search

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).
MIT License
1.73k stars 784 forks source link

Information about prediction and training on my dataset #254

Open martirius opened 6 years ago

martirius commented 6 years ago

Hi @experiencor , Thanks for your great work. I have some question about prediction and training. About training: i've trained on my own dataset(one class, images and bound boxes took from image-net), after 23 fo 50 epoch, the train process stopped saying this: Epoch 00023: val_loss did not improve Epoch 00023: early stopping /home/mattia/Tensorflow/keras-yolo2/utils.py:198: RuntimeWarning: overflow encountered in exp return 1. / (1. + np.exp(-x)) n01443537 0.0000 mAP: 0.0000 and is not encouraging. However the val_loss and loss were enough low, loss: 0.1420 - val_loss: 0.1616 so i tried the trained weights and surprisingly it works quite well, tried with other classes images and no false positive detected. I've also modified predict.py for reading frames from a camera and it works but in a strange way: if the camera is really close to the picture(5-6 cm) it detects the object, but if i outdistance the camera from the picture(10 cm and more) it doesn't detect anything. Is this the normal behavior of your implementation of YoloV2? Because i didn't found this problem with darknet

experiencor commented 6 years ago

There a limitation with the custom sigmoid in utitls. Try to use https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.special.expit.html instead.

It's not working well if you train it on big objects only. Try to include small objects in the dataset.