Semantic Segmenataion Model Problem

vision
21

@MariosOreo Thanks for replying. I use CosineAnnealingLR.
it gives good results upto 5 epoches but after that loss again starts decreasing slow.

optimizer = optim.Adam(model.parameters(),lr=0.0001,eps=0.00001)#,betas=(0.9,0.999))
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, len(train_img), eta_min=0)

Loss is like

EPOCH 1
ALL ABOUT LOSS(training)-------- 0.457618580705883
ALL ABOUT LOSS(valid)-------- 12.778249708088962
 
EPOCH 2
ALL ABOUT LOSS(training)-------- 0.4419868331245242
ALL ABOUT LOSS(valid)-------- 12.325974919579245
 
EPOCH 3
ALL ABOUT LOSS(training)-------- 0.4364099946170994
ALL ABOUT LOSS(valid)-------- 12.118723771788858
 
EPOCH 4
ALL ABOUT LOSS(training)-------- 0.43276495430995787
ALL ABOUT LOSS(valid)-------- 11.967909010973843
 
EPOCH 5
ALL ABOUT LOSS(training)-------- 0.43006896931733046
ALL ABOUT LOSS(valid)-------- 11.857121261683377
 
EPOCH 6
ALL ABOUT LOSS(training)-------- 0.42755556564907254
ALL ABOUT LOSS(valid)-------- 11.79303228855133
 
EPOCH 7
ALL ABOUT LOSS(training)-------- 0.425640314745813
ALL ABOUT LOSS(valid)-------- 11.791005622256886
 
EPOCH 8
ALL ABOUT LOSS(training)-------- 0.4244025208636571
ALL ABOUT LOSS(valid)-------- 11.793241988528859
 
EPOCH 9

ALL ABOUT LOSS(training)-------- 0.4226630027867902
ALL ABOUT LOSS(valid)-------- 11.805133386091752
22

@ptrblck and @MariosOreo I think I got my answer. I just modify my model with batchNorm and scheduler and learning rate and it’s going good. Thnx for your help :grinning: . Anything you feel to help me more during training, please suggest.

1 Like