Centermask2
Opened this issue · 5 comments
@davamix
Have you tested training using Centermask2
Yes, I tested but the results were not really good in comparison with Detectron2, as you can see in the image.
I used the VoVNetV2-99 to train the model for 4200 iterations instead of 700 as I did with Detectron2.
I uploaded the changes in train.py and inference.py to this Gist. If you want to test it, clone the Centermask2 repository and copy those files into the root folder, also copy the same loader.py and the data folder with the images.
I don't know why these results, maybe there is something wrong in the configuration, maybe the dataset is to small or it need to be trained for more iterations, it's a matter of more experiments.
hmmm...
i am currently training an centermask_lite_V_19_slim_dw_eSE_FPN_ms_4x
model, it seems to take a long time to converge, but i will continue training.
stay tuned for results in the future
hmmmm....
I trained almost 70,000 iterations and the total_loss: 0.6
, it wont go below 0.6
Tested training both V_39
and V_39
.
I posted an issue to the developer, awaiting his reply
youngwanLEE/centermask2#21
you are right, Centermask2 is a waste of time