pred = torch.argmax(input[:, :2], dim=1)

[Note-Liu]

In model.py:

def accuracy_gender(input, targs):
pred = torch.argmax(input[:, :2], dim=1)
y = targs[:, 0]
return torch.sum(pred == y)

I want to know why not pred = input[:,:1] ?
I think first two columns of input are genders and races.

[sajjjadayobi]

first two columns are logits for (male, female) and the second one is for Age

as you can see I've used last index (aka 3) for age

def l1loss_age(input, targs):
    return F.l1_loss(input[:, -1], targs[:, -1]).mean()

[Note-Liu]

first two columns are logits for (male, female) and the second one is for Age

as you can see I've used last index (aka 3) for age

def l1loss_age(input, targs):
    return F.l1_loss(input[:, -1], targs[:, -1]).mean()

Thanks for your reply!

[Note-Liu]

first two columns are logits for (male, female) and the second one is for Age
as you can see I've used last index (aka 3) for age

def l1loss_age(input, targs):
    return F.l1_loss(input[:, -1], targs[:, -1]).mean()

Thanks for your reply!

In Age & Gender Estimation,did you use the original image of UTKFace Dataset to train ShufflenetFull ?

[sajjjadayobi]

Although I do not know what you mean by original, I train on UTKFace Dataset with ShufflenetFull as the backbone

[Note-Liu]

Although I do not know what you mean by original, I train on UTKFace Dataset with ShufflenetFull as the backbone

Thanks!I see what you mean.
I have another question to ask：

def multitask_loss(input, target):
input_gender = input[:, :2]
input_age = input[:, -1]
loss_gender = F.cross_entropy(input_gender, target[:, 0].long())
loss_age = F.l1_loss(input_age, target[:, 2])

return loss_gender / (.16) + loss_age * 2

I want to know "loss_gender / (.16) + loss_age * 2"，how is the ratio determined? Is it from the experiment？

[sajjjadayobi]

Yep, the loss ratio results from some experiments
Depending on how much you care about gender detection or age estimation.

[Note-Liu]

Yep, the loss ratio results from some experiments
Depending on how much you care about gender detection or age estimation.

Thank you for your patience！