-
In readme i didn't find any steps for VQ-VAE-encoder + WaveNet decoder .Please can any one help me to proceed further.
-
Hi,
I was wondering how I can obtain the bits / dim in the VQ-VAE notebook. I tried the following:
use MSE loss to calculate the likelihood under a standard gaussian, then add the KL term to obt…
-
thanks for this work! But i still have a question:
When I run the code train_vqvae.py, it will report an error
"AttributeError: module 'torch.nn.functional' has no attribute 'one_hot'"
cdq14 updated
5 years ago
-
I noticed that in train_vqvae.py, there are only two terms for the loss function; however, there appears to be three terms in the paper. I suppose the second term in the code refers to the commitment …
-
Thank you for implementing this in pytorch.
According to the paper (Generating Diverse High-Fidelity Images with VQ-VAE-2), the input to the model is a 256 × 256 image that is compressed to quantized…
-
I did something like
```
model_base_path = str(os.sep).join(paths.model_path().split(os.sep)[:-1])
step_base_path = str(os.sep).join(paths.step_path().split(os.sep)[:-1])
…
-
Hello,
I am successful in building a VCTK model by executing train.py as per your instructions. Thank you.
But I have the following issue while generating:
Python 3.6
Chainer 5.1.0
python gen…
-
Hi,
Do you plan subj?
thanks
-
在图像的风格转移中，通过迭代处理随机初始化图像的像素值，得到最优的风格生成图像，所以处理的关键数据是图像的像素值（也就是3D中的RGB值）；但是在语音的风格转移中，声音信号被转换成2D spectrum，横轴是时间，纵轴是频率，2D图中的每个值是对应时间、对应频率的能量。
所以我想问：
1. 在语音的风格转移中，处理（迭代更新）的关键数据是2D spectru…
-
It takes a real-valued Tensor as input (in VQVAEs, encoder output) and returns the quantized vectors whose Tensor optionally has EMA control dependencies.
+ It adds commitment loss to the layer.
+…