CoinCheung
About BatchNormalization
hi by running your project,I can reach high mIoU,but when the Batchnormlization is setting all SYNC in ResNet,the mIoU can just reach 70%,why is that?
@Cynicsss How did you control the BN sync or not sync? I found the bn in resnet is original all sync already? What do you mean by ,but when the Batchnormlization is setting all SYNC in ResNet,the mIoU can just reach 70%
@Cynicsss How did you control the BN sync or not sync? I found the bn in resnet is original all sync already? What do you mean by
,but when the Batchnormlization is setting all SYNC in ResNet,the mIoU can just reach 70%
https://github.com/CoinCheung/DeepLab-v3-plus-cityscapes/blob/e31cc7567f87c44e7ef449eec738e8bb18c3605a/models/resnet.py#L51
@Cynicsss How did you control the BN sync or not sync? I found the bn in resnet is original all sync already? What do you mean by
,but when the Batchnormlization is setting all SYNC in ResNet,the mIoU can just reach 70%https://github.com/CoinCheung/DeepLab-v3-plus-cityscapes/blob/e31cc7567f87c44e7ef449eec738e8bb18c3605a/models/resnet.py#L51
Do you know why the author need to use the original nn.BatchNorm2drather than synced bn here ?
@Cynicsss Hi, how long did it take you to reach 70% mIoU? I tried the code, but it takes one night just training about 12000 iterations.
@Cynicsss I use 3 *1080ti to train on CityScapes without any change, but just reaching 70.23% mIoU. Is there something I should change in order to reach higher mIoU? BTW, it costed me about 2.5 days to train 41000 iterations..so sad.
@Cynicsss I use 3 *1080ti to train on CityScapes without any change, but just reaching 70.23% mIoU. Is there something I should change in order to reach higher mIoU? BTW, it costed me about 2.5 days to train 41000 iterations..so sad.
It takes me 4 hours to train 15000 iterations with 2 RTX2080 and finally I got 78.2 mIoU.
@LetsGoFir can you please confirm if this repository is reproducible for cityscapes? I'm also using 2RTX 2080 GPUs and facing issues getting 75%+ mIoU on CS. Also any idea if there are any differences between this implementation and the official implementation?
