search

ooooverflow / BiSeNet

BiSeNet based on pytorch
394 stars 77 forks source link

train val accuracy is not as high as mentioned. Plus res101 accuracy curve is not stable #7

Closed 1093842024 closed 5 years ago

1093842024 commented 5 years ago

i train with context_path: resnet101 and resnet 18. first question: both the validation accuracy hardly reach 0.9, Mostly stop as 0.88-0.89. second question: while the resnet101 training, the validation accuracy will fluctuate a lot after epoch 100, and will drop a lot after about epoch 180+. could you please share your training parameter like lr,batchsize,GPU num,crop_height,crop_width and some detailed trick?

hubutui commented 5 years ago

Try another optimizer? The authors use SGD, and this repo use RMSprop, not sure why. Why not have a try with Adam, or just the simple SGD with momentum >= 0.9, and adjust your learning rate according to learning curve in tensorboardx.

JunjieZhouwust commented 5 years ago

Which dataset did you train? I train resnet101 with CamVid, train loss is close to 0.3, precision_val is close to 0.25 after 35 epoch. Do you know what is the problem?