fsx950223 / mobilenetv2-yolov3

yolov3 with mobilenetv2 and efficientnet

Mobilenetv2-Yolov3

Tensorflow implementation mobilenetv2-yolov3 and efficientnet-yolov3 inspired by keras-yolo3

---

Update

Backend:

• MobilenetV2
• Efficientnet
• Darknet53

Callback:

• mAP
• Tensorboard extern callback

Loss:

• MSE
• GIOU
• Adversarial loss

Train:

• Cosine learning rate
• Auto augment

Tensorflow:

• Tensorflow2 Ready
• tf.data pipeline
• Convert model to tensorflow lite model
• Multi GPU training
• TPU support
• TensorRT support

Serving:

• Tensorflow Serving warm up request
• Tensorflow Serving JAVA Client
• Tensorflow Serving Python Client
• Tensorflow Serving Service Control Client
• Tensorflow Serving Server Build and Plugins develop

---

Usage

Install:

pip install -r requirements.txt

Get help info:

python main.py --help

Train:

1. Format file name like [name]_[number].[extension]
   Example:
   

voc_train_3998.txt

2. If you are using txt dataset, please format records like [image_path] [,[xmin ymin xmax ymax class]]
(for convenience, you can modify voc_text.py to parse your data to specific data format), else you should modify voc_annotation.py, then run

python voc_annotation.py

to parse your data to tfrecords.
Example:

/image/path 179 66 272 290 14 172 38 317 349 14 276 2 426 252 14 1 32 498 365 13

3. Run:

python main.py --mode=TRAIN --train_dataset_glob=<your dataset glob> --epochs=50 --epochs=50 --mode=TRAIN

Predict:

python main.py --mode=IMAGE --model=<your_model_path>

MAP:

python main.py --mode=MAP --model=<your_model_path> --test_dataset_glob=<your dataset glob>

Export serving model:

python main.py --mode=SERVING --model=<your_model_path>

Use custom config file:

python main.py --config=mobilenetv2.yaml

---

Set up tensorflow.js model (Live Demo: https://fsx950223.github.io/mobilenetv2-yolov3/tfjs/)

1. Create a web server on project folder
   
2. Open browser and enter [your_url:your_port]/tfjs
   

---

Resources

• Download pascal tfrecords from here.
• Download pre-trained mobilenetv2-yolov3 model(VOC2007) here
• Download pre-trained efficientnet-yolov3 model(VOC2007) here
• Download pre-trained efficientnet-yolov3 model(VOC2007+2012) here

---

Performance

Network: Mobilenetv2+Yolov3
Input size: 416*416
Train Dataset: VOC2007
Test Dataset: VOC2007
mAP:

aeroplane ap:  0.6721874861775297
bicycle ap:  0.7844226664948993
bird ap:  0.6863393529648882
boat ap:  0.5102715372530052
bottle ap:  0.4098093697072679
bus ap:  0.7646277543282962
car ap:  0.8000339732789448
cat ap:  0.8681120849855787
chair ap:  0.4021823009684314
cow ap:  0.6768311030872428
diningtable ap:  0.626045232887253
dog ap:  0.8293983813984888
horse ap:  0.8315961581768014
motorbike ap:  0.771283337747543
person ap:  0.7298645793931624
pottedplant ap:  0.3081565644702266
sheep ap:  0.6510012751038824
sofa ap:  0.6442699680945367
train ap:  0.8025086962000969
tvmonitor ap:  0.6239227675451299
mAP:  0.6696432295131602

GPU inference time (GTX1080Ti): 19ms
CPU inference time (i7-8550U): 112ms
Model size: 37M

Network: Efficientnet+Yolov3
Input size: 380*380
Train Dataset: VOC2007
Test Dataset: VOC2007
mAP:

aeroplane ap:  0.7770436248733187
bicycle ap:  0.822183784348553
bird ap:  0.7346967323068865
boat ap:  0.6142903989882571
bottle ap:  0.4518063126765959
bus ap:  0.782237197681936
car ap:  0.8138978890046222
cat ap:  0.8800232369515162
chair ap:  0.4531520519719176
cow ap: 0.6992367978932157
diningtable ap:  0.6765065569475968
dog ap:  0.8612118810883834
horse ap:  0.8559580684256001
motorbike ap:  0.8027311717682002
person ap:  0.7280218883512792
pottedplant ap:  0.35520418960051925
sheep ap:  0.6833401035128458
sofa ap:  0.6753841073186044
train ap:  0.8107647793504738
tvmonitor ap:  0.6726791558585905
mAP:  0.7075184964459456

GPU inference time (GTX1080Ti): 23ms
CPU inference time (i7-8550U): 168ms
Model size: 77M

Network: Efficientnet+Yolov3
Input size: 380*380
Train Dataset: VOC2007+VOC2012
Test Dataset: VOC2007
mAP:

aeroplane ap:  0.8572154850266848
bicycle ap:  0.8129962658687486
bird ap:  0.8325678324285539
boat ap:  0.7061501348114156
bottle ap:  0.5603823420846883
bus ap:  0.8536452418769342
car ap:  0.8395446870008888
cat ap:  0.9200504816535645
chair ap:  0.514644868267842
cow ap:  0.8202171886452714
diningtable ap:  0.7370149790284737
dog ap:  0.900374518831019
horse ap:  0.8632567146990895
motorbike ap:  0.8147344820261591
person ap:  0.7690434789031615
pottedplant ap:  0.4576271726152926
sheep ap:  0.8006580581981677
sofa ap:  0.7478146395952494
train ap:  0.8783508559769437
tvmonitor ap:  0.6923886096918628
mAP:  0.7689339018615006

GPU inference time (GTX1080Ti): 23ms
CPU inference time (i7-8550U): 168ms
Model size: 77M

---

Reference

paper:

• YOLOv3: An Incremental Improvement
  
• An Analysis of Scale Invariance in Object Detection - SNIP
  
• Scale-Aware Trident Networks for Object Detection
  
• Understanding the Effective Receptive Field in Deep Convolutional Neural Networks
  
• Bag of Freebies for Training Object Detection Neural Networks
  
• Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression
  
• MobileNetV2: Inverted Residuals and Linear Bottlenecks