Replicating the Pretrained Semantic Segmentation

I would like to reproduce the reported accuracies of pre-trained segmentation models.

I executed the training code from references under default parameters with dataset coco. The trained DeepLabV3 ResNet101 reaches in the validation-set about 60% IoU (with and without the auxiliary loss). This is much lower than the reported 67.4%.

I do appreciate any help on this issue. Thanks a lot in advance :).

Hi, were you able to resolve your problem?
I am also facing a similar issue, I have removed an aditional 3000 images from training set, as they were overlapping with my validation and test set, total training images after filtering are 92,034. Best Mean IOU accuray of 53.5 occured at 8th epoch, and it has been fluctuating between 50-53 since 2nd epoch.
I am using the reference scripts for segmentation and following command line for training

OMP_NUM_THREADS=4 CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch --nproc_per_node=4 --use_env --lr 0.02 --dataset coco -b 4 -j 4 --model deeplabv3_resnet101 --aux-loss

Hi Kanishk,
No, I haven’t been able to resolve this. In PyTorch GitHub it’s mentioned all models have been trained on 8x V100 GPUs with the following command

python -m torch.distributed.launch --nproc_per_node=8 --use_env --lr 0.02 --dataset coco -b 4 --model fcn_resnet50 --aux-loss

I only have access to two GPUs. As a work around, I tried with nproc_per_node=2 and -b 16, but still couldn#t reproduce the reported accuracies.