ModuleNotFoundError: No module named 'upsnet.bbox.bbox' about upsnet

[YuwenXiong]

From your information you only compile bbox cython module with python 3.7, make sure you also use python 3.7 when you run experiments.

[YuwenXiong]

Also please note that config files w/o 4gpu suffix are with horovod in distributed setting

[WenFuLee]

�We only have 1 GPU in our environment now. Is it still compatible with the distributed setting?

[YuwenXiong]

Then please edit the 4gpu config, reduce lr by 4x / enlarge #iter by 4x, etc. BTW I don't think you can run coco dataset w/ 1 GPU in a acceptable time, please try cityscapes instead.

[WenFuLee]

I am not so sure to make the setting upsnet_resnet50_cityscapes_4gpu.yaml to fit my current environment with only 1 GPU. I have done the change below.

====

train:
gpus: '0,1,2,3' -> '0'
lr: 0.005 -> 0.00125
max_iteration: 48000 -> 192000
decay_iteration:

• 36000 -> 144000
  warmup_iteration: 1500 -> 6000

====

test:
test_iteration: 48000 -> 192000

====

Did I miss anything? Thanks.

[YuwenXiong]

looks good to me

[WenFuLee]

I know training takes a long time. Could you please give me a rough time how long it may take to finish training a model using only 1 GPU based on the above setting? Thank you!

[YuwenXiong]

It will take ~3 days to train on 1 1080Ti GPU

[WenFuLee]

I see. What if I don't need a model as accurate as your best one? Could I trade off performance for less training time? For example, I can decrease the iteration numbers in order to get a trained model more quickly, which may have lower accuracy. If this is acceptable, what parameters I can try first, iteration number or else? Thank you!

[YuwenXiong]

You can reduce #iter, e.g. 72k/96k rather than 144k/192k

[wshilaji]

兄弟请问怎么自定义自己的coco panoptic格式的数据集