[西安电子科技大学][高校贡献][Pytorch迁移1.8][Baseline_RawNet2]-初次提交-ModelZoo-PyTorch-AtomGit

MPyTorch/contrib/audio/baseline-rawnet/README.md+192-105

	@@ -1,130 +1,217 @@
1	-# RawNet2	1	+# Baseline-Rawnet for PyTorch
2		2
3	-This implements training of RawNet2 on the VoxCeleb1&2 datasets of YouTube.	3	+- [概述](概述.md)
		4	+- [准备训练环境](准备训练环境.md)
		5	+- [开始训练](开始训练.md)
		6	+- [训练结果](训练结果.md)
		7	+- [版本说明](版本说明.md)
4		8
5	-- Reference implementation:	9	+# 概述
6		10
7	-```	11	+## 简述
8	-url=https://github.com/Jungjee/RawNet
9	-dir=RawNet-master/python/RawNet2
10	-```
11		12
12	-## Baseline-RawNet2 Detail	13	+最近，使用深度神经网络对原始波形进行直接建模已被广泛研究用于音频领域的许多任务。然而，在说话人验证中，原始波形的利用处于初步阶段，需要进一步研究。在这项研究中，我们探索了输入原始波形以改进各个方面的端到端深度神经网络：前端说话人嵌入提取，包括模型架构、预训练方案、附加目标函数和后端分类。使用预训练方案调整模型架构可以提取说话人嵌入，从而显着提高性能。
13		14
14	-As of the current date, Ascend-Pytorch is still inefficient for contiguous operations. Therefore, RawNet2 is re-implemented using semantics such as custom OP.	15	+- 参考实现：
15		16
16	-## Requirements	17	+ ```
		18	+ url=https://github.com/Jungjee/RawNet
		19	+ commit_id=585ce2c4fb80ae6ab236f79f06911e2f8bef180c
		20	+ ```
17		21
18	-- Install PyTorch ([pytorch.org](http://pytorch.org))	22	+- 适配昇腾 AI 处理器的实现：
19	-- `pip install -r requirements.txt`
20	- Note: pillow recommends installing a newer version. If the corresponding torchvision version cannot be installed directly, you can use the source code to install the corresponding version. The source code reference link: https://github.com/pytorch/vision，
21	-Suggestion the pillow is 9.1.0 and the torchvision is 0.6.0
22		23
23	-## DataSet	24	+ ```
		25	+ url=https://gitee.com/ascend/ModelZoo-PyTorch.git
		26	+ code_path=PyTorch/contrib/audio
		27	+ ```
24		28
25	-```	29	+- 通过Git获取代码方法如下：
26	-url: http://www.robots.ox.ac.uk/~vgg/data/voxceleb/
27	-```
28		30
29	-- The training datasets are VoxCeleb2, the evaluation dataset is VoxCeleb1H & VoxCeleb1E. The datasets are large. Please ensure sufficient hard disk space when downloading and decompressing.	31	+ ```
30	-- Besides, the data in the VoxCeleb2 downloaded from the url above is in the format of .m4a. If you do not use the dataset which is converted already, you should firstly run the `m4a2wav.py`.	32	+ git clone {url} # 克隆仓库的代码
31	-- You need to follow directory structure of the `data` as below. If you connect to the prepared data folder, you don't need to build the following directory.	33	+ cd {code_path} # 切换到模型代码所在路径，若仓库下只有该模型，则无需切换
		34	+ ```
32		35
33	-```	36	+- 通过单击“立即下载”，下载源码包。
34	-${RawNet}/DB/VoxCeleb1/
35	-├── dev_wav
36	-│ ├── id10001
37	-│ │ ├── 1zcIwhmdeo4
38	-│ │ │ ├── 00001.wav
39	-│ │ │ ├── 00002.wav
40	-│ │ │ └── 00003.wav
41	-│ │ ├── 5ssVY9a5X-M
42	-│ │ │ ├── 00001.wav
43	-│ │ │ ├── 00002.wav
44	-│ │ │ ├── 00003.wav
45	-│ │ │ └── 00003.wav
46	-│ └── ...
47	-├── eval_wav
48	-│ ├── id10270
49	-│ │ ├── 5r0dWxy17C8
50	-│ │ │ ├── 00001.wav
51	-│ │ │ ├── 00002.wav
52	-│ │ │ ├── 00003.wav
53	-│ │ │ ├── 00004.wav
54	-│ │ │ └── 00005.wav
55	-│ └── ...
56	-│ ├── _z_BR0ERa9g
57	-│ ├── 00001.wav
58	-│ ├── 00002.wav
59	-│ └── 00003.wav
60	-├── val_trial.txt
61	-└── veri_test.txt
62		37
63	-${RawNet}/DB/VoxCeleb2/	38	+# 准备训练环境
64	-└── wav
65	- ├── id00012
66	- │ ├── 21Uxsk56VDQ
67	- │ │ ├── 00001.wav
68	- │ │ ├── ...
