search

tjuskyzhang / Scaled-YOLOv4-TensorRT

Got 100fps on TX2. Got 500fps on GeForce GTX 1660 Ti. If the project is useful to you, please Star it.
178 stars 41 forks source link

Scaled YOLOv4-tiny 3l #16

Closed casper-hansen closed 3 years ago

casper-hansen commented 3 years ago

@tjuskyzhang hi,

Can you guide with making your repository work with the official tiny 3 yolo layer version?

Here is my config:

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

learning_rate=0.00261
burn_in=1000
max_batches = 6000
policy=steps
steps=4800,5400
scales=.1,.1

[convolutional]
batch_normalize=1
filters=32
size=3
stride=2
pad=1
activation=mish

[convolutional]
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=mish

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

[route_lhalf]
layers=-1

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

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

[route]
layers = -1,-2

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

[route]
layers = -6,-1

[maxpool]
size=2
stride=2

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

[route_lhalf]
layers=-1

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

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

[route]
layers = -1,-2

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

[route]
layers = -6,-1

[maxpool]
size=2
stride=2

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

[route_lhalf]
layers=-1

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

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

[route]
layers = -1,-2

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

[route]
layers = -6,-1

[maxpool]
size=2
stride=2

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

##################################

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

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

[convolutional]
size=1
stride=1
pad=1
filters=24
activation=linear

[yolo]
mask = 6,7,8
anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
classes=3
num=9
jitter=.3
scale_x_y = 1.05
cls_normalizer=1.0
iou_normalizer=0.07
iou_loss=ciou
ignore_thresh = .7
truth_thresh = 1
random=0
#resize=1.5
nms_kind=greedynms
beta_nms=0.6

[route]
layers = -4

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

[upsample]
stride=2

[route]
layers = -1, 23

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

[convolutional]
size=1
stride=1
pad=1
filters=24
activation=linear

[yolo]
mask = 3,4,5
anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
classes=3
num=9
jitter=.3
scale_x_y = 1.05
cls_normalizer=1.0
iou_normalizer=0.07
iou_loss=ciou
ignore_thresh = .7
truth_thresh = 1
random=0
#resize=1.5
nms_kind=greedynms
beta_nms=0.6

[route]
layers = -3

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

[upsample]
stride=2

[route]
layers = -1, 15

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

[convolutional]
size=1
stride=1
pad=1
filters=24
activation=linear

[yolo]
mask = 0,1,2
anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
classes=3
num=9
jitter=.3
scale_x_y = 1.05
cls_normalizer=1.0
iou_normalizer=0.07
iou_loss=ciou
ignore_thresh = .7
truth_thresh = 1
random=0
#resize=1.5
nms_kind=greedynms
beta_nms=0.6
casper-hansen commented 3 years ago

Never mind, your repository is so good that I could modify the CSP version to work with my config and weights! Even though I have never touched C++. Thank you for this awesome repo!

philipp-schmidt commented 3 years ago

@casperbh96 have you removed the group feature on purpose?

[route]
layers=-1
groups=2
group_id=1

I can not find a way to convert the weights for tiny or tiny-3l because of those parameters. They seem to split the feature map in half. You simply removed it and retrained the network, correct? Or did you start without this group feature in the first place?

philipp-schmidt commented 3 years ago

Ah you replaced them with [route_lhalf], I get it. Thanks.

casper-hansen commented 3 years ago

@casperbh96 have you removed the group feature on purpose?

[route]
layers=-1
groups=2
group_id=1

I can not find a way to convert the weights for tiny or tiny-3l because of those parameters. They seem to split the feature map in half. You simply removed it and retrained the network, correct? Or did you start without this group feature in the first place?

This was because I used the original Scaled YOLOv4 implementation in PyTorch where the author used route_lhalf. I might replace them at some point and completely covert to Darknet.

philipp-schmidt commented 3 years ago

Actually you can't, conversion fails with an error because the python darknet implementation doesn't know about the groupid field. But it seems the implementations are compatible.