Spaces:
Running on Zero
Running on Zero
| import torch | |
| class Normalization(): | |
| """ | |
| This class is for normalizing the spectrograms batch by batch. | |
| The normalization used is min-max, two modes 'framewise' and 'imagewise' can be selected. | |
| In this paper, we found that 'imagewise' normalization works better than 'framewise' | |
| If framewise is used, then X must follow the shape of (B, F, T) | |
| """ | |
| def __init__(self, min, max, mode='imagewise'): | |
| if mode == 'framewise': | |
| def normalize(x): | |
| size = x.shape | |
| x_max = x.max(1, keepdim=True)[0] # Finding max values for each frame | |
| x_min = x.min(1, keepdim=True)[0] | |
| x_std = (x-x_min)/(x_max-x_min) # If there is a column with all zero, nan will occur | |
| x_std[torch.isnan(x_std)]=0 # Making nan to 0 | |
| x_scaled = x_std * (max - min) + min | |
| return x_scaled | |
| elif mode == 'imagewise': | |
| def normalize(x): | |
| # x_max = x.view(size[0], size[1]*size[2]).max(1, keepdim=True)[0] | |
| # x_min = x.view(size[0], size[1]*size[2]).min(1, keepdim=True)[0] | |
| x_max = x.flatten(1).max(1, keepdim=True)[0] | |
| x_min = x.flatten(1).min(1, keepdim=True)[0] | |
| x_max = x_max.unsqueeze(1) # Make it broadcastable | |
| x_min = x_min.unsqueeze(1) # Make it broadcastable | |
| x_std = (x-x_min)/(x_max-x_min) | |
| x_scaled = x_std * (max - min) + min | |
| x_scaled[torch.isnan(x_scaled)]=min # if piano roll is empty, turn them to min | |
| return x_scaled | |
| elif mode == 'rowwise': | |
| def normalize(x): | |
| x_max = x.max(2)[0].unsqueeze(2) | |
| x_min = x.min(2)[0].unsqueeze(2) | |
| x_std = (x-x_min)/(x_max-x_min) | |
| x_scaled = x_std * (max - min) + min | |
| x_scaled[torch.isnan(x_scaled)]=min | |
| return x_scaled | |
| else: | |
| print(f'please choose the correct mode') | |
| self.normalize = normalize | |
| def __call__(self, x): | |
| return self.normalize(x) | |