69	- │ │ └── 00059.wav
70	- │ ├── 00-qODbtozw
71	- │ │ ├── ...
72	- │ │ ├── 00079.wav
73	- │ │ └── 00080.wav
74	- ├── ...
75	- │ └── zw-4DTjqIA0
76	- │ ├── 00108.wav
77	- │ └── 00109.wav
78	- └── id09272
79	- └── u7VNkYraCw0
80	- ├── ...
81	- └── 00027.wav
82	-```
83		39
84	-- You need to follow directory structure of the `output` as below.	40	+## 准备环境
85		41
86	-```	42	+- 当前模型支持的固件与驱动、 CANN 以及 PyTorch 如下表所示。
87	-${RawNet}/train/train_${device_count}P
88	-\|-- DNNS/${name}/
89	-\| \|-- models
90	-\| \| \|--best_opt_eval.pt ## The best perfomance model is saved here
91	-\| \| \|--TA_${epoch}_${eer}.pt ##The other model is saved here
92	-\| \|-- results
93	-\| \|-- log
94	-\| \| \|-- eval_epoch${epoch}.txt ## The training log is saved here
95	-\| \|-- prof
96	-\| \|-- eers.txt ##The eers is saved here
97	-\| \|-- f_params.txt ##The params of the model are saved here
98	-```
99		43
100	-## Training #	44	+ 表 1 版本配套表
101		45
102	-- Note that the `output` folder under the `test` directory will also save the code running log.	46	+ \| 配套 \| 版本 \|
103	-- To run the model, you should cd to the directory of test	47	+ \| ---------- \| ------------------------------------------------------------ \|
104	-- To train a model, run `train_1p.py` or `train_8p.py`:	48	+ \| 硬件 \| [1.0.16](https://www.hiascend.com/hardware/firmware-drivers?tag=commercial) \|
		49	+ \| 固件与驱动 \| [5.1.RC2](https://www.hiascend.com/hardware/firmware-drivers?tag=commercial) \|
		50	+ \| CANN \| [5.1.RC2](https://www.hiascend.com/software/cann/commercial?version=5.1.RC2) \|
		51	+ \| PyTorch \| [1.8.1](https://gitee.com/ascend/pytorch/tree/master/) \|
105		52
106	-```bash	53	+- 环境准备指导。
107	-# 1p train perf
108	-nohup bash train_performance_1p.sh --data_path=xxx &
109		54
110	-# 8p train perf	55	+ 请参考《[Pytorch框架训练环境准备](https://www.hiascend.com/document/detail/zh/ModelZoo/pytorchframework/ptes)》。
111	-nohup bash train_performance_8p.sh --data_path=xxx &
112		56
113	-# 1p train full	57	+- 安装依赖。
114	-nohup bash train_full_1p.sh --data_path=xxx &
115		58
116	-# 8p train full	59	+ ```
117	-nohup bash train_full_8p.sh --data_path=xxx &	60	+ pip install -r requirements.txt
		61	+ ```
118		62
119	-```	63	+## 准备数据集
120		64
121	-## RawNet2 training result	65	+1. 准备数据集
122		66
123	-\| eer(percentage) \| FPS(aver) \| Npu_nums \| Epochs \| AMP_Type \|	67	+ 请用户自行准备好数据集，包含训练集和验证集两部分，可选用的数据集包括 [VoxCeleb2，VoxCeleb1](http://www.robots.ox.ac.uk/~vgg/data/voxceleb/) 等。在源码包根目录下建立“DB”/文件夹，将准备好的数据集上传至“DB/”文件夹中并解压。数据集很大。下载和解压缩时，请确保有足够的硬盘空间。
124	-\| :-------------------------: \| :-------: \| :------: \| :----: \| :------: \|
125	-\| 0.14 \| 7760 \| 1 \| 1 \| O2 \|
126	-\| 0.038(aver) and 0.035(high) \| 8573 \| 8 \| 20 \| O2 \|
127		68
128	-### Testing	69	+ 解压后评估数据集和训练数据集分别位于“DB/VoxCeleb1”和“DB/VoxCeleb2”文件夹路径下，该目录下每个文件夹代表一个类别，且同一文件夹下的所有图片都有相同的标签。当前提供的训练脚本中，是以VoxCeleb1和VoxCeleb2数据集为例。在使用其他数据集时，修改数据集路径。
129		70
130	-The testing data in the paper is about the VoxCeleb1H and VoxCeleb1E. And here we use the dataset of the VoxCeleb1H, and the target of the eer in the paper is 0.0489.	71	+2. 数据预处理
		72	+
		73	+ 从上面的url下载的VoxCeleb2中的数据格式为.m4a。如果不使用已转换的数据集，则应首先执行数据预处理脚本，将数据格式转为.wav。
		74	+
		75	+ ```
		76	+ python m4a2wav.py
		77	+ ```
		78	+
		79	+ 数据集目录结构参考：
		80	+
