is there any way to get confidence score for prediction

[AnhMinhTran]

Question as per title, with model trained through timm, is there any way to check for confidence score where top 5 add up to 1.0?

[amaarora]

If I understand your question correctly, I think if you're doing classification, all you pretty much need to do is add a nn.Softmax() layer on top of the a trained model.

So consider you do something like:

import torch 
import timm 
import torch.nn as nn


class ProbabilityModel(nn.Module):
    def __init__(self, model_name, **kwargs):
        super().__init__()
        self.model = timm.create_model(model_name, **kwargs)
        self.softmax = nn.Softmax()

    def forward(self, x):
         out = self.model(x)
         probabilities = self.softmax(out)
         return probabilities

# random input image
x = torch.randn(1, 3, 224, 224)
m = ProbabilityModel('resnet34', pretrained=True, num_classes=5)
m(x)

>> torch.tensor([[0.1633, 0.1588, 0.2028, 0.2260, 0.2491]], grad_fn=<SoftmaxBackward>)

[amaarora]

Or rather you could just do,

model = get_trained_model_with_weights()

out = model(input)
probabilities = nn.Softmax()(out)