DIGITS
DIGITS copied to clipboard
NVIDIA
Detectnet for several classes
I saw how to train detectnet for 1 and 2 classes and i managed to train them (https://github.com/NVIDIA/DIGITS/issues/1359). Can I train 3, 4 etc multiclass detectnet classifier? And if yes how? Thanks
Hi.I tried the network for 2 classes for Kitty dataset.It is only learning the first class of car.How can i solve it.Please help me to solve.
@csadom @shinaushin @sulth I wrote two blog posts about multiclass custom object detection. Here you go:
https://labs.coria.com/blog/computer-vision/PreparingDataForCustomObjectDetectionUsingNvidiaDigits?sc_camp=33AFA8630062426190B5760C8FDF17CF https://labs.coria.com/en/blog/computer-vision/TrainingACustomMulticlassObjectDetectionModel?sc_camp=33AFA8630062426190B5760C8FDF17CF
I hope it helps. Let me know if you have any questions.
Hi @elaith9 I was reading your posts, and a lot of posts. Please if someone already solved this problem please help me. There are very interesting. But I'm having a big problem. In fact is the same problem that you were covering in posts. I'm trying to have a two detectnet detection. However my training process shows only one of my classes with a map different from Zero the other class remains in Zero after 60 epochs of training as you can see in the image below.
And the configuration on the model is this:
Hi @kala855, did you solve the problem?
I'm facing the same problem as you, should we use different network define?
Here is my problem, mAP of class 1 = 0
@dqthebt24 make your current model pretrained and train again for 300-600 epochs. Success will also depend on quality of your training data and labels.
@kala855 60 epochs is not enough to train your model. Consider training it for at least 600 epochs.
@elaith9 Thank you, I did with KITTI dataset for 300 epoches. mAP of class 1 always equals to 0 through all epoches, (mAP of class 0 is #0)
@dqthebt24 please read my blog posts. It took me 600 epochs with first model which I later made pretrained and used it's weights to train for another 300 epochs.
@elaith9 thanks for your answer. @dqthebt24 I was trying with detectnet but I suppose that in fact I didn't use enough epochs. Im going to test again.
@elaith9 , @kala855 There are lines in the .prototxt file:
data_param { backend: LMDB source: "examples/kitti/kitti_train_images.lmdb" batch_size: 10 }
I checked my pc, it had no path "examples/kitti/kitti_train_images.lmdb".
Do we need above lines in .prototxt file? or if yes, how to generate the path of ".lmdb"?
I wrote a small article on how to crate a dataset and train a two class DetectNet on Digits. You can read my post here
https://www.deeplearning-blog.com/2020/01/31/how-to-train-a-two-class-detectnet-neural-network-on-digits/
Hope it helps Cheers Marco
@dqthebt24 please read my blog posts. It took me 600 epochs with first model which I later made pretrained and used it's weights to train for another 300 epochs.
@elaith9 could you please tell me what was the learning rate for the next 300 epochs after making a pretrained model?
