Benchmarking on Cityscapes
koda12344505 opened this issue · comments
Hello,
Do you have any results on Cityscapes datasets?
I just wonder rmi loss will bring better performance on cityscapes datset
Thank you
We did not test RMI on Cityscapes.
Our GPU resource (4 GTX 1080s) did not allow us to run DeepLabv3+ on Cityscapes datasets with batch_size=16, even with output_stride=16. So we choose the CamVid dataset,
which contains similar images with Cityscapes, i.e., the street view images.
We also hope someone can test RMI on Cityscapes and report the results.
no response for a long time
@mzhaoshuai Nice work!
I want to try your method on Cityscapes with batch-size=16. I am curious how should I modify the hyperparameters to adapt to the training on the Cityscapes with crop size 1024x512.
Currently, I find the RMI loss values based on the set of default parameters is negative values (the red line marked values after decode_1.loss_seg: )
_euler_num = 2.718281828 # euler number
_pi = 3.14159265 # pi
_ln_2_pi = 1.837877 # ln(2 * pi)
_CLIP_MIN = 1e-6 # min clip value after softmax or sigmoid operations
_CLIP_MAX = 1.0 # max clip value after softmax or sigmoid operations
_POS_ALPHA = 5e-4 # add this factor to ensure the AA^T is positive definite
_IS_SUM = 1 # sum the loss per channel