Can you please provide some information about how the supervised model and the untrained encoder are trained in the paper? My assumption is that the supervised model is trained on both pretext group + train group and the encoder model is trained on only the train group. When I did run the supervised.py script I was getting an accuracy of 78.28 and a kappa score of only 0.3871, clearly, the model is performing badly due to fewer training subjects and I was using the same random seed as yours. I have been experimenting on sleepEDF dataset.

And in the paper, why there is a difference in metrics between the table values and graph values. Do they point to the same results?

Is it a good idea to parameterize the weights where you calculate using the dot product between two latent vectors in the contraWR+. If we can how do you think doing it is a good idea. Let me know your thoughts on it, I shall experiment with them.

def noise_channel(ts, mode, degree, bound):
    """
    Add noise to ts
    
    mode: high, low, both
    degree: degree of noise, compared with range of ts    
    
    Input:
        ts: (n_length)
    Output:
        out_ts: (n_length)
        
    """

noise_channel需要一个单通道时间序列,但是调用它的add_noise和remove_noise传给noise_channel是x[i,:]，函数说明中说 x: (n_length, n_channel)，那每次传给noise_channel的x不是变成了某个采样点的全通道信号吗，请问一下作者这里是否是有什么问题

    def add_noise(self, x, ratio):
        """
        Add noise to multiple ts
        Input: 
            x: (n_length, n_channel)
        Output: 
            x: (n_length, n_channel)
        """
        for i in range(self.n_channels):
            if np.random.rand() > ratio:
                mode = np.random.choice(['high', 'low', 'both', 'no'])
                x[i,:] = noise_channel(x[i,:], mode=mode, degree=0.05, bound=self.bound)
        return x
    
    def remove_noise(self, x, ratio):
        """
        Remove noise from multiple ts
        Input: 
            x: (n_length, n_channel)
        Output: 
            x: (n_length, n_channel)
        """
        for i in range(self.n_channels):
            rand = np.random.rand()
            if rand > 0.75:
                x[i, :] = denoise_channel(x[i, :], self.bandpass1, self.signal_freq, bound=self.bound) +\
                        denoise_channel(x[i, :], self.bandpass2, self.signal_freq, bound=self.bound)
            elif rand > 0.5:
                x[i, :] = denoise_channel(x[i, :], self.bandpass1, self.signal_freq, bound=self.bound)
            elif rand > 0.25:
                x[i, :] = denoise_channel(x[i, :], self.bandpass2, self.signal_freq, bound=self.bound)
            else:
                pass

        return x