Keras-yolo3: problem in train my own dateset

[douxiao]

Hello, the code you wrote is very good. When I used to train my own dataset, the following problems occurred. My dataset was only one class and the number of pictures was 1000, and the training pre-training weight was yolo_weights.h5. The final loss value always stops at about 14, not falling. Do you have any good suggestions for me? Many thanks!

1/50 [..............................] - ETA: 26s - loss: 14.6514
2/50 [>.............................] - ETA: 25s - loss: 14.7824
3/50 [>.............................] - ETA: 25s - loss: 14.8408
4/50 [=>............................] - ETA: 25s - loss: 14.7723
5/50 [==>...........................] - ETA: 24s - loss: 14.8877
6/50 [==>...........................] - ETA: 24s - loss: 14.7377
.......................
49/50 [============================>.] - ETA: 1s - loss: 14.8550
50/50 [==============================] - 61s 1s/step - loss: 14.8404 - val_loss: 14.7532
Epoch 00034: early stopping

[qqwweee]

What about the performance? You can try a gradually decreasing learning rate.

[douxiao]

How to change the code can reduce the learning rate. I'm also trying to reduce the learning rate, but I don't know how to change it? In the train.py file? Should change that part?

[qqwweee]

See the latest code. keras.callbacks.ReduceLROnPlateau is used.

[franklu323]

Hi, I was training my own data and after 50 epochs, it doing the unfreeze process but I met the gpu ran out of memory. how can I avoid this problem? And is that ok if I just skip this step??

@@
@qqwweee

[sherlockchou86]

@franklu323 set smaller batch_size

[xiongfeiliu]

Hi I want know what setting changed when you train you own data.just fix path and yolo3.cfg?

[LamzZ2]

性能如何？您可以尝试逐渐降低学习率。

Hello, I used the five categories in the voc2012 dataset. About 8700 images were trained 100 times from the beginning. I did not use the pre-training weights, and the final loss was about 14. But the effect of the test is not ideal. Some test images don't even have a box. I want to ask what is the situation. Is it still necessary to continue training to reduce the loss?

[eain3314]

修改的效果可以了吗？

[eain3314]

微信 lyb864770486 能讨论一下吗