/mxnet-yolov2

YOLO: You only look once real-time object detector

Primary LanguagePythonMIT LicenseMIT

YOLO-v2: Real-Time Object Detection

Todo list

  • Training different backbone from scratch
  • Get new network weights by training on imagenet directly or transfer learning
  • Loss function
  • Adding --num_layers in train.py for choosing different network

History of mAP (on VOC2007)

Resnet-50:

Exp1

  • lr=0.001, steps=(90, 180), data-shape=416. However, adding the first 80 epoch, in fact the steps = (90 + 80, 180 + 80)
  • first: just runing in 80 epoch then stop because of batch-size=30!
python train-416.py --gpus 0,1 --network resnet50_yolo --data-shape 416 --pretrained model/resnet-50 \ 
                    --epoch 0 --log train_416.log --min-random-shape 320 --batch-size 30
  • second: continuing to train based on the 80-th weight in the first, meanwhile tuning batch-size to 28!
python train-416.py --gpus 0,1 --network resnet50_yolo --data-shape 416 --pretrained model/yolo2_resnet50_416 \
                    --epoch 80 --log train_416_80.log --min-random-shape 320 --batch-size 28
  • 2018/06/15: Epoch[236] Validation-mAP=0.759613
  • 2018/06/14: Epoch[174] Validation-mAP=0.759229
  • 2018/06/14: Epoch[126] Validation-mAP=0.757837
  • 2018/06/14: Epoch[105] Validation-mAP=0.757725
  • 2018/06/14: Epoch[104] Validation-mAP=0.757814
  • 74 mAP by original repo (https://github.com/zhreshold/mxnet-yolo)
  • 76.8 mAP at (416 x 416) by original paper.

Exp2

  • tuning random-shape-epoch from 10 to 1, other almost same with original repo.
  • Command:
python train-416.py --network resnet50_yolo --batch-size 28 --pretrained model/resnet-50 \
                    --epoch 0 --gpus 0,1 --begin-epoch 0 --end-epoch 240 --data-shape 416 \
                    --random-shape-epoch 1 --min-random-shape 320 --max-random-shape 608 \
                    --lr 0.001 --lr-steps 90,180 --lr-factor 0.1 --log train-exp2.log \
                    --num-class 20 --num-example 16551 --nms 0.45 --overlap 0.5
  • 2018/06/15: Epoch[236] Validation-mAP=0.707528

Exp3

  • lr-steps 180,360
  • Command:
python train-416.py --network resnet50_yolo --batch-size 28 --pretrained model/resnet-50 \
                    --epoch 0 --gpus 0,1 --begin-epoch 0 --end-epoch 540 --data-shape 416 \
                    --random-shape-epoch 10 --min-random-shape 320 --max-random-shape 608 \
                    --lr 0.001 --lr-steps 180,360 --lr-factor 0.1 --log train-exp3.log \
                    --num-class 20 --num-example 16551 --nms 0.45 --overlap 0.5
  • 2018/06/18: Epoch[256] Validation-mAP=0.758914

Exp4

  • lr-factor=0.5，lr-steps=[90，180，270，360，450]，lr=[0.001，0.0005，0.00025，0.000125，0.0000625，0.00003125]
  • Command:
python train-416.py --network resnet50_yolo --batch-size 28 --pretrained model/resnet-50 \
                    --epoch 0 --gpus 0,1 --begin-epoch 0 --end-epoch 540 --data-shape 416 \
                    --random-shape-epoch 10 --min-random-shape 320 --max-random-shape 608 \
                    --lr 0.001 --lr-steps 90,180,270,360,450 --lr-factor 0.5 --log train-exp4.log \
                    --num-class 20 --num-example 16551 --nms 0.45 --overlap 0.5

python train-416.py --network resnet50_yolo --batch-size 28 --pretrained model/resnet-50 \
                    --epoch 0 --gpus 0,1 --begin-epoch 0 --end-epoch 540 --data-shape 416 \
                    --random-shape-epoch 10 --min-random-shape 320 --max-random-shape 608 \
                    --lr 0.001 --lr-steps 90,180,270,360,450 --lr-factor 0.5 --log train-exp4.log \
                    --num-class 20 --num-example 16551 --nms 0.45 --overlap 0.5 --resume 288
  • 2018/06/21: Epoch[477] Validation-mAP=0.759190
  • 2018/06/20: Epoch[408] Validation-mAP=0.759149
  • 2018/06/19: Epoch[276] Validation-mAP=0.757242
  • 2018/06/19: Epoch[233] Validation-mAP=0.755262

