Confusion about Experimental Discrepancies

[DevKPro]

Hello, author. I ran your open-source code on the Cityscapes dataset in my single GPU (3080Ti) environment, and the program ran smoothly. Since the paper used 8 GPUs, and I used only one, I accordingly reduced the learning rate to 1/8 of the original value. However, the experimental results are far from the results provided in your paper. Could you please explain the reason for this? I don't believe that single GPU execution is the main cause of this issue. Are there any other possibilities? The test results after running 80k iterations of the code are as follows: ` [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 500/500, 4.0 task/s, elapsed: 126s, ETA: 0s

2024-02-04 02:55:51,145 - mmseg - INFO - per class results: 2024-02-04 02:55:51,147 - mmseg - INFO - +---------------+-------+-------+ | Class | IoU | Acc | +---------------+-------+-------+ | road | 90.24 | 96.64 | | sidewalk | 39.15 | 51.54 | | building | 75.24 | 91.19 | | wall | 15.67 | 17.24 | | fence | 2.16 | 2.39 | | pole | 6.68 | 7.13 | | traffic light | 0.06 | 0.06 | | traffic sign | 11.86 | 12.65 | | vegetation | 81.37 | 90.16 | | terrain | 44.97 | 57.3 | | sky | 86.01 | 96.76 | | person | 35.78 | 55.09 | | rider | 0.0 | 0.0 | | car | 67.18 | 86.92 | | truck | 0.0 | 0.0 | | bus | 0.35 | 0.35 | | train | 0.0 | 0.0 | | motorcycle | 0.0 | 0.0 | | bicycle | 19.81 | 23.84 | +---------------+-------+-------+ 2024-02-04 02:55:51,147 - mmseg - INFO - Summary: 2024-02-04 02:55:51,148 - mmseg - INFO - +-------+-------+-------+ | aAcc | mIoU | mAcc | +-------+-------+-------+ | 85.38 | 30.34 | 36.28 | +-------+-------+-------+ 2024-02-04 02:55:51,149 - mmseg - INFO - Exp name: regproxy-s16-sub4+implicit-mid-2+768x768+80k+adamw-poly+cityscapes.py 2024-02-04 02:55:51,150 - mmseg - INFO - Iter(val) [500] aAcc: 0.8538, mIoU: 0.3034, mAcc: 0.3628, IoU.road: 0.9024, IoU.sidewalk: 0.3915, IoU.building: 0.7524, IoU.wall: 0.1567, IoU.fence: 0.0216, IoU.pole: 0.0668, IoU.traffic light: 0.0006, IoU.traffic sign: 0.1186, IoU.vegetation: 0.8137, IoU.terrain: 0.4497, IoU.sky: 0.8601, IoU.person: 0.3578, IoU.rider: 0.0000, IoU.car: 0.6718, IoU.truck: 0.0000, IoU.bus: 0.0035, IoU.train: 0.0000, IoU.motorcycle: 0.0000, IoU.bicycle: 0.1981, Acc.road: 0.9664, Acc.sidewalk: 0.5154, Acc.building: 0.9119, Acc.wall: 0.1724, Acc.fence: 0.0239, Acc.pole: 0.0713, Acc.traffic light: 0.0006, Acc.traffic sign: 0.1265, Acc.vegetation: 0.9016, Acc.terrain: 0.5730, Acc.sky: 0.9676, Acc.person: 0.5509, Acc.rider: 0.0000, Acc.car: 0.8692, Acc.truck: 0.0000, Acc.bus: 0.0035, Acc.train: 0.0000, Acc.motorcycle: 0.0000, Acc.bicycle: 0.2384 `