aishell的deploy的问题

[yyhlvdl]

我直接使用你们发布的aishell模型，执行python deploy/demo_server.py，然后出现了错误：

root@095d9ada1b1d:/DeepSpeech# python deploy/demo_server.py
-----------  Configuration Arguments -----------
alpha: 2.15
beam_size: 500
beta: 0.35
cutoff_prob: 1.0
cutoff_top_n: 40
decoding_method: ctc_beam_search
host_ip: localhost
host_port: 8086
lang_model_path: models/lm/zh_giga.no_cna_cmn.prune01244.klm
mean_std_path: asset/preprocess/mean_std.npz
model_path: asset/train/params.tar.gz
num_conv_layers: 2
num_rnn_layers: 3
rnn_layer_size: 2048
share_rnn_weights: False
specgram_type: linear
speech_save_dir: demo_cache
use_gpu: True
use_gru: True
vocab_path: asset/preprocess/vocab.txt
warmup_manifest: asset/preprocess/test
------------------------------------------------
I1205 10:14:34.175657    15 Util.cpp:166] commandline:  --use_gpu=True --trainer_count=1 
[INFO 2017-12-05 10:14:35,626 layers.py:2606] output for __conv_0__: c = 32, h = 81, w = 54, size = 139968
[INFO 2017-12-05 10:14:35,626 layers.py:3133] output for __batch_norm_0__: c = 32, h = 81, w = 54, size = 139968
[INFO 2017-12-05 10:14:35,627 layers.py:7224] output for __scale_sub_region_0__: c = 32, h = 81, w = 54, size = 139968
[INFO 2017-12-05 10:14:35,627 layers.py:2606] output for __conv_1__: c = 32, h = 41, w = 54, size = 70848
[INFO 2017-12-05 10:14:35,628 layers.py:3133] output for __batch_norm_1__: c = 32, h = 41, w = 54, size = 70848
[INFO 2017-12-05 10:14:35,628 layers.py:7224] output for __scale_sub_region_1__: c = 32, h = 41, w = 54, size = 70848
-----------------------------------------------------------
Warming up ...
('Warm-up Test Case %d: %s', 0, u'asset/data/aishell/wav/test/S0765/BAC009S0765W0205.wav')
[INFO 2017-12-05 10:14:42,337 model.py:230] begin to initialize the external scorer for decoding
[INFO 2017-12-05 10:14:50,941 model.py:241] language model: is_character_based = 1, max_order = 5, dict_size = 0
[INFO 2017-12-05 10:14:50,941 model.py:242] end initializing scorer. Start decoding ...
Traceback (most recent call last):
  File "deploy/demo_server.py", line 224, in <module>
    main()
  File "deploy/demo_server.py", line 220, in main
    start_server()
  File "deploy/demo_server.py", line 204, in start_server
    num_test_cases=3)
  File "deploy/demo_server.py", line 143, in warm_up_test
    (finish_time - start_time, transcript))
UnicodeEncodeError: 'ascii' codec can't encode characters in position 40-94: ordinal not in range(128)

于是，我将transcript注释掉，重新执行，然后可以继续了。只是

[INFO 2017-12-05 10:46:41,054 model.py:230] begin to initialize the external scorer for decoding
[INFO 2017-12-05 10:46:42,193 model.py:241] language model: is_character_based = 1, max_order = 5, dict_size = 0
[INFO 2017-12-05 10:46:42,193 model.py:242] end initializing scorer. Start decoding ...
Response Time: 1174.020508
('Warm-up Test Case %d: %s', 1, u'asset/data/aishell/wav/test/S0767/BAC009S0767W0141.wav')

一个文件就需要1174s，这么长的时间，请问，有办法可以提速吗？

[yyhlvdl]

而且，在按照新的更改，就是简化了数据并行处理的新的代码后，这个时间更是长的夸张，难道，真的是必须要很高的硬件配置吗？我的一个gpu感觉完全看不到尽头

[kuke]