		81	+ ```
		82	+ ${RawNet}/DB/VoxCeleb1/
		83	+ ├── dev_wav
		84	+ │ ├── id10001
		85	+ │ │ ├── 1zcIwhmdeo4
		86	+ │ │ │ ├── 00001.wav
		87	+ │ │ │ ├── 00002.wav
		88	+ │ │ │ └── 00003.wav
		89	+ │ │ ├── 5ssVY9a5X-M
		90	+ │ │ │ ├── 00001.wav
		91	+ │ │ │ ├── 00002.wav
		92	+ │ │ │ ├── 00003.wav
		93	+ │ │ │ └── 00003.wav
		94	+ │ └── ...
		95	+ ├── eval_wav
		96	+ │ ├── id10270
		97	+ │ │ ├── 5r0dWxy17C8
		98	+ │ │ │ ├── 00001.wav
		99	+ │ │ │ ├── 00002.wav
		100	+ │ │ │ ├── 00003.wav
		101	+ │ │ │ ├── 00004.wav
		102	+ │ │ │ └── 00005.wav
		103	+ │ └── ...
		104	+ │ ├── _z_BR0ERa9g
		105	+ │ ├── 00001.wav
		106	+ │ ├── 00002.wav
		107	+ │ └── 00003.wav
		108	+ ├── val_trial.txt
		109	+ └── veri_test.txt
		110	+
		111	+ ${RawNet}/DB/VoxCeleb2/
		112	+ └── wav
		113	+ ├── id00012
		114	+ │ ├── 21Uxsk56VDQ
		115	+ │ │ ├── 00001.wav
		116	+ │ │ ├── ...
		117	+ │ │ └── 00059.wav
		118	+ │ ├── 00-qODbtozw
		119	+ │ │ ├── ...
		120	+ │ │ ├── 00079.wav
		121	+ │ │ └── 00080.wav
		122	+ ├── ...
		123	+ │ └── zw-4DTjqIA0
		124	+ │ ├── 00108.wav
		125	+ │ └── 00109.wav
		126	+ └── id09272
		127	+ └── u7VNkYraCw0
		128	+ ├── ...
		129	+ └── 00027.wav
		130	+ ```
		131	+
		132	+# 开始训练
		133	+
		134	+1. 进入运行脚本目录下
		135	+
		136	+ ```
		137	+ cd /${模型文件夹名称}/test
		138	+ ```
		139	+
		140	+2. 运行训练脚本
		141	+
		142	+ 该模型支持单机单卡训练和单机8卡训练。
		143	+
		144	+ * 单机单卡训练
		145	+
		146	+ 启动单卡训练
		147	+
		148	+ ```
		149	+ # 1p train perf
		150	+ bash train_performance_1p.sh --data_path=./DB
		151	+ # 1p train full
		152	+ bash train_full_1p.sh --data_path=./DB
		153	+ ```
		154	+
		155	+ * 参数说明：
		156	+
		157	+ * data_path：填写数据集路径
		158	+
		159	+ * 代码运行日志保存在test路径下的output文件夹
		160	+
		161	+ * 单机多卡训练
		162	+
		163	+ 启动8卡训练
		164	+
		165	+ ```
		166	+ # 8p train perf
		167	+ bash train_performance_8p.sh --data_path=./DB
		168	+ # 8p train full
		169	+ bash train_full_8p.sh --data_path=./DB
		170	+ ```
		171	+
		172	+ * 参数说明：
		173	+
		174	+ * data_path：填写数据集路径
		175	+
		176	+ * 代码运行日志保存在test路径下的output文件夹
		177	+
		178	+ 遵循“output”的目录结构如下：
		179	+
		180	+ ```
		181	+ ${RawNet}/train/train_${device_count}P
		182	+ \|-- DNNS/${name}/
		183	+ \| \|-- models
		184	+ \| \| \|--best_opt_eval.pt ## The best perfomance model is saved here
		185	+ \| \| \|--TA_${epoch}_${eer}.pt ##The other model is saved here
		186	+ \| \|-- results
		187	+ \| \|-- log
		188	+ \| \| \|-- eval_epoch${epoch}.txt ## The training log is saved here
		189	+ \| \|-- prof
		190	+ \| \|-- eers.txt ##The eers is saved here
		191	+ \| \|-- f_params.txt ##The params of the model are saved here
		192	+ ```
		193	+
		194	+# 训练结果展示
		195	+
		196	+表 2 训练结果展示表
		197	+
		198	+\| eer(percentage) \| FPS(aver) \| Npu_nums \| Epochs \| AMP_Type \| Torch \|
		199	+\| :-------------------------: \| :-------: \| :------: \| :----: \| :------: \| ----- \|
		200	+\| 0.14 \| 7760 \| 1 \| 1 \| O2 \| 1.5 \|
		201	+\| - \| 7912 \| 1 \| 1 \| O2 \| 1.8 \|
		202	+\| 0.038(aver) and 0.035(high) \| 8573 \| 8 \| 20 \| O2 \| 1.5 \|
		203	+\| 0.038(aver) and 0.035(high) \| 12575 \| 8 \| 20 \| O2 \| 1.8 \|
		204	+
		205	+
		206	+
		207	+# 版本说明
		208	+
		209	+## 变更
		210	+
		211	+2022.03.18：首次发布
		212	+
		213	+2022.11.24：更新pytorch1.8版本，重新发布。
		214	+
		215	+## 已知问题
		216	+
		217	+无。

