training is good, eval is worse

[ailias]

When I trained hg model using 300w datasets, and the training result showed is good, but is worse when I use eval mode. Maybe it's because of the bn layer. Have you met this problem? What should I do for it?

[raymon-tian]

@ailias No, I didn't meet the problem. Did you set net.eval() before evaluation？

[thruster1996]

@ailias ，I meet the same problem just like you. After finishing training, if I set the net to eval, then output of the net seems to be much more worse than the train mode:(

[TingmanYan]

I have met the same problem when training the hg model on MPII dataset.
I'm training on a Titan Xp GPU, Pytorch 0.4.1.
Changing hg.py, line 22-23

self.res2 = Residual(128, 128)
self.res3 = Residual(128, self._nFeats)

to

self.res2 = Residual(128, self._nFeats)
self.res3 = Residual(self._nFeats, self._nFeats)

solved my problem.
Hope it will help you too.

[4Statistics]

When I trained hg model using 300w datasets, and the training result showed is good, but is worse when I use eval mode. Maybe it's because of the bn layer. Have you met this problem? What should I do for it?

Can you tell me how to run the code ?
python train.py

Traceback (most recent call last):
File "train.py", line 123, in
net = KFSGNet()

TypeError: new() received an invalid combination of arguments - got (float, int, int, int), but expected one of:

• (torch.device device)
• (torch.Storage storage)
• (Tensor other)
• (tuple of ints size, torch.device device)
• (object data, torch.device device)

---

Thank you very much .
Best wishes.

[silvercherry]

请问能将训练集分享出来吗

[silvercherry]

When I trained hg model using 300w datasets, and the training result showed is good, but is worse when I use eval mode. Maybe it's because of the bn layer. Have you met this problem? What should I do for it?

Can you tell me how to run the code ?

python train.py
Traceback (most recent call last):
File "train.py", line 123, in
net = KFSGNet()

TypeError: new() received an invalid combination of arguments - got (float, int, int, int), but expected one of:

• (torch.device device)
• (torch.Storage storage)
• (Tensor other)
• (tuple of ints size, torch.device device)
• (object data, torch.device device)

Thank you very much .
Best wishes.

have you solve this problem?

[ilaij0810]

in models.py, from line 64, should be modified:
nn.Conv2d(ins,int(outs//2),1),
nn.BatchNorm2d(int(outs//2)),
nn.ReLU(inplace=True),
nn.Conv2d(int(outs//2),int(outs//2),3,1,1),
nn.BatchNorm2d(int(outs//2)),
nn.ReLU(inplace=True),
nn.Conv2d(int(outs//2),outs,1)