facebookresearch--audiocraft
Alexandre Défossez
21 行
630 B
Python
21 行
630 B
Python
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
|
# All rights reserved.
|
|
#
|
|
# This source code is licensed under the license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
|
|
import torch
|
|
|
|
from audiocraft.quantization.vq import ResidualVectorQuantizer
|
|
|
|
|
|
class TestResidualVectorQuantizer:
|
|
|
|
def test_rvq(self):
|
|
x = torch.randn(1, 16, 2048, requires_grad=True)
|
|
vq = ResidualVectorQuantizer(n_q=8, dimension=16, bins=8)
|
|
res = vq(x, 1.)
|
|
assert res.x.shape == torch.Size([1, 16, 2048])
|
|
res.x.sum().backward()
|
|
assert torch.allclose(x.grad.data, torch.ones(1))
|