MPyTorch/contrib/audio/baseline-rawnet/test/env_npu.sh+6-8

	@@ -1,19 +1,19 @@
1	#!/bin/bash	1	#!/bin/bash
		2	+# 配置CANN相关环境变量
2	CANN_INSTALL_PATH_CONF='/etc/Ascend/ascend_cann_install.info'	3	CANN_INSTALL_PATH_CONF='/etc/Ascend/ascend_cann_install.info'
3		4
4	if [ -f $CANN_INSTALL_PATH_CONF ]; then	5	if [ -f $CANN_INSTALL_PATH_CONF ]; then
5	- CANN_INSTALL_PATH=$(cat $CANN_INSTALL_PATH_CONF \| grep Install_Path \| cut -d "=" -f 2)	6	+ CANN_INSTALL_PATH=$(cat $CANN_INSTALL_PATH_CONF \| grep Install_Path \| cut -d "=" -f 2)
6	else	7	else
7	- CANN_INSTALL_PATH="/usr/local/Ascend"	8	+ CANN_INSTALL_PATH="/usr/local/Ascend/"
8	fi	9	fi
9		10
10	-if [ -d ${CANN_INSTALL_PATH}/ascend-toolkit/latest ]; then	11	+if [ -d ${CANN_INSTALL_PATH}/ascend-toolkit/latest ];then
11	- source ${CANN_INSTALL_PATH}/ascend-toolkit/set_env.sh	12	+ source ${CANN_INSTALL_PATH}/ascend-toolkit/set_env.sh
12	else	13	else
13	- source ${CANN_INSTALL_PATH}/nnae/set_env.sh	14	+ source ${CANN_INSTALL_PATH}/nnae/set_env.sh
14	fi	15	fi
15		16
16	-#设置device侧日志登记为error
17	msnpureport -g error -d 0	17	msnpureport -g error -d 0
18	msnpureport -g error -d 1	18	msnpureport -g error -d 1
19	msnpureport -g error -d 2	19	msnpureport -g error -d 2
	@@ -23,8 +23,6 @@ msnpureport -g error -d 5
23	msnpureport -g error -d 6	23	msnpureport -g error -d 6
24	msnpureport -g error -d 7	24	msnpureport -g error -d 7
25		25
26	-export DUMP_GRAPH_LEVEL=0
27	-export DUMP_GE_GRAPH=0
28		26
29	#将Host日志输出到串口,0-关闭/1-开启	27	#将Host日志输出到串口,0-关闭/1-开启
30	export ASCEND_SLOG_PRINT_TO_STDOUT=0	28	export ASCEND_SLOG_PRINT_TO_STDOUT=0

