Fail to train in mini-Imagenet about pytorch-classification-uncertainty HOT 5 OPEN

I use the edl loss to train in mini-imagenet dataset with 64 classes, but the loss can't converge and the accuracy is very low.

I think you can modify the parameters:

def edl_loss(self, func, y, alpha, annealing_step, device="cuda"):
        y = self.one_hot_embedding(y)
        y = y.to(device)
        alpha = alpha.to(device)
        S = torch.sum(alpha, dim=1, keepdim=True)
        
        A = torch.sum(y * (func(S) - func(alpha)), dim=1, keepdim=True)

        # annealing_coef = torch.min(
        #     torch.tensor(1.0, dtype=torch.float32),
        #     torch.tensor(self.epoch / annealing_step, dtype=torch.float32),
        # )
        annealing_coef = 0.1

        kl_alpha = (alpha - 1) * (1 - y) + 1
        kl_div = annealing_coef * self.kl_divergence(kl_alpha, device=device)
       
        return A + kl_div

set annealing_coef as 0.1 or lower setting will work, do not set 1, it's too large.

from pytorch-classification-uncertainty.

I use the edl loss to train in mini-imagenet dataset with 64 classes, but the loss can't converge and the accuracy is very low.

I think you can modify the parameters:

def edl_loss(self, func, y, alpha, annealing_step, device="cuda"):
        y = self.one_hot_embedding(y)
        y = y.to(device)
        alpha = alpha.to(device)
        S = torch.sum(alpha, dim=1, keepdim=True)
        
        A = torch.sum(y * (func(S) - func(alpha)), dim=1, keepdim=True)

        # annealing_coef = torch.min(
        #     torch.tensor(1.0, dtype=torch.float32),
        #     torch.tensor(self.epoch / annealing_step, dtype=torch.float32),
        # )
        annealing_coef = 0.1

        kl_alpha = (alpha - 1) * (1 - y) + 1
        kl_div = annealing_coef * self.kl_divergence(kl_alpha, device=device)
       
        return A + kl_div

set annealing_coef as 0.1 or lower setting will work, do not set 1, it's too large.

Hi, thank you for your answer. I try to set 'annealing_coef' as 0.1 and 0.05 respectively, but it still not works. Do you let it works successfully?

from pytorch-classification-uncertainty.

I use the edl loss to train in mini-imagenet dataset with 64 classes, but the loss can't converge and the accuracy is very low.

I think you can modify the parameters:

def edl_loss(self, func, y, alpha, annealing_step, device="cuda"):
        y = self.one_hot_embedding(y)
        y = y.to(device)
        alpha = alpha.to(device)
        S = torch.sum(alpha, dim=1, keepdim=True)
        
        A = torch.sum(y * (func(S) - func(alpha)), dim=1, keepdim=True)

        # annealing_coef = torch.min(
        #     torch.tensor(1.0, dtype=torch.float32),
        #     torch.tensor(self.epoch / annealing_step, dtype=torch.float32),
        # )
        annealing_coef = 0.1

        kl_alpha = (alpha - 1) * (1 - y) + 1
        kl_div = annealing_coef * self.kl_divergence(kl_alpha, device=device)
       
        return A + kl_div

set annealing_coef as 0.1 or lower setting will work, do not set 1, it's too large.

Hi, thank you for your answer. I try to set 'annealing_coef' as 0.1 and 0.05 respectively, but it still not works. Do you let it works successfully?

I haven't tried this repo, but I have tried to train an 8631 ID face recognition network, we use resnet-100 is ok, can get the same accuracy on the in-distribution dataset as the softmax training method, resnet-50 can't get to convergence. Another thing is that we all find the KL loss will damage the accuracy, try to decrease the coefficients or just remove it.

from pytorch-classification-uncertainty.

I use the edl loss to train in mini-imagenet dataset with 64 classes, but the loss can't converge and the accuracy is very low.

I think you can modify the parameters:

def edl_loss(self, func, y, alpha, annealing_step, device="cuda"):
        y = self.one_hot_embedding(y)
        y = y.to(device)
        alpha = alpha.to(device)
        S = torch.sum(alpha, dim=1, keepdim=True)
        
        A = torch.sum(y * (func(S) - func(alpha)), dim=1, keepdim=True)

        # annealing_coef = torch.min(
        #     torch.tensor(1.0, dtype=torch.float32),
        #     torch.tensor(self.epoch / annealing_step, dtype=torch.float32),
        # )
        annealing_coef = 0.1

        kl_alpha = (alpha - 1) * (1 - y) + 1
        kl_div = annealing_coef * self.kl_divergence(kl_alpha, device=device)
       
        return A + kl_div

set annealing_coef as 0.1 or lower setting will work, do not set 1, it's too large.

Hi, thank you for your answer. I try to set 'annealing_coef' as 0.1 and 0.05 respectively, but it still not works. Do you let it works successfully?

I haven't tried this repo, but I have tried to train an 8631 ID face recognition network, we use resnet-100 is ok, can get the same accuracy on the in-distribution dataset as the softmax training method, resnet-50 can't get to convergence. Another thing is that we all find the KL loss will damage the accuracy, try to decrease the coefficients or just remove it.

Thank you so much. It still don't works. I think I may need to fine-tune other hyperparameters.

from pytorch-classification-uncertainty.

I use the edl loss to train in mini-imagenet dataset with 64 classes, but the loss can't converge and the accuracy is very low.

I think you can modify the parameters:

def edl_loss(self, func, y, alpha, annealing_step, device="cuda"):
        y = self.one_hot_embedding(y)
        y = y.to(device)
        alpha = alpha.to(device)
        S = torch.sum(alpha, dim=1, keepdim=True)
        
        A = torch.sum(y * (func(S) - func(alpha)), dim=1, keepdim=True)

        # annealing_coef = torch.min(
        #     torch.tensor(1.0, dtype=torch.float32),
        #     torch.tensor(self.epoch / annealing_step, dtype=torch.float32),
        # )
        annealing_coef = 0.1

        kl_alpha = (alpha - 1) * (1 - y) + 1
        kl_div = annealing_coef * self.kl_divergence(kl_alpha, device=device)
       
        return A + kl_div

set annealing_coef as 0.1 or lower setting will work, do not set 1, it's too large.

Hi, thank you for your answer. I try to set 'annealing_coef' as 0.1 and 0.05 respectively, but it still not works. Do you let it works successfully?

I haven't tried this repo, but I have tried to train an 8631 ID face recognition network, we use resnet-100 is ok, can get the same accuracy on the in-distribution dataset as the softmax training method, resnet-50 can't get to convergence. Another thing is that we all find the KL loss will damage the accuracy, try to decrease the coefficients or just remove it.

Thank you so much. It still don't works. I think I may need to fine-tune other hyperparameters.

maybe you can refer this https://github.com/RuoyuChen10/FaceTechnologyTool/blob/master/FaceRecognition/evidential_learning.py, I have tried this on Face Recognition. I'm also failed before. I conclude it's mainly about:

1. remove KL loss
2. Learning Rate is important.
3. The depth of the network.

Maybe the learning rate and depth of the network has few influences on softmax and Cross-Entropy Loss training method.

from pytorch-classification-uncertainty.