@john81529 这个时间不正常，通常来讲一段语言的识别时间短于1s。你可以先跑一下run_infer_golden.sh这个脚本，看平均处理一条语言的时间是多少。

[yyhlvdl]

嗯。感觉aishell模型的确是不如librisoeech的模型，非常期待cn1.2k模型。很期待可以实现中文的自己语音的识别。像librispeech那样的显得很完整的体系。

[pkuyym]

@john81529 是这样的，你把cutoff_prob设成0.99看看？应该可以提速很多，cn1.2k的cer表现还可以，但是对真实人声的识别效果不太稳定，推测跟training数据的关系较大

[yyhlvdl]

嗯。中文语音数据的确是没有非常好的，至少开源的比较少。

[yyhlvdl]

root@b203ab126266:/DeepSpeech# python deploy/demo_server.py
-----------  Configuration Arguments -----------
alpha: 2.15
beam_size: 500
beta: 0.35
cutoff_prob: 1.0
cutoff_top_n: 40
decoding_method: ctc_beam_search
host_ip: localhost
host_port: 8086
lang_model_path: models/lm/zh_giga.no_cna_cmn.prune01244.klm
mean_std_path: asset/preprocess/mean_std.npz
model_path: asset/train/params.tar.gz
num_conv_layers: 2
num_rnn_layers: 3
rnn_layer_size: 2048
share_rnn_weights: False
specgram_type: linear
speech_save_dir: demo_cache
use_gpu: True
use_gru: True
vocab_path: asset/preprocess/vocab.txt
warmup_manifest: asset/preprocess/test
------------------------------------------------
I1205 11:37:54.343217    15 Util.cpp:166] commandline:  --use_gpu=True --trainer_count=1 
[INFO 2017-12-05 11:37:55,704 layers.py:2606] output for __conv_0__: c = 32, h = 81, w = 54, size = 139968
[INFO 2017-12-05 11:37:55,705 layers.py:3133] output for __batch_norm_0__: c = 32, h = 81, w = 54, size = 139968
[INFO 2017-12-05 11:37:55,705 layers.py:7224] output for __scale_sub_region_0__: c = 32, h = 81, w = 54, size = 139968
[INFO 2017-12-05 11:37:55,706 layers.py:2606] output for __conv_1__: c = 32, h = 41, w = 54, size = 70848
[INFO 2017-12-05 11:37:55,706 layers.py:3133] output for __batch_norm_1__: c = 32, h = 41, w = 54, size = 70848
[INFO 2017-12-05 11:37:55,706 layers.py:7224] output for __scale_sub_region_1__: c = 32, h = 41, w = 54, size = 70848
-----------------------------------------------------------
Warming up ...
('Warm-up Test Case %d: %s', 0, u'asset/data/aishell/wav/test/S0765/BAC009S0765W0205.wav')
[INFO 2017-12-05 11:37:59,694 model.py:230] begin to initialize the external scorer for decoding
[INFO 2017-12-05 11:37:59,770 model.py:241] language model: is_character_based = 1, max_order = 5, dict_size = 0
[INFO 2017-12-05 11:37:59,770 model.py:242] end initializing scorer. Start decoding ...
Response Time: 1542.750508
('Warm-up Test Case %d: %s', 1, u'asset/data/aishell/wav/test/S0767/BAC009S0767W0141.wav')
Response Time: 1605.342387
('Warm-up Test Case %d: %s', 2, u'asset/data/aishell/wav/test/S0908/BAC009S0908W0175.wav')
Response Time: 1449.373175
-----------------------------------------------------------
ASR Server Started.