MPyTorch/contrib/audio/baseline-rawnet/test/train_full_1p.sh+7-7

	@@ -12,7 +12,7 @@ export RANK_SIZE=1
12	data_path=""	12	data_path=""
13		13
14	# 训练epoch	14	# 训练epoch
15	-train_epochs=80	15	+train_epochs=20
16	# 指定训练所使用的npu device卡id	16	# 指定训练所使用的npu device卡id
17	device_id=1	17	device_id=1
18	# 加载数据进程数	18	# 加载数据进程数
	@@ -80,7 +80,7 @@ source test/env_npu.sh
80	rm -f nohup.out	80	rm -f nohup.out
81	cd train/train-1P	81	cd train/train-1P
82		82
83	-nohup python train_RawNet2.py > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &	83	+nohup python train_RawNet2.py -epoch ${train_epochs} > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
84		84
85	wait	85	wait
86		86
	@@ -93,13 +93,14 @@ e2e_time=$(( $end_time - $start_time ))
93	#结果打印，不需要修改	93	#结果打印，不需要修改
94	echo "------------------ Final result ------------------"	94	echo "------------------ Final result ------------------"
95	#输出性能FPS，需要模型审视修改	95	#输出性能FPS，需要模型审视修改
96	-FPS=`grep -a 'FPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'`	96	+FPS=`grep -a 'INFOFPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'\|awk '{sum += $1} END {printf "%3.3f",sum/NR}'`
97	#打印，不需要修改	97	#打印，不需要修改
98	echo "Final Performance images/sec : $FPS"	98	echo "Final Performance images/sec : $FPS"
99		99
100	#输出训练精度,需要模型审视修改	100	#输出训练精度,需要模型审视修改
101	-eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:" '{print $NF}'`	101	+eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:\|/epoch" '{print $2}'\|awk 'END{print}'`
102	#打印，不需要修改	102	#打印，不需要修改
		103	+train_accuracy=${eer}
103	echo "Final Train Accuracy : ${eer}"	104	echo "Final Train Accuracy : ${eer}"
104	echo "E2E Training Duration sec : $e2e_time"	105	echo "E2E Training Duration sec : $e2e_time"
105		106
	@@ -116,10 +117,9 @@ ActualFPS=${FPS}
116	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`	117	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
117		118
118	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视	119	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
119	-awk -F "INFOEpoch:" '{print $2}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log >> /opt/npu/Vox_Data/test/output/$ASCEND_DEVICE_ID/train_loss_temp.txt	120	+cat ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log \| grep "INFOEpoch" \| awk -F "INFOEpoch:" '{print $2}' \| awk -F "loss =" '{print $2}' \| awk -F "]" '{print $1}' \| sed 's/\s*//g' > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${CaseName}_loss.txt
120	-awk '!/^$/' $ASCEND_DEVICE_ID/output/$ASCEND_DEVICE_ID/train_loss_temp.txt >> ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
121	#最后一个迭代loss值，不需要修改	121	#最后一个迭代loss值，不需要修改
122	-#ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`	122	+ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
123		123
124	#关键信息打印到${CaseName}.log中，不需要修改	124	#关键信息打印到${CaseName}.log中，不需要修改
125	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log	125	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log

MPyTorch/contrib/audio/baseline-rawnet/test/train_full_8p.sh+7-8

	@@ -12,7 +12,7 @@ export RANK_SIZE=8
12	data_path=""	12	data_path=""
13		13
14	# 训练epoch	14	# 训练epoch
15	-train_epochs=80	15	+train_epochs=20
16	# 指定训练所使用的npu device卡id	16	# 指定训练所使用的npu device卡id
17	device_id=0,1,2,3,4,5,6,7	17	device_id=0,1,2,3,4,5,6,7
18	# 加载数据进程数	18	# 加载数据进程数
	@@ -80,7 +80,7 @@ source test/env_npu.sh
80	rm -f nohup.out	80	rm -f nohup.out
81	cd train/train-8P	81	cd train/train-8P
82		82
83	-nohup python3.7 -m torch.distributed.launch --nproc_per_node=8 train_RawNet2.py > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &	83	+nohup python3 -m torch.distributed.launch --nproc_per_node=8 train_RawNet2.py -epoch ${train_epochs} > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
		haiyan82022年8月1日是否python3.7更合适
84		84
85	wait	85	wait
86		86
	@@ -93,14 +93,15 @@ e2e_time=$(( $end_time - $start_time ))
93	#结果打印，不需要修改	93	#结果打印，不需要修改
94	echo "------------------ Final result ------------------"	94	echo "------------------ Final result ------------------"
95	#输出性能FPS，需要模型审视修改	95	#输出性能FPS，需要模型审视修改
96	-FPS=`grep -a 'FPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'`	96	+FPS=`grep -a 'INFOFPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'\|awk '{sum += $1} END {printf "%3.3f",sum/NR}'`
97	#打印，不需要修改	97	#打印，不需要修改
98	echo "Final Performance images/sec : $FPS"	98	echo "Final Performance images/sec : $FPS"
99		99
100	#输出训练精度,需要模型审视修改	100	#输出训练精度,需要模型审视修改
101	-eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:" '{print $NF}'`	101	+eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:\|/epoch" '{print $2}'\|awk 'END{print}'`
102	train_accuracy=${eer}	102	train_accuracy=${eer}
103	#打印，不需要修改	103	#打印，不需要修改
		104	+train_accuracy=${eer}
104	echo "Final Train Accuracy : ${eer}"	105	echo "Final Train Accuracy : ${eer}"
105	echo "E2E Training Duration sec : $e2e_time"	106	echo "E2E Training Duration sec : $e2e_time"
106		107
	@@ -117,11 +118,9 @@ ActualFPS=${FPS}
117	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`	118	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
118		119
119	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视	120	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
120	-awk -F "INFOEpoch:" '{print $2}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log >> ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss_temp.txt	121	+cat ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log \| grep "INFOEpoch" \| awk -F "INFOEpoch:" '{print $2}' \| awk -F "loss =" '{print $2}' \| awk -F "]" '{print $1}' \| sed 's/\s*//g' > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${CaseName}_loss.txt
121	-awk '!/^$/' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss_temp.txt >> ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
122	-
123	#最后一个迭代loss值，不需要修改	122	#最后一个迭代loss值，不需要修改
124	-#ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`	123	+ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
125		124
126	#关键信息打印到${CaseName}.log中，不需要修改	125	#关键信息打印到${CaseName}.log中，不需要修改
127	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log	126	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log

