lukasc-ch / pytorch-yolo2

Convert https://pjreddie.com/darknet/yolo/ into pytorch

pytorch-yolo2

Convert https://pjreddie.com/darknet/yolo/ into pytorch. This repository is trying to achieve the following goals.

• implement RegionLoss, MaxPoolStride1, Reorg, GolbalAvgPool2d
• implement route layer
• detect, partial, valid functions
• load darknet cfg
• load darknet saved weights
• save as darknet weights
• fast evaluation
• pascal voc validation
• train pascal voc
• LMDB data set
• Data augmentation
• load/save caffe prototxt and weights
• reproduce darknet's training results
• convert weight/cfg between pytorch caffe and darknet
• add focal loss

---

Detection Using A Pre-Trained Model

wget http://pjreddie.com/media/files/yolo.weights
python detect.py cfg/yolo.cfg yolo.weights data/dog.jpg

You will see some output like this:

layer     filters    size              input                output
    0 conv     32  3 x 3 / 1   416 x 416 x   3   ->   416 x 416 x  32
    1 max          2 x 2 / 2   416 x 416 x  32   ->   208 x 208 x  32
    ......
   30 conv    425  1 x 1 / 1    13 x  13 x1024   ->    13 x  13 x 425
   31 detection
Loading weights from yolo.weights... Done!
data/dog.jpg: Predicted in 0.014079 seconds.
truck: 0.934711
bicycle: 0.998013
dog: 0.990524

---

Real-Time Detection on a Webcam

python demo.py cfg/tiny-yolo-voc.cfg tiny-yolo-voc.weights

---

Training YOLO on VOC

Get The Pascal VOC Data

wget https://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
wget https://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
wget https://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar
tar xf VOCtrainval_11-May-2012.tar
tar xf VOCtrainval_06-Nov-2007.tar
tar xf VOCtest_06-Nov-2007.tar

Generate Labels for VOC

wget http://pjreddie.com/media/files/voc_label.py
python voc_label.py
cat 2007_train.txt 2007_val.txt 2012_*.txt > voc_train.txt

Modify Cfg for Pascal Data

Change the cfg/voc.data config file

train  = train.txt
valid  = 2007_test.txt
names = data/voc.names
backup = backup

Download Pretrained Convolutional Weights

Download weights from the convolutional layers

wget http://pjreddie.com/media/files/darknet19_448.conv.23

or run the following command:

python partial.py cfg/darknet19_448.cfg darknet19_448.weights darknet19_448.conv.23 23

Train The Model

python train.py cfg/voc.data cfg/yolo-voc.cfg darknet19_448.conv.23

Evaluate The Model

python valid.py cfg/voc.data cfg/yolo-voc.cfg yolo-voc.weights
python scripts/voc_eval.py results/comp4_det_test_

mAP test on released models

yolo-voc.weights 544 0.7682 (paper: 78.6)
yolo-voc.weights 416 0.7513 (paper: 76.8)
tiny-yolo-voc.weights 416 0.5410 (paper: 57.1)

---

Focal Loss

A implementation of paper Focal Loss for Dense Object Detection

We get the results by using Focal Loss to replace CrossEntropyLoss in RegionLosss.

gama	training set	val set	mAP@416	mAP@544	Notes
0	VOC2007+2012	VOC2007	73.05	74.69	std-Cross Entropy Loss
1	VOC2007+2012	VOC2007	73.63	75.26	Focal Loss
2	VOC2007+2012	VOC2007	74.08	75.49	Focal Loss
3	VOC2007+2012	VOC2007	73.73	75.20	Focal Loss
4	VOC2007+2012	VOC2007	73.53	74.95	Focal Loss

---

Problems

1. Running variance difference between darknet and pytorch

Change the code in normalize_cpu to make the same result

normalize_cpu:
x[index] = (x[index] - mean[f])/(sqrt(variance[f] + .00001f));

Training on your own data

1. Padding your images into square size and produce the corresponding label files.
2. Modify the resize strageties in listDataset. Currently the resize scales range from 320 ~ 608, and the resize intervals is 64, which should be equal to batch_size or several times of batch_size.
3. Add warm up learning rate (scales=0.1,10,.1,.1)
4. Train your model as VOC does.

---

License

MIT License (see LICENSE file).

Contribution

Thanks for the contributions from @iceflame89 for the image augmentation and @huaijin-chen for focal loss.