- Colab
- Pytorch >=1.5
- Clone into repository
https://github.com/ultralytics/yolov3 - install
pip install -U -r requirements.txt
Created train, test and valid folder with correspondence to the training,test and valid.txt files.
#separate images based on train,val and test and put corresponding files into
#respective folder having images and labels as subfolders in it
def copy(filepath,dest_dir):
with open(filepath) as fp:
for line in fp:
os.makedirs(dest_dir, exist_ok=True) # succeeds even if directory exists.
shutil.copy(line.replace('\n', ''), dest_dir+"images") #image
shutil.copy(line.replace('.jpg\n', '.txt'), dest_dir+"labels") ##label
print("Done: "+ filepath)
copy('./training.txt','yolo_train/')
copy('./validation.txt','yolo_val/')
copy('./test.txt','yolo_test/')
The images and annotations/labels should look like this:
Dataset/yolo_test/images/0000.jpg #image
Dataset/yolo_test/label/0000.txt #label
Note: Each image's label should be locatable by replacing /image/*.jpg with /label/*.txt
then created .txt for annotations with following code:
def copy(filepath,newfile,binder):
with open(filepath) as fp:
print(filepath)
for line in fp:
in_file = line[2:]
new = binder+in_file
with open(newfile, "a") as f:
f.write(new)
copy('training.txt','Dataset/yolo_train/yolo_train.txt',
"/content/yolov3/Dataset/yolo_train/")
copy('test.txt','Dataset/yolo_test/yolo_test.txt',
"/content/yolov3/Dataset/yolo_test/")
copy('validation.txt','Dataset/yolo_val/yolo_val.txt',
"/content/yolov3/Dataset/yolo_val/")
Create classes.names file, Here we are using Dataset/sims_classes.names
car
bus
truck
van
longvehicle
...
Create sims.txt for passing data for training. It should look like this:
classes=15
train=/content/yolov3/Dataset/yolo_train/yolo_train.txt
valid=/content/yolov3/Dataset/yolo_val/yolo_val.txt
names=/content/yolov3/Dataset/sims_classes.names
For testing, create sims_test.txt and it should look like this:
classes=15
valid=/content/yolov3/Dataset/yolo_test/yolo_test.txt
names=/content/yolov3/Dataset/sims_classes.names
Update *.cfg in all yolov3 layers. Filters in these layers should be filters=[5 + n] * 3 where n = number of classes
Example dataset provided in Dataset folder
Summary availble here in txt file link
Summary availble here in txt file link
Form more details please refer to Yolov3_baseline.ipynb or 'Yolov3_SPP_improved.ipynbavailble in current directory. <br /> For training on custom dataset, earlier createddata.txt` will be passed with following command:
Pre-Trained weights can be download for transfer learning by:
!. weights/download_yolov3_weights.sh
#start training
!python3 train.py --batch 8 --epochs 60 --img-size 512 --data Dataset/sims.txt --cache-images --rec --cfg yolov3.cfg --name from_yolov3 --weights weights/yolov3.pt
Run command:
python3 test.py --data Dataset/sims_test.txt --cfg yolov3.cfg --batch-size 8 --weights weights/last_from_yolov3.pt --save-json --img-size 512
For detections run:
python3 detect.py --names Dataset/sims_classes.names --cfg yolov3.cfg --weights weights/best_from_yolov3.pt
| Model | Validation mAP | Test mAP |
|---|---|---|
| Yolov3 | 0.608 | 0.634 |
| Yolov3-SPP | 0.653 | 0.679 |
Image Size = 512
Batch_size = 8
iterations_per_epoch = 1000
LR = 0.01 , final = 0.0005
LR Scheduler = cosine scheduler
config file = cfg/Yolov3.cfg
Image Size = 512
Batch_size = 8
iterations_per_epoch = 1000
LR = 0.01 , final = 0.0005
LR Scheduler = cosine scheduler
config file = cfg/Yolov3-spp.cfg
Class Images Targets P R mAP@0.5 F1: 100% 94/94 [00:25<00:00, 3.76it/s]
all 748 7.98e+03 0.549 0.697 0.634 0.581
car 748 3.68e+03 0.768 0.836 0.834 0.801
truck 748 446 0.497 0.78 0.674 0.607
van 748 874 0.499 0.747 0.626 0.598
longvehicle 748 222 0.395 0.856 0.6 0.541
bus 748 366 0.733 0.847 0.823 0.786
airliner 748 161 0.696 0.981 0.962 0.815
propeller 748 25 0.254 0.92 0.689 0.398
trainer 748 49 0.533 0.863 0.782 0.659
chartered 748 103 0.677 0.854 0.838 0.755
fighter 748 8 0.53 0.625 0.657 0.574
other 748 85 0.0956 0.259 0.0944 0.14
stairtruck 748 83 0.395 0.47 0.328 0.429
pushbacktruck 748 55 0.255 0.145 0.13 0.185
helicopter 748 9 1 0.352 0.54 0.52
boat 748 1.82e+03 0.898 0.918 0.936 0.908
Speed: 19.6/2.7/22.3 ms inference/NMS/total per 512x512 image at batch-size 8
Class Images Targets P R mAP@0.5 F1: 100% 94/94 [00:17<00:00, 5.24it/s]
all 748 7.98e+03 0.586 0.712 0.679 0.636
car 748 3.68e+03 0.785 0.856 0.859 0.819
truck 748 446 0.57 0.767 0.727 0.654
van 748 874 0.643 0.778 0.743 0.704
longvehicle 748 222 0.488 0.833 0.632 0.615
bus 748 366 0.757 0.883 0.864 0.815
airliner 748 161 0.904 0.969 0.968 0.935
propeller 748 25 0.658 0.88 0.903 0.753
trainer 748 49 0.687 0.98 0.941 0.807
chartered 748 103 0.635 0.963 0.927 0.765
fighter 748 8 0.75 0.75 0.821 0.75
other 748 85 0.0909 0.235 0.0898 0.131
stairtruck 748 83 0.32 0.283 0.22 0.3
pushbacktruck 748 55 0.17 0.109 0.115 0.133
helicopter 748 9 0.437 0.432 0.421 0.434
boat 748 1.82e+03 0.9 0.955 0.952 0.927
Speed: 9.1/2.6/11.8 ms inference/NMS/total per 512x512 image at batch-size 8