MPyTorch/contrib/audio/baseline-rawnet/test/train_performance_1p.sh+9-8

	@@ -12,7 +12,8 @@ export RANK_SIZE=1
12	data_path=""	12	data_path=""
13		13
14	# 训练epoch	14	# 训练epoch
15	-train_epochs=80	15	+
		16	+train_epochs=1
16	# 指定训练所使用的npu device卡id	17	# 指定训练所使用的npu device卡id
17	device_id=1	18	device_id=1
18	# 加载数据进程数	19	# 加载数据进程数
	@@ -80,7 +81,7 @@ source test/env_npu.sh
80	rm -f nohup.out	81	rm -f nohup.out
81	cd train/train-1P	82	cd train/train-1P
82		83
		haiyan82022年8月1日是否python3.7更合适
83	-nohup python3.7 train_RawNet2.py > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &	84	+nohup python3 train_RawNet2.py -epoch ${train_epochs} > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
84		85
85	wait	86	wait
86		87
	@@ -93,13 +94,14 @@ e2e_time=$(( $end_time - $start_time ))
93	#结果打印，不需要修改	94	#结果打印，不需要修改
94	echo "------------------ Final result ------------------"	95	echo "------------------ Final result ------------------"
95	#输出性能FPS，需要模型审视修改	96	#输出性能FPS，需要模型审视修改
96	-FPS=`grep -a 'FPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'`	97	+FPS=`grep -a 'INFOFPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'\|awk '{sum += $1} END {printf "%3.3f",sum/NR}'`
97	#打印，不需要修改	98	#打印，不需要修改
98	echo "Final Performance images/sec : $FPS"	99	echo "Final Performance images/sec : $FPS"
99		100
100	#输出训练精度,需要模型审视修改	101	#输出训练精度,需要模型审视修改
101	-eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:" '{print $NF}'`	102	+eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:\|/epoch" '{print $2}'\|awk 'END{print}'`
102	#打印，不需要修改	103	#打印，不需要修改
		104	+train_accuracy=${eer}
103	echo "Final Train Accuracy : ${eer}"	105	echo "Final Train Accuracy : ${eer}"
104	echo "E2E Training Duration sec : $e2e_time"	106	echo "E2E Training Duration sec : $e2e_time"
105		107
	@@ -107,7 +109,7 @@ echo "E2E Training Duration sec : $e2e_time"
107	#训练用例信息，不需要修改	109	#训练用例信息，不需要修改
108	BatchSize=${batch_size}	110	BatchSize=${batch_size}
109	DeviceType=`uname -m`	111	DeviceType=`uname -m`
110	-CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'acc'	112	+CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'perf'
111		113
112	##获取性能数据，不需要修改	114	##获取性能数据，不需要修改
113	#吞吐量	115	#吞吐量
	@@ -116,10 +118,9 @@ ActualFPS=${FPS}
116	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`	118	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
117		119
118	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视	120	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
119	-awk -F "INFOEpoch:" '{print $2}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log >> /opt/npu/Vox_Data/test/output/$ASCEND_DEVICE_ID/train_loss_temp.txt	121	+cat ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log \| grep "INFOEpoch" \| awk -F "INFOEpoch:" '{print $2}' \| awk -F "loss =" '{print $2}' \| awk -F "]" '{print $1}' \| sed 's/\s*//g' > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${CaseName}_loss.txt
120	-awk '!/^$/' $ASCEND_DEVICE_ID/output/$ASCEND_DEVICE_ID/train_loss_temp.txt >> ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
121	#最后一个迭代loss值，不需要修改	122	#最后一个迭代loss值，不需要修改
122	-#ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`	123	+ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
123		124
124	#关键信息打印到${CaseName}.log中，不需要修改	125	#关键信息打印到${CaseName}.log中，不需要修改
125	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log	126	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log

