search

MehmetAygun / fusenet-pytorch

Other
80 stars 17 forks source link

How could I get a better result? #15

Closed zeng32 closed 4 years ago

zeng32 commented 4 years ago

Hi,

When I run the code successfully(with the required environment), below is the result on SUNRGBD dataset: epoch 400 test loss: 3.094 glob acc : 64.37, mean acc : 25.19, IoU : 18.14 It's far from the posted results: 75.4% | 46.48% | 35.69%

So, would you please enlighten me on how can I improve the result?

BTW, below is the command I use with 4 RTX 2070 GPUs: python train.py --dataroot datasets/sunrgbd --dataset sunrgbd --name sunrgbd --gpu_ids 0,1,2,3 --display_id 1 | tee ./run.log

MehmetAygun commented 4 years ago

Hey, it seems that it is just last epoch result, where model could be overfit until then. In log file you can inspect the situation better.

zeng32 commented 4 years ago

Thanks for the reply. I find it interesting that if I set --gpu_ids 0,1,2,3, then 4 GPUs are utilized, but the IoU result is bad:18%. But with the default setting(only on GPU 0), the result could be as high as 35%. Would you please help me figure out why changing in gpu_ids effects the result?

hazirbas commented 4 years ago

Hi, have you tried different learning or a smaller batch size? SGD-based optimization schemes are sensitive to hyper-params for multi-gpu trainings.

zeng32 commented 4 years ago

Right now my options are: batch_size: 4 lr: 0.001
lr_decay_epochs: 25
lr_decay_iters: 5000000
lr_gamma: 0.9
lr_policy: lambda I'll try different settings. Thanks again!