search

TuSimple / TuSimple-DUC

Understanding Convolution for Semantic Segmentation
https://arxiv.org/abs/1702.08502
Apache License 2.0
605 stars 118 forks source link

Reproduce failed. #22

Open acgtyrant opened 6 years ago

acgtyrant commented 6 years ago

I write resnet.symbol by myself and use it as pretrained model to train TuSimple-DUC, and keep cfg consistent with the paper except cropped size. See my fork.

However, train_IoU is only 0.563 while I can not validate temporarily #21 , I think the val_IoU should not be too far away:

01-03 02:23:28 Epoch[19] Train-acc_ignore=0.932954
01-03 02:23:28 Epoch[19] Train-IoU=0.563348
01-03 02:23:28 Epoch[19] Train-SoftmaxLoss=0.212817
01-03 02:23:28 Epoch[19] Time cost=10693.555
01-03 02:23:29 Saved checkpoint to "../models/ResNet_DUC_HDC_CityScapes/2017_12_31_15:19:05/ResNet_DUC_HDC_CityScapes-0020.params"

Anyone reproduce successfully?

shipengai commented 6 years ago

Hi , what is acc_ignore? I have the same question. Iou is lower much

LeiyuanMa commented 5 years ago

I came into the same question as you,then I find out that I should use the ResNet_DUC_HDC_CityScapes-0020.params as the pretrained model,and set the begin epoch to 20, rather than the init.params and begin the train at epoch 0..hope it helps