-
In usual wav2vec training, it says max audio length is 150000, which creates **935** discrete representations. I am still not clear what is the maximum sequence length of the transformer architectures…
-
Hi! Thanks for your great work.
I wonder about the effect of discrete textural features. It seems that these discrete textural features do not enroll in the image generation part. Are the benefits…
-
I am currently on the second stage to train the pixelsnail model. Do I train the top and bottom models separately, or do I first train the top first and use that checkpoint to train the bottom, since …
-
- News
- [MS의 530B Turing-NLG](https://www.microsoft.com/en-us/research/blog/using-deepspeed-and-megatron-to-train-megatron-turing-nlg-530b-the-worlds-largest-and-most-powerful-generative-language-…
-
is this normal?
df720 updated
2 years ago
-
## 🐛 Bug
practically, I'm doing the following:
```srun python train.py --num-nodes 2 --gpus 4```
Which results in the following hanging state, where multiple processes are assigned the same global …
-
Hi, thank you for your great work!
When training the vqvae network, the training time is very long, could you tell me how to accelerate? Thanks for replying!
-
Hi Phil,
Wanted to bring a bit more rigor and testing to your great work. Ran a few quick experiments with a simple dataset as mentioned in another thread, and noticed a large disparity in results …
-
想請問一下關於此篇論文數學的部分:
1. 加入文字之後的 ELBO 有更詳細的推導過程嗎? 是不是單純把只有 image 的 ELBO 不等式兩邊各加上一個 NLL loss for text 而已?
因為我看起來 text 並沒有在 VQVAE 訓練過程中起到作用,而這個 ELBO 是給 VQVAE 的,不太懂為什麼會有 text 那項 loss
2. 不太了解 式(2) 如何變成 式…
-
Hello,
Thank you for your great work on the BEIT paper. I had a quick question. In the paper, you say that you tokenize the image from 224px to 14x14 image patches using the DALL-E VQ-VAE. However,…