MPyTorch/contrib/audio/baseline-rawnet/test/train_performance_8p.sh+8-8

	@@ -12,7 +12,7 @@ export RANK_SIZE=8
12	data_path=""	12	data_path=""
13		13
14	# 训练epoch	14	# 训练epoch
15	-train_epochs=80	15	+train_epochs=1
16	# 指定训练所使用的npu device卡id	16	# 指定训练所使用的npu device卡id
17	device_id=0,1,2,3,4,5,6,7	17	device_id=0,1,2,3,4,5,6,7
18	# 加载数据进程数	18	# 加载数据进程数
	@@ -80,7 +80,7 @@ source test/env_npu.sh
80	rm -f nohup.out	80	rm -f nohup.out
81	cd train/train-8P	81	cd train/train-8P
82		82
83	-nohup python3.7 train_RawNet2.py ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &	83	+nohup python3 -m torch.distributed.launch --nproc_per_node=8 train_RawNet2.py -epoch ${train_epochs} > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
84		84
85	wait	85	wait
86		86
	@@ -92,13 +92,14 @@ e2e_time=$(( $end_time - $start_time ))
92	#结果打印，不需要修改	92	#结果打印，不需要修改
93	echo "------------------ Final result ------------------"	93	echo "------------------ Final result ------------------"
94	#输出性能FPS，需要模型审视修改	94	#输出性能FPS，需要模型审视修改
95	-FPS=`grep -a 'FPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'`	95	+FPS=`grep -a 'INFOFPS' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F 'INFOFPS:\|/epoch' '{print $2}'\|awk '{sum += $1} END {printf "%3.3f",sum/NR}'`
96	#打印，不需要修改	96	#打印，不需要修改
97	echo "Final Performance images/sec : $FPS"	97	echo "Final Performance images/sec : $FPS"
98		98
99	#输出训练精度,需要模型审视修改	99	#输出训练精度,需要模型审视修改
100	-eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:" '{print $NF}'`	100	+eer=`grep -a 'eer' ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F "eer:\|/epoch" '{print $2}'\|awk 'END{print}'`
101	#打印，不需要修改	101	#打印，不需要修改
		102	+train_accuracy=${eer}
102	echo "Final Train Accuracy : ${eer}"	103	echo "Final Train Accuracy : ${eer}"
103	echo "E2E Training Duration sec : $e2e_time"	104	echo "E2E Training Duration sec : $e2e_time"
104		105
	@@ -106,7 +107,7 @@ echo "E2E Training Duration sec : $e2e_time"
106	#训练用例信息，不需要修改	107	#训练用例信息，不需要修改
107	BatchSize=${batch_size}	108	BatchSize=${batch_size}
108	DeviceType=`uname -m`	109	DeviceType=`uname -m`
109	-CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'acc'	110	+CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'perf'
110		111
111	##获取性能数据，不需要修改	112	##获取性能数据，不需要修改
112	#吞吐量	113	#吞吐量
	@@ -115,10 +116,9 @@ ActualFPS=${FPS}
115	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`	116	TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
116		117
117	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视	118	#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
118	-awk -F "INFOEpoch:" '{print $2}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log >> /opt/npu/Vox_Data/test/output/$ASCEND_DEVICE_ID/train_loss_temp.txt	119	+cat ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log \| grep "INFOEpoch" \| awk -F "INFOEpoch:" '{print $2}' \| awk -F "loss =" '{print $2}' \| awk -F "]" '{print $1}' \| sed 's/\s*//g' > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${CaseName}_loss.txt
119	-awk '!/^$/' $ASCEND_DEVICE_ID/output/$ASCEND_DEVICE_ID/train_loss_temp.txt >> ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
120	#最后一个迭代loss值，不需要修改	120	#最后一个迭代loss值，不需要修改
121	-#ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`	121	+ActualLoss=`awk 'END {print}' ${test_path_dir}/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
122		122
123	#关键信息打印到${CaseName}.log中，不需要修改	123	#关键信息打印到${CaseName}.log中，不需要修改
124	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log	124	echo "Network = ${Network}" > ${test_path_dir}/output/$ASCEND_DEVICE_ID/${CaseName}.log

