search

juhongm999 / hsnet

Official PyTorch Implementation of Hypercorrelation Squeeze for Few-Shot Segmentation, ICCV 2021
231 stars 43 forks source link

about effect #2

Closed ily666666 closed 3 years ago

ily666666 commented 3 years ago

I download your code,and train it according to the readme.txt, but when I train it 150 epoches,Its mIOU up to 82.47%.So ,can you tell me how should I train to achieve the effect of the essay.

ily666666 commented 3 years ago

backbone is resnet50

juhongm999 commented 3 years ago

Can you provide us the command that you run? On which benchmark are you training the model? Also, which mIoU are you talking about (training/validation)?

ily666666 commented 3 years ago

python train.py --backbone resnet50 --benchmark fss --lr 1e-3 --bsz 10 And this is one info in log.txt Validation [@Epoch 138] Avg L: 0.12781 mIoU: 82.47 FB-IoU: 89.57

ily666666 commented 3 years ago

Sorry ,I don't mean anything else, I am just puzzled

juhongm999 commented 3 years ago

Training HSNet on FSS-1000 dataset takes approximately 3 days until convergence (when trained with four 2080 Ti GPUs). If you are training with one GPU only, it will take longer. As you see in our tensorboard log we provided in the README.md (https://drive.google.com/file/d/1i9WlwCEqK4XOdBRh0nShtxxEdWe-1q_r/view), the best model is retrieved when the validation mIoU is the highest (82.85) at epoch 731 (as seen below).

스크린샷 2021-04-13 오후 3 09 18

Also FSS-1000 consists of training/validation/test splits. The model (having val mIoU of 82.85) should be further tested using command below:

python test.py --backbone resnet50 --benchmark fss --nshot 1 --load "path_to_trained_model/best_model.pt"

which gives 1-shot results of 85.5 (mIoU) on the test split of FSS-1000. If you have any further issues, please leave us a comment.