I tried creating a training model and FANet-18 with cityscapes dataset. I replaced the InPlaceABN layers with normal BN followed by Activation as I needed to parse the trained model to ONNX for deploying in my application.
These are my training configurations highly adapted from the paper:
Mini-Batch SGD with Batch Size 4 as I only have 8 gigs of GPU Memory, weight decay = 5e-4, momentum = 0.9
Initial Learning Rate(LR) = 1e-2, with update to LR multiplied by a factor (1-(iter/max_iter)pow(2))
Data Augmentation - Horizontal Flipping, random scaling (0.75 to 2)
Training iterations 80000
I resulted with a OHEM Cross Entropy loss of 0.3941 in the final iteration
I am yet to check the mIOU.
As a preliminary discussion I would like to compare it with BiseNet which was trained in a similar fashion but with auxiliary losses and resulted with a OHEM Cross Entropy Loss of 0.2947 which resulted in a mIOU of 0.63
Could you please give me more details on the training especially
What is the number of Iterations you trained the model for?
What was the Final Cross Entropy loss you ended up with?
Did you use Auxiliary losses as well for better Convergence resulting in lower loss?
Did you have any specific modified version of Cross Entropy Loss to achieve better convergence?
Is there any other thing that I am missing out to achieve better results
Hello once again,
I tried creating a training model and FANet-18 with cityscapes dataset. I replaced the InPlaceABN layers with normal BN followed by Activation as I needed to parse the trained model to ONNX for deploying in my application.
These are my training configurations highly adapted from the paper:
4as I only have 8 gigs of GPU Memory, weight decay =5e-4, momentum =0.91e-2, with update to LR multiplied by a factor(1-(iter/max_iter)pow(2))0.75 to 2)80000I resulted with a OHEM Cross Entropy loss of
0.3941in the final iterationI am yet to check the mIOU.
As a preliminary discussion I would like to compare it with BiseNet which was trained in a similar fashion but with auxiliary losses and resulted with a OHEM Cross Entropy Loss of
0.2947which resulted in a mIOU of0.63Could you please give me more details on the training especially
Is there any other thing that I am missing out to achieve better results