請問如何重現EncNet的結果

[wuhuikai]

請問如何重現EncNet的結果

[wuhuikai]

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset pcontext \
    --model encnet --dilated --aux --se-loss \
    --backbone resnet50 --checkname encnet_res50_pcontext

[HolmesShuan]

Sorry to bother you again. I wonder how to reproduce the results of DeepLabv3 and PSPNet, reported in Table 1. I use the following code to train DeepLabv3, but it converges pretty slow:

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset pcontext \
    --model deeplab --aux --se-loss\
    --backbone resnet50 --checkname deeplab_res50_pcontext --workers 8 --no-val

[wuhuikai]

@HolmesShuan Training for 80 epochs is enough, which should finish within 1 day on 4 Titan-Xp GPUs.

[HolmesShuan]

Thanks for your responses, and should --se-loss be used in ASPP and PSPNet?

[wuhuikai]

@HolmesShuan No, it's used in EncNet only.

[henry8527]

may I ask that for reproducing original EncNet, when to use the args['lateral'] ?

[wuhuikai]

The official code shows that they use lateral when the dataset is Pascal. (link)

[henry8527]

Thanks, but why your command didn't use --lateral?
Since it's also using Pascal family dataset (Pascal-Context).
#16 (comment)

[wuhuikai]

Ops, I forgot lateral in the above command.

[henry8527]

How about the correct script for reproducing PSPNet baseline without jpu? thanks a lot.

[wuhuikai]

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset pcontext \
    --model psp --dilated --aux \
    --backbone resnet50 --checkname psp_res50_pcontext