mit-han-lab
Hyperparameter request for reproducibility
I'm training the segmentation EfficientViT B1 on Cityscapes, and achieving ~0.6 mIoU, however the reported results are around 0.8 mIoU.
Would you be able to offer some guidance or share more details around the hyperparameters?
My setup is:
- 1024 x 2048 resolution
- Backbone starts from the ImageNet checkpoints you have provided
- Num Epochs: 100
- LR: 0.005 with cosine annealing to 0
- Batch Size: 2 (I'm limited by hardware at this resolution)
- AdamW optimizer
- Focal Loss w/ equal class weights
Any augmentations? Anything else that can help?
I'm training the segmentation EfficientViT B1 on Cityscapes, and achieving ~0.6 mIoU, however the reported results are around 0.8 mIoU.
Would you be able to offer some guidance or share more details around the hyperparameters?
My setup is:
- 1024 x 2048 resolution
- Backbone starts from the ImageNet checkpoints you have provided
- Num Epochs: 100
- LR: 0.005 with cosine annealing to 0
- Batch Size: 2 (I'm limited by hardware at this resolution)
- AdamW optimizer
- Focal Loss w/ equal class weights
Any augmentations? Anything else that can help?
Hi @ovunctuzel-bc , this seems good hyperparameters for training. First of all one thing to ask, there is no official release of training code for segmentation EfficientViT except for SAM variant right? How did u get the code reference. If you just guide then it would be very useful for me as well.
A fairly standard pytorch training loop seems to work fine. The results are satisfactory but not quite at the level of the pretrained model.