MPyTorch/contrib/audio/baseline-rawnet/train/train-1P/parser1.py+1-1

	@@ -54,7 +54,7 @@ def get_args():
54	#hyper-params	54	#hyper-params
55	parser.add_argument('-frame', type = int, default = 135395880960) # the number of frames of all the .wav file	55	parser.add_argument('-frame', type = int, default = 135395880960) # the number of frames of all the .wav file
56	parser.add_argument('-bs', type = int, default = 128)	56	parser.add_argument('-bs', type = int, default = 128)
57	- parser.add_argument('-lr', type = float, default = 0.001)	57	+ parser.add_argument('-lr', type = float, default = 0.0015)
58	parser.add_argument('-nb_samp', type = int, default = 59049)	58	parser.add_argument('-nb_samp', type = int, default = 59049)
59	parser.add_argument('-window_size', type = int, default = 11810)	59	parser.add_argument('-window_size', type = int, default = 11810)
60		60

MPyTorch/contrib/audio/baseline-rawnet/train/train-1P/train_RawNet2.py+2-0

	@@ -28,6 +28,8 @@ import json
28	import numpy as np	28	import numpy as np
29		29
30	import torch	30	import torch
		31	+if torch.__version__ >="1.8":
		32	+ import torch_npu
31	import torch.nn as nn	33	import torch.nn as nn
32	import torch.distributed as dist	34	import torch.distributed as dist
33	from torch.nn.parallel import DistributedDataParallel as DDP	35	from torch.nn.parallel import DistributedDataParallel as DDP

MPyTorch/contrib/audio/baseline-rawnet/train/train-1P/trainer.py+3-0

	@@ -121,6 +121,9 @@ def train_model(model, db_gen, optimizer, epoch, args, device, lr_scheduler, cri
121		121
122		122
123	def time_augmented_evaluate_model(mode, model, db_gen, l_utt, save_dir, epoch, l_trial, args, device):	123	def time_augmented_evaluate_model(mode, model, db_gen, l_utt, save_dir, epoch, l_trial, args, device):
		124	+ with open(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch), mode = 'w') as ff:
		125	+ logger = get_logger(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch))
		126	+
124	f_log = open(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch), 'a', buffering = 1)	127	f_log = open(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch), 'a', buffering = 1)
125	args = get_args()	128	args = get_args()
126	if mode not in ['val','eval']: raise ValueError('mode should be either "val" or "eval"')	129	if mode not in ['val','eval']: raise ValueError('mode should be either "val" or "eval"')

MPyTorch/contrib/audio/baseline-rawnet/train/train-8P/parser1.py+1-1

	@@ -54,7 +54,7 @@ def get_args():
54	#hyper-params	54	#hyper-params
55	parser.add_argument('-frame', type = int, default = 135395880960) # the number of frames of all the .wav file	55	parser.add_argument('-frame', type = int, default = 135395880960) # the number of frames of all the .wav file
56	parser.add_argument('-bs', type = int, default = 1024)	56	parser.add_argument('-bs', type = int, default = 1024)
57	- parser.add_argument('-lr', type = float, default = 0.001)	57	+ parser.add_argument('-lr', type = float, default = 0.0015)
58	parser.add_argument('-nb_samp', type = int, default = 59049)	58	parser.add_argument('-nb_samp', type = int, default = 59049)
59	parser.add_argument('-window_size', type = int, default = 11810)	59	parser.add_argument('-window_size', type = int, default = 11810)
60		60

MPyTorch/contrib/audio/baseline-rawnet/train/train-8P/train_RawNet2.py+2-0

	@@ -28,6 +28,8 @@ import json
28	import numpy as np	28	import numpy as np
29		29
30	import torch	30	import torch
		31	+if torch.__version__ >="1.8":
		32	+ import torch_npu
31	import torch.nn as nn	33	import torch.nn as nn
32	import torch.distributed as dist	34	import torch.distributed as dist
33	from torch.nn.parallel import DistributedDataParallel as DDP	35	from torch.nn.parallel import DistributedDataParallel as DDP

MPyTorch/contrib/audio/baseline-rawnet/train/train-8P/trainer.py+3-0

	@@ -106,6 +106,9 @@ def train_model(model, db_gen, optimizer, epoch, args, device, lr_scheduler, cri
106		106
107		107
108	def time_augmented_evaluate_model(mode, model, db_gen, l_utt, save_dir, epoch, l_trial, args, device):	108	def time_augmented_evaluate_model(mode, model, db_gen, l_utt, save_dir, epoch, l_trial, args, device):
		109	+ with open(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch), mode = 'w') as ff:
		110	+ logger = get_logger(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch))
		111	+
109	f_log = open(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch), 'a', buffering = 1)	112	f_log = open(args.save_dir + args.name + '/' + args.save_log + 'TA_{}.log'.format(epoch), 'a', buffering = 1)
110	args = get_args()	113	args = get_args()
111	if mode not in ['val','eval']: raise ValueError('mode should be either "val" or "eval"')	114	if mode not in ['val','eval']: raise ValueError('mode should be either "val" or "eval"')