How to do transfer learning for yolo object detection algorithm? - neural-network

Question

I have successfully trained yolo for predicting my own images with this article. In there I changed classes = 5 (I trained for 5 classes) and filters to 50 in line 224 in my cfg file.
What I want is I want to do transfer learning for yolo with training the last fully connected layer and the softmax layer.
My cfg file is as follows.
[net]
batch=64
subdivisions=8
height=416
width=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
learning_rate=0.0001
max_batches = 45000
policy=steps
steps=100,25000,35000
scales=10,.1,.1
[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
[convolutional]
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky
[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
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[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
[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]
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
[maxpool]
size=2
stride=2
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
#######
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[route]
layers=-9
[reorg]
stride=2
[route]
layers=-1,-3
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=50
activation=linear
[region]
anchors = 1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52
bias_match=1
classes=5
coords=4
num=5
softmax=1
jitter=.2
rescore=1
object_scale=5
noobject_scale=1
class_scale=1
coord_scale=1
absolute=1
thresh = .6
random=0

Answer 1

Actually it is filters=(classes + 5)*5
Reference: Here

Answer 2

With 5 classes, you need to set filters to 30 not 50.
filters = (number of classes+1 )* 5

Answer 3

I am guessing you are using the pjreddie/darknet framework for the YOLO implementation.
If that's the case, then you set an additional parameter stopbackward=1 at the layer above which you don't need the update.
In parse.c file, at line 724:
l.stopbackward = option_find_int_quiet(options, "stopbackward", 0);
So this means its a parameter at every layer, just like batch_normalize=1 you can specify stopbackward=1. So any layers above this will not be updated. This is also seen in file network.c at Line 272:
if(l.stopbackward) break;