search

qqwweee / keras-yolo3

A Keras implementation of YOLOv3 (Tensorflow backend)
MIT License
7.14k stars 3.44k forks source link

bottlenecks features nan #710

Open ouarda253 opened 4 years ago

ouarda253 commented 4 years ago

I use a dataset of 800 images , using CPU . During training I had a val loss and loss :NAN when I looked for the cause i found bottlenecks features are NAN . Any Help ?? @qqwweee

datduonguva commented 4 years ago

Check if there is box that are too thin. If an original image has a very thin box, after scaling, one of it dimensions become 0, loss would be 0 in that case.

On Mon, Jul 20, 2020, 4:10 PM ouarda253 notifications@github.com wrote:

I use a dataset of 800 images , using CPU . During training I had a val loss and loss :NAN when I looked for the cause i found bottlenecks features are NAN . Any Help ?? @qqwweee https://github.com/qqwweee

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/qqwweee/keras-yolo3/issues/710, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFSNTUM6R6H3QPAHFCE7GEDR4SQFBANCNFSM4PCXYWVA .

ouarda253 commented 4 years ago

I found that is because the "process_true_boxes" function return y_true=np.zeros y_true[l][b, j, i, k, 0:4] = true_boxes[b,t, 0:4] y_true[l][b, j, i, k, 4] = 1 y_true[l][b, j, i, k, 5+c] = 1 return y_true @datduonguva @qqwweee what can be the probleme ?

monicatao commented 4 years ago

Check if there is box that are too thin. If an original image has a very thin box, after scaling, one of it dimensions become 0, loss would be 0 in that case. … On Mon, Jul 20, 2020, 4:10 PM ouarda253 @.***> wrote: I use a dataset of 800 images , using CPU . During training I had a val loss and loss :NAN when I looked for the cause i found bottlenecks features are NAN . Any Help ?? @qqwweee https://github.com/qqwweee — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#710>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFSNTUM6R6H3QPAHFCE7GEDR4SQFBANCNFSM4PCXYWVA .

Hi the train data images will be the original images without the boxes right? Why will there be boxes that are too thin? Can you explain?

monicatao commented 4 years ago

I found that is because the "process_true_boxes" function return y_true=np.zeros y_true[l][b, j, i, k, 0:4] = true_boxes[b,t, 0:4] y_true[l][b, j, i, k, 4] = 1 y_true[l][b, j, i, k, 5+c] = 1 return y_true @datduonguva @qqwweee what can be the probleme ?

Hi I have the same problem. Did you solve it?