其实还好，也不是太费时间。

[yyhlvdl]

只是，在客户端出现这样的错误：
python -u deploy/demo_client.py
ALSA lib pcm.c:2266:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.rear
ALSA lib pcm.c:2266:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.center_lfe
ALSA lib pcm.c:2266:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.side
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map
Start Recording ... Traceback (most recent call last):
File "deploy/demo_client.py", line 58, in callback
sock.connect((args.host_ip, args.host_port))
File "/usr/lib/python2.7/socket.py", line 228, in meth
return getattr(self._sock,name)(*args)
socket.error: [Errno 111] Connection refused
我是在命令端打开的这个文件，因为docker没有pynput这个模块。同时，我的服务器和客户端是一台机器。

[pkuyym]

这个看起来是硬件问题或者驱动问题？

ALSA lib pcm.c:2266:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.rear
ALSA lib pcm.c:2266:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.center_lfe
ALSA lib pcm.c:2266:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.side
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map
ALSA lib pcm_route.c:867:(find_matching_chmap) Found no matching channel map

这个话，最好确认，client连server是没问题的？

File "deploy/demo_client.py", line 58, in callback
sock.connect((args.host_ip, args.host_port))
File "/usr/lib/python2.7/socket.py", line 228, in meth
return getattr(self._sock,name)(*args)
socket.error: [Errno 111] Connection refused

[yyhlvdl]

前者还好说，alsa报错应该不影响使用。主要是后者，client链接server好像不成功，我现在使用了127.0.0.1来用，正在等待程序完成。

[yyhlvdl]

依然还是链接不成功。

[yyhlvdl]

我觉得，按照代码，宿主机无法访问docker造成的，请问，你们当时测试的时候，是什么样的情况？我的服务器是在docker中启动，而客户端是在宿主机中启动。

[pkuyym]

这个应该是docker问题了，host怎么跟docker通信，你可以搜索一下看看

[yyhlvdl]

嗯。应该是这个问题，不过，真心希望paddlepaddle可以开发出更舒适的安装模式，虽然docker下运行的确很方便。

[pkuyym]

@john81529 我们本地都setup过ds2的环境，都没有遇到很大的问题

[yyhlvdl]

就是说，你们不是在docker环境中运行的deepspeech,而是用源码编译或者deb安装的paddlepaddle?然而，不得不说，我从官网上下载源码和deb，从来没有成功过，上次您给我的cur链接也是下载不了，所以，我只能选择docker下运行。

[pkuyym]

嗯，是的，paddlepaddle是源码编译或者pip install，deepspeech从源码编译，对平台兼容度还不错

[yyhlvdl]

不是说，pip install的版本比较低吗？而且，最新版的paddle的deb包居然没有，我还是试试源码编译吧。

[pkuyym]

@john81529 pip install的版本并不低，我记得之前给过你链接，从里面下载就可以。

[yyhlvdl]

实不相瞒，我从官网和您给的链接里，各种版本都下载不成功，我先源码编译。

[pkuyym]

@john81529 可能哪里出错了，我们这里测试：源码编译和pip install都工作的很好。

[kuke]

@john81529 你是否是用git clone https://github.com/PaddlePaddle/Paddle.git的方式获取的源码？其它方式都不能保证获取的源码是最新的。

[pkuyym]

@john81529 当然不是，你的错误很明显呀，就是下载mkldnn的包出错了

CMakeFiles/extern_mklml.dir/build.make:88: recipe for target 'third_party/mklml/src/extern_mklml-stamp/extern_mklml-download' failed

可以参考 PaddlePaddle/Paddle#5508

以及，不同的问题可以开新的issue哈

[pkuyym]

Paddle的编译问题，建议你去Paddle下面发issue哈
https://github.com/PaddlePaddle/Paddle/issues

以及，你这么长的log，最好加引用把格式抹掉，否则显示很糟糕

[Pelhans]

@yyhlvdl 请问你是怎么解决docker和主机通信的问题的？ 我的报错和您一样，都是
socket.error: [Errno 111] Connection refused

[yyhlvdl]

@Pelhans 在命令中加这么一句话net=host，

[Pelhans]

@yyhlvdl 谢谢～