
activation=elu
from=-3

# Residual Block increase dim,stride=2,filters=64
[convolutional]
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=elu

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

[shortcut]
activation=elu
from=-3

# Residual Block increase dim,stride=1  ,filters=64
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=elu

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

[shortcut]
activation=elu
from=-3

# Residual Block increase dim,stride=2 ,filters=128
[convolutional]
batch_normalize=1
filters=128
size=3
stride=2
pad=1
activation=elu

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

[shortcut]
activation=elu
from=-3

# Residual Block increase dim,stride=1 ,filters=128
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=elu

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

[shortcut]
activation=elu
from=-3

[dropout]
probability=.4

# full connected layer
[connected] 
output=128
activation=elu

# L2Norm layer
[l2norm]
[connected]
output=1501
activation=elu
# softmax layer
[softmax] 
groups=1

Data

The MARS dataset consist of 1501 class with about one milion of image.

I've prepared my data file according to cifar10 example as below

trainval.txt

/home/ss/data/MOTData/train_in_one/0105C6T0044F121_0105.jpg
/home/ss/data/MOTData/train_in_one/0717C3T0008F005_0717.jpg
/home/ss/data/MOTData/train_in_one/0811C3T0021F075_0811.jpg
/home/ss/data/MOTData/train_in_one/0603C3T0001F031_0603.jpg
...

mot.names

0003
0005
0007
0009
0011
0013
0015
0017
0019
0021
...

labels

Exactly the same as mot.names

mot.data

classes=1501
train=data/mot/mot_train_new.txt
valid=data/mot/mot_test_new.txt
labels=data/mot/mot_labels.names
names =data/mot/mot_labels.names
backup=backupMOT/

Run

I've ran training command as below

./darknet classifier train data/mot/mot.data cfg/mot.cfg

output error

I got error like

1
layer     filters    size              input                output
    0 conv     32  3 x 3 / 1   128 x 256 x   3   ->   128 x 256 x  32  0.057 BFLOPs
    1 conv     32  3 x 3 / 1   128 x 256 x  32   ->   128 x 256 x  32  0.604 BFLOPs
    2 max          3 x 3 / 2   128 x 256 x  32   ->    64 x 128 x  32
    3 conv     32  3 x 3 / 1    64 x 128 x  32   ->    64 x 128 x  32  0.151 BFLOPs
    4 conv     32  3 x 3 / 1    64 x 128 x  32   ->    64 x 128 x  32  0.151 BFLOPs
    5 res    2                  64 x 128 x  32   ->    64 x 128 x  32
    6 conv     32  3 x 3 / 1    64 x 128 x  32   ->    64 x 128 x  32  0.151 BFLOPs
    7 conv     32  3 x 3 / 1    64 x 128 x  32   ->    64 x 128 x  32  0.151 BFLOPs
    8 res    5                  64 x 128 x  32   ->    64 x 128 x  32
    9 conv     64  3 x 3 / 2    64 x 128 x  32   ->    32 x  64 x  64  0.075 BFLOPs
   10 conv     64  3 x 3 / 1    32 x  64 x  64   ->    32 x  64 x  64  0.151 BFLOPs
   11 res    8                  64 x 128 x  32   ->    32 x  64 x  64
   12 conv     64  3 x 3 / 1    32 x  64 x  64   ->    32 x  64 x  64  0.151 BFLOPs
   13 conv     64  3 x 3 / 1    32 x  64 x  64   ->    32 x  64 x  64  0.151 BFLOPs
   14 res   11                  32 x  64 x  64   ->    32 x  64 x  64
   15 conv    128  3 x 3 / 2    32 x  64 x  64   ->    16 x  32 x 128  0.075 BFLOPs
   16 conv    128  3 x 3 / 1    16 x  32 x 128   ->    16 x  32 x 128  0.151 BFLOPs
   17 res   14                  32 x  64 x  64   ->    16 x  32 x 128
   18 conv    128  3 x 3 / 1    16 x  32 x 128   ->    16 x  32 x 128  0.151 BFLOPs
   19 conv    128  3 x 3 / 1    16 x  32 x 128   ->    16 x  32 x 128  0.151 BFLOPs
   20 res   17                  16 x  32 x 128   ->    16 x  32 x 128
   21 dropout       p = 0.40               65536  ->  65536
   22 connected                            65536  ->   128
   23 l2norm                                          128
   24 connected                             128  ->  1501
   25 softmax                                        1501
Learning Rate: 0.001, Momentum: 0.9, Decay: 0.0005
509914
128 256
Segmentation fault (core dumped)

It was ran on a Nvidia 1080ti machine,i think batch=64 and subdivisions=64 is far enough to train such a small network. Any idea? Thank you in advance!

[AlexeyAB]

@shartoo Hi,

• What parameters do you use in the Makefile?

• Do you use this repo https://github.com/pjreddie/darknet or this https://github.com/AlexeyAB/darknet ?

• What versions of CUDA and cuDNN do you use?

[shartoo]

GPU=1
CUDNN=1
opencv=1

I'm using pjreddie's darknet and my cuda version is 8.0. I can run cifar10 model ,it works .

[AlexeyAB]

@shartoo

May be there is some bug.

Just try to localize it:

• try to use square network size 256x256
• try to find empty last line in the mot_train_new.txt
• try to cooment l2norm and dropout layers

[shartoo]

Thank you for your advice, all of the image are 128x256 thus i can't change this. I've tried to commented l2norm it make no sense. I'll find if there were sth wrong in mot_train.txt

[shartoo]

I‘ve found the problem now, there are only 625 classes in training data though the number ranges from 1 to 1501 but not continuous.I have to create new label file.Thank you！