Exp5

  • lr-factor=0.1，lr-steps=[90，180，270，360，450]，lr=[0.001，0.0001，0.00001，0.000001，0.0000001，0.00000001]
  • Command:
python train-416.py --network resnet50_yolo --batch-size 28 --pretrained model/resnet-50 \
                    --epoch 0 --gpus 0,1 --begin-epoch 0 --end-epoch 540 --data-shape 416 \
                    --random-shape-epoch 10 --min-random-shape 320 --max-random-shape 608 \
                    --lr 0.001 --lr-steps 90,180,270,360,450 --lr-factor 0.1 --log train-exp5.log
  • 2018/06/22:

Exp6

  • lr-factor=0.1，lr-steps=[90，180]，lr=[0.001，0.0001，0.00001]
  • Command:
python train-416-resnet152.py --network resnet152_yolo --batch-size 28 --pretrained scratch\
                    --epoch 0 --gpus 0,1 --begin-epoch 0 --end-epoch 540 --data-shape 416 \
                    --random-shape-epoch 10 --min-random-shape 320 --max-random-shape 608 \
                    --lr 0.001 --lr-steps 90,180 --lr-factor 0.1 --log train-exp6.log
  • 2018/06/22:

Exp7

  • data-shape != 416
  • The results are not good becouse the pretrained models are trained for data-shape 416?

Darkent19:

Difference between official Mxnet and the one used in the repo

added the following operators:

  • yolo_output.* to src/operator/contrib

zhreshold#7

Changing network

  • create file: mxnet-yolov2/symbol/symbol_resnet152_yolo.py
  • get weights file for resnet152
  • change num_layers (from 50 to 152) in symbol_resnet152_yolo.py:
def get_symbol(num_classes=20, nms_thresh=0.5, force_nms=False, **kwargs):
    #body = resnet.get_symbol(num_classes, 50, '3,224,224')
    body = resnet.get_symbol(num_classes, 152, '3,224,224')
  • using resnet152_yolo as --network in train.py

Disclaimer

Re-implementation of original yolo-v2 which is based on darknet.

The arXiv paper is available here.

Demo

demo1

Getting started

  • Build from source, this is required because this example is not merged, some custom operators are not presented in official MXNet. Instructions
  • Install required packages: cv2, matplotlib

Try the demo

  • Download the pretrained model(darknet as backbone), or this model(resnet50 as backbone) and extract to model/ directory.
  • Run
# cd /path/to/mxnet-yolo
python demo.py --cpu
# available options
python demo.py -h

Train the model

cd /path/to/where_you_store_datasets/
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
# Extract the data.
tar -xvf VOCtrainval_11-May-2012.tar
tar -xvf VOCtrainval_06-Nov-2007.tar
tar -xvf VOCtest_06-Nov-2007.tar
ln -s /path/to/VOCdevkit /path/to/mxnet-yolo/data/VOCdevkit
  • Create packed binary file for faster training
# cd /path/to/mxnet-ssd
bash tools/prepare_pascal.sh
# or if you are using windows
python tools/prepare_dataset.py --dataset pascal --year 2007,2012 --set trainval --target ./data/train.lst
python tools/prepare_dataset.py --dataset pascal --year 2007 --set test --target ./data/val.lst --shuffle False
  • Start training
python train.py --gpus 0,1,2,3 --epoch 0
# choose different networks, such as resnet50_yolo
python train.py --gpus 0,1,2,3 --network resnet50_yolo --data-shape 416 --pretrained model/resnet-50 --epoch 0
# choose different networks, such as resnet152_yolo
python train.py --gpus 0,1,2,3 --network resnet152_yolo --data-shape 416 --pretrained model/resnet-50 --epoch 0
# see advanced arguments for training
python train.py -h