| from vector_quantize_pytorch import FSQ |
| import torch |
| import torch.nn as nn |
| class FSQWrapper(nn.Module): |
| def __init__(self, dim, levels): |
| super().__init__() |
| level_dim = len(levels) |
| self.quantizer = FSQ( |
| levels = levels |
| ) |
| self.down = nn.Linear(dim, level_dim) |
| self.up = nn.Linear(level_dim, dim) |
| def forward(self, x): |
| x = x.transpose(1,2) |
| x = self.down(x) |
| z_q, indices = self.quantizer(x) |
| x_q = self.up(z_q) |
| x_q = x_q.transpose(1,2) |
| loss = torch.tensor(0.0).to(x.device) |
| return x_q, indices, loss |
| |
| def decode(self, indices): |
| z_q = self.quantizer.indices_to_codes(indices) |
| x_q = self.up(z_q) |
| return x_q.transpose(1,2) |
| |
| if __name__ == '__main__': |
| model = FSQWrapper(512, [8,4,4,4]) |
| x = torch.randn(1, 512, 10000) |
| x_q, indices, loss = model(x) |
| x_q = model.decode(indices) |
| print(indices.min(), indices.max()) |
| print(loss) |
