dourflow is a keras/tensorflow implementation of the state-of-the-art object detection system You only look once.
Original paper and github: YOLO9000: Better, Faster, Stronger & Darknet.
- Clone repository:
git clone https://github.com/ksanjeevan/dourflow.git
python3 dourflow.py bird.jpg- Use on webcam (press 'q' to quit):
python3 dourflow.py camRunning python3 dourflow.py --help:
usage: dourflow.py [-h] [-m MODEL] [-c CONF] [-t THRESHOLD] [-w WEIGHT_FILE]
[--gif]
action
dourflow: a keras YOLO V2 implementation.
positional arguments:
action what to do: 'train', 'validate', 'cam' or pass a
video, image file/dir.
optional arguments:
-h, --help show this help message and exit
-m MODEL, --model MODEL
path to input yolo v2 keras model
-c CONF, --conf CONF path to configuration file
-t THRESHOLD, --threshold THRESHOLD
detection threshold
-w WEIGHT_FILE, --weight_file WEIGHT_FILE
path to weight file
--gif video output stored as gif also
Pass what to do with dourflow:
- A path to an image file/dir or video: Run inference on those file(s).
- 'cam': Run inference on webcam ('cams' to store the output).
- 'validate': Perform validation on a trained model.
- 'train': Perform training on your own dataset.
Pass the keras input model h5 file (could be to perform inference, validate against or for transfer learning).
Pretrained COCO/VOC keras models can be downloaded here. Alternatively, you can download the weights from here and generate the model file using YAD2K.
Pass a config.json file that looks like this (minus the comments!):
{
"model" : {
"input_size": 416, #Net input w,h size in pixels
"grid_size": 13, #Grid size
"true_box_buffer": 10, #Maximum number of objects detected per image
"iou_threshold": 0.5, #Intersection over union detection threshold
"nms_threshold": 0.3 #Non max suppression threhsold
},
"config_path" : {
"labels": "models/coco/labels_coco.txt", #Path to labels file
"anchors": "models/coco/anchors_coco.txt", #Path to anchors file
"arch_plotname": "" #Model output name (leave empty for none, see result_plots/yolo_arch.png for an example)
},
"train": {
"out_model_name": "", #Trained model name (saved during checkpoints)
"image_folder": "", #Training data, image directory
"annot_folder": "", #Training data, annotations directory (use VOC format)
"batch_size": 16, #Training batch size
"learning_rate": 1e-4, #Training learning rate
"num_epochs": 20, #Number of epochs to train for
"object_scale": 5.0 , #Loss function constant parameter
"no_object_scale": 1.0, #Loss function constant parameter
"coord_scale": 1.0, #Loss function constant parameter
"class_scale": 1.0, #Loss function constant parameter
"verbose": 1 #Training verbosity
},
"valid": {
"image_folder": "", #Validation data, image directory
"annot_folder": "", #Validation data, annotation directory
"pred_folder": "", #Validation data, predicted images directory (leave empty for no predicted image output)
}
}
Pass the confidence threshold used for detection (default is 30%).
Will generate a file in the same directory with an '_pred' name extension. Example:
python3 dourflow.py theoffice.png -m coco_model.h5 -c coco_config.json -t 0.35
Will create a directory named out/ in the current one and output all the images with the same name.
Example:
python3 dourflow.py images/ -m coco_model.h5 -c coco_config.json -t 0.35
Allows to evaluate the performance of a model by computing its mean Average Precision in the task of object detection (mAP WRITE UP COMING SOON).
Example:
python3 dourflow.py validate -m voc_model.h5 -c voc_config.jsonOutput:
Batch Processed: 100%|████████████████████████████████████████████| 4282/4282 [01:53<00:00, 37.84it/s]
AP( cat ): 0.908
AP( train ): 0.907
AP( dog ): 0.899
AP( bird ): 0.814
AP( aeroplane ): 0.810
AP( cow ): 0.810
AP( bus ): 0.806
AP( motorbike ): 0.792
AP( person ): 0.737
AP( sheep ): 0.719
AP( tvmonitor ): 0.718
AP( sofa ): 0.701
AP( bicycle ): 0.683
AP( diningtable ): 0.665
AP( car ): 0.641
AP( boat ): 0.617
AP( horse ): 0.575
AP( pottedplant ): 0.568
AP( chair ): 0.528
AP( bottle ): 0.487
-------------------------------
mAP: 0.719
Script to generate training/testing splits.
python3 split_dataset.py -p 0.75 --in_ann VOC2012/Annotations/ --in_img VOC2012/JPEGImages/ --output ~/Documents/DATA/VOC
Running:
python3 dourflow.py genp -c config.json
Will store the custom bounding box priors wherever the path indicates in the config file under config['config_path']['anchors'] with the prefix 'custom_' (so as to not overwrite accidentally).
Training will create directory logs/ which will store metrics and checkpoints for all the different training runs.
Model passed is used for transfer learning by randomizing the last layer of the network (with the appropiate size of the target classes).
Example:
python3 dourflow.py train -m models/logo/coco_model.h5 -c confs/config_custom.json
Then, in another terminal tab you can run tensorboard --logdir=logs/run_X and open a browser page at http://localhost:6006/ to monitor the loss, mAP, recall:
- Multiclass Non Max Suppression
- Anchor generation for custom datasets
- mAP write up
- Add webcam support
- Data Augmentation
- TensorBoard metrics