yihengzhang-cv / lps-net Goto Github PK

这篇论文的结果是在1080ti +TensorRT上对比的，也没说TensorRT采用的什么精度，而DDRNet等是在pytorch上跑的，不能直接比较

I trained the lpsnet-m on ImageNet 1k for 100 epoch with batch size 1024.The detials are as follows:
Optimizer:

Data augments:

Eval dataset:

result:

Then I use the pre-trained model to train a lpsnet-m on cityscapes for 180k iterations with batch size 16. The detials are as follows:
Loss function:

lr_scheduler:

Optimizer:

Data augments:

Eval dataset:

Result:

it didn't upsample to original input size, is this a bug ?

hello could you tell me this is why

I'm wondering that did the miou score reported by table 5 in the paper use the ImageNet pre-trained?
And can you provide the exact training settings of ImageNet pre-train and cityscapes, including but not limited to data augments、optimizer settings and so on.
Thank you in advance!