For suggestions to reproduce on cityspaces

[hustcc19860606]

I try to reproduce on cityspaces based on 4XP100 GPUs. And try the combination of loss function. The experimental parameters are as follows：
1、without aux and se-loss
CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname 1-dranet101 --base-size 2048 --crop-size 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8
The outcome of training stage:
=>Epoch 239, learning rate = 0.0000, previous best = 0.7571
Train loss: 0.172: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 371/371 [10:34<00:00, 1.71s/it]
pixAcc: 0.941, mIoU: 0.567: 100%|
The outcome of test stage after 240 epochs::
██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 125/125 [04:37<00:00, 2.22s/it]
pixAcc: 0.9393, mIoU: 0.6456
2、without se-loss
CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname dranet101 --base-size 2048 --crop-size 768 --epochs 450 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8 --aux
The outcome of test stage after 240 epochs:
pixAcc: 0.9350, mIoU: 0.6230: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 125/125 [04:43<00:00, 2.27s/it]
pixAcc: 0.9350, mIoU: 0.6230

3、use aux and se-loss
CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname 3-dranet101 --base-size 2048 --crop-size 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8 --aux --se-loss
The return value is not enough：
File "/opt/conda/lib/python3.7/site-packages/torch_encoding-1.2.2b20201207-py3.7.egg/encoding/nn/loss.py", line 89, in forward
pred1, se_pred, pred2, target = tuple(inputs)
ValueError: not enough values to unpack (expected 4, got 3)

can you give me some suggestions to reproduce the outcome of paper?

@junfu1115

[sundongwei]

only add --aux, you did not error?

[sundongwei]

i think you modify the loss function ,otherwise , you only add --aux will occur error yes?

[hustcc19860606]

i think you modify the loss function ,otherwise , you only add --aux will occur error yes?

I only add some 'print' statements in the code, and i find use 'CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname dranet101 --base-size 2048 --cropize 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8 --aux ' , then, args.aux= True and args.se-loss= False.

If i try to reproduce, what's the statement of args.aux and args.se-loss? I find if args.aux= True and args.se-loss= True, there will be error for not enough inputs.

[sundongwei]

i think you modify the loss function ,otherwise , you only add --aux will occur error yes?

I only add some 'print' statements in the code, and i find use 'CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname dranet101 --base-size 2048 --cropize 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8 --aux ' , then, args.aux= True and args.se-loss= False.

If i try to reproduce, what's the statement of args.aux and args.se-loss? I find if args.aux= True and args.se-loss= True, there will be error for not enough inputs.

i mean i use the same args like you ,"CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname dranet101 --base-size 2048 --cropize 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8 --aux " is error .so the model actually use another loss function. you can check it in my issue 39
#124

[hustcc19860606]

i think you modify the loss function ,otherwise , you only add --aux will occur error yes?

I only add some 'print' statements in the code, and i find use 'CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname dranet101 --base-size 2048 --cropize 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8 --aux ' , then, args.aux= True and args.se-loss= False.
If i try to reproduce, what's the statement of args.aux and args.se-loss? I find if args.aux= True and args.se-loss= True, there will be error for not enough inputs.

i mean i use the same args like you ,"CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset citys --model dran --backbone resnet101 --checkname dranet101 --base-size 2048 --cropize 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 8 --multi-grid --multi-dilation 4 8 16 --os 8 --aux " is error .so the model actually use another loss function. you can check it in my issue 39
#124

I think the loss funtion is the same as version-0.5. Do you reproduce the accuracy of paper?

[sundongwei]

i do not reproduce the accuracy of paper, i modify the code for my project

[wlj567]

Hello, do you have the complete original code? Is it convenient to add a contact information? Thank you!