# Main code for Confidence and Dispersity Speak: Characterising Prediction Matrix for Unsupervised Accuracy Estimation (ICML'23)

# MIT license

# Please use TIMM (https://github.com/fastai/timmdocs/tree/master/) to test models on test datasets

# transform model's output logits to softmax score matrix (Temp T is a hyper-parameter)
# logits size: N x C, where N is number of test samples and C is the number of classes
T = 0.4
softmax_outputs = F.softmax(logits/T, dim=1)

# nuclear norm of softmax score matrix
raw_nuclear = torch.norm(softmax_outputs, p='nuc')

# upper bound of nuclear norm
n_norm = np.sqrt(softmax_outputs[0] * np.min([softmax_outputs[0], softmax_outputs[1]]))

# final score
nuclear_score = raw_nuclear / n_norm

# All Done! Enjoy!