Paddle可以說是各種坑，但支持國產，含淚試用了百度飛漿的Speech。

1. 坑點

Docker develop-gpu-cuda10.2-cudnn7-latest 缺失:
1. libsndfile1-dev
2. CUDA_VISIBLE_DEVICES

2. 安裝教程

也沒什麼安裝教程。下載docker鏡像和項目源碼。

docker pull paddlecloud/paddlespeech:develop-gpu-cuda10.2-cudnn7-latest

　　docker鏡像

git clone https://github.com/PaddlePaddle/PaddleSpeech

　　項目源碼

3. 測試顯示卡

nvidia-docker run --runtime=nvidia -it -v ~/PycharmProjects/PaddleSpeech:/home paddlecloud/paddlespeech:develop-gpu-cuda10.2-cudnn7-latest /bin/bash

　　進入Docker鏡像

nvidia-smi

nvcc -V

./test_cuda

4. Bug

cd demos/style_fs2 && ./run.sh

　　出現error

4.1 缺失 libsndfile1-dev

Traceback (most recent call last):
  File "style_syn.py", line 22, in <module>
    import soundfile as sf
  File "/usr/local/python3.7.0/lib/python3.7/site-packages/soundfile.py", line 192, in <module>
    _snd = _ffi.dlopen(_explicit_libname)
OSError: cannot load library 'libsndfile.so': libsndfile.so: cannot open shared object file: No such file or directory

　　我猜測，這是CUDA10.2官方docker的bug，里面缺少了libsndfile1-dev。

apt install libsndfile1-dev

　　成功解決

4.2 缺失 環境參數CUDA_VISIBLE_DEVICES

Traceback (most recent call last):
  File "style_syn.py", line 228, in <module>
    main()
  File "style_syn.py", line 209, in main
    paddle.set_device("gpu")
  File "/usr/local/python3.7.0/lib/python3.7/site-packages/paddle/device/__init__.py", line 313, in set_device
    place = _convert_to_place(device)
  File "/usr/local/python3.7.0/lib/python3.7/site-packages/paddle/device/__init__.py", line 204, in _convert_to_place
    place = core.CUDAPlace(ParallelEnv().dev_id)
OSError: (External) CUDA error(100), no CUDA-capable device is detected.
  [Hint: 'cudaErrorNoDevice'. This indicates that no CUDA-capable devices were detected by the installed CUDA driver. ] (at /paddle/paddle/phi/backends/gpu/cuda/cuda_info.cc:66)

　　因缺失 環境參數CUDA_VISIBLE_DEVICES，導致PaddleSpeech失敗獲取GPU。

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

　　當設置CUDA_VISIBLE_DEVICES=0後，成功解決。

5.0 成功