Hello, in the RGB SOD you provided, saliency maps of gateNet_resnext101 is incomplete, could you send the complete maps to me? Thank you.

It will generate an image like this.

Because author might forgot to uncomment the line 71

prediction = F.sigmoid(prediction)

First of all, thank you for the wonderful code shared.

I am glad that I could execute it well with all modules implemented. All the evaluation metric values also have been obtained correctly. As found in your paper, I would like to plot the PR curve. Can you kindly share an advise on plotting the PR curve?

Thank you in advance.

The parameters of this model definition
"net = GateNet().cuda()"
are missing.
Could you please provide the parameters?

hello!
I want to know what's the meaning of the ckpt_path and exp_name in the generate_sal.py. how to define it?

Hello, I created the same environment as you, but I still can't succeed when running the build salmap, he prompts me that I need to enter two parameters. Can you help me?

File "generate_sal.py", line 83, in
main()
File "generate_sal.py", line 41, in main
net = GateNet().cuda()
TypeError: init() missing 2 required positional arguments: 'block1' and 'layers'

When I'm done training on my own dataset，there was a problem while testing。Can you tell me why?

net = GateNet().cuda()
TypeError: init() missing 2 required positional arguments: 'block1' and 'layers'

Thanks for your sharing.

When I test the model, I have a problem in the second step, which is "run the generate_visfeamaps.py". The code that is "net = RGB_sal().cuda()", what is the "RGB_sal"? If it is your "GateNet", however, there should be 4 outputs while your GateNet only have one output. Could you please tell you how to deal with it?

Hi, thanks for sharing your implementations and results.
But I find the url for RGB-D SOD results is inaccessible now.
Would you mind share it again?

hello,

i trained the model at first at 50k iterations.

[iter 49997], [total loss 0.03742],[loss1 0.01920],[loss1 0.01822],[lr 0.0000002054854] 
[iter 49998], [total loss 0.03780],[loss1 0.01941],[loss1 0.01839],[lr 0.0000001586120] 
[iter 49999], [total loss 0.03794],[loss1 0.01949],[loss1 0.01845],[lr 0.0000001101169] 
[iter 50000], [total loss 0.03843],[loss1 0.01973],[loss1 0.01870],[lr 0.0000000590102] 

i have resumed training it to reach 100k iterations but in the logs it starts at 0 iterations.

{'iter_num': 100000, 'train_batch_size': 4, 'last_iter': 0, 'lr': 0.001, 'lr_decay': 0.9, 'weight_decay': 0.0005, 'momentum': 0.9, 'snapshot': '50000'}

[iter 1], [total loss 0.04253],[loss1 0.02118],[loss1 0.02136],[lr 0.0010000000000] 

isnt it supposed to start at 50k going up ? or i can just reduce my iter_num to be 50k again ?

what do you advice ?

Thanks for the valuable contribution!
I'm reproducing generate_ visfeamaps.py, there is a mistake. I don't know how to correct it. Could you help me to answer it?
The error is as follows:
output1,output2,output3,output4 = net(img_var)
ValueError : not enough values to unpack (expected 4, got 1)

I have seen the GateNet network structure, and there is only one output, I modify the code as follows
output = net(img_var)
And I got only one graph, like a heat map, not salient map.

A part of the training log:

[iter 99500], [total loss 0.04293],[loss1 0.02237],[loss1 0.02055],[lr 0.0000085085186]
[iter 99600], [total loss 0.04290],[loss1 0.02238],[loss1 0.02052],[lr 0.0000069635378]
[iter 99700], [total loss 0.04236],[loss1 0.02209],[loss1 0.02027],[lr 0.0000053791028]
[iter 99800], [total loss 0.04294],[loss1 0.02237],[loss1 0.02057],[lr 0.0000037400418]
[iter 99900], [total loss 0.04292],[loss1 0.02237],[loss1 0.02055],[lr 0.0000020132107]
[iter 100000], [total loss 0.04314],[loss1 0.02248],[loss1 0.02066],[lr 0.0000000316228]

hello,

i trained a model at 50k iterations , now am testing and the maps generated are very weird.

我在数据集DUTS上进行10000次迭代的情况下，分别进行了无Fold-ASPP模块和有Fold-ASPP模块的训练；
进行无Fold-ASPP的修改为将self.dem1修改为如下代码
self.dem1 = nn.Sequential(nn.Conv2d(2048, 512, kernel_size=3, padding=1), nn.BatchNorm2d(512), nn.PReLU())
同时将self.layer4的stride调整为2
self.layer4 = self._make_layer(block1, 512, layers[3], stride=2)

我的预期效果是有Fold-ASPP的训练好于无Fold-ASPP的训练效果，但结果恰好相反，请问可能的原因是什么？(训练得到的测试结果如下)
迭代10000次，无fold-ASPP

DUTS-TE MAE: 0.0511 maxF: 0.8533 avgF: 0.7539 wfm: 0.7512 Sm: 0.8570 Em: 0.8677

迭代10000次，有Fold-ASPP

DUTS-TE MAE: 0.0526 maxF: 0.8468 avgF: 0.7622 wfm: 0.7505 Sm: 0.8523 Em: 0.8735