search

qqwweee / keras-yolo3

A Keras implementation of YOLOv3 (Tensorflow backend)
MIT License
7.14k stars 3.45k forks source link

Retrain w/ own Dataset: Choice of Parameters #587

Open syikmis opened 4 years ago

syikmis commented 4 years ago

Yeehawww guys and girls, has someone successfully retrained the model w/ another dataset then COCO or similar? I am retraining the model w/ Crowdhuman dataset for 50 epochs with frozen layers optimizer=Adam(lr=3e-3) and then another 50 epochs for all layers optimizer=Adam(lr=1e-3) - just like it is in train.py. Training takes guite time (15k images), so i am not able to play around with various parameters.

After 60 epochs of training i have a val_loss = 95.39, which is obviously not very good..

Has someone suggestions for an improved training procedure or better parametervalues?

I already use the anchors produced by the k-means algorithm and got an avg IoU of 69,45%.

Thx very much!

tabmoo commented 4 years ago

Could you please share the annotations in the format of this repository so I could try Crowdhuman myself and compare the loss with yours?