[东北大学软件学院][高校贡献][PyTorch离线推理][SKNet50]-初次提交-ModelZoo-PyTorch-AtomGit

MACL_PyTorch/contrib/cv/classfication/Sknet50/README.md+95-81

	@@ -1,4 +1,4 @@
1	-# SK-ResNet50 Onnx 模型端到端推理指导	1	+# SK-ResNet50 Onnx 模型端到端推理指导
2		2
3	- [1. 模型概述](#1)	3	- [1. 模型概述](#1)
4	- [论文地址](#11)	4	- [论文地址](#11)
	@@ -6,6 +6,7 @@
6	- [2. 环境说明](#2)	6	- [2. 环境说明](#2)
7	- [深度学习框架](#21)	7	- [深度学习框架](#21)
8	- [python第三方库](#22)	8	- [python第三方库](#22)
		9	+ - [环境导入命令](#23)
9	- [3. 模型转换](#3)	10	- [3. 模型转换](#3)
10	- [pth转onnx模型](#31)	11	- [pth转onnx模型](#31)
11	- [4. 数据预处理](#4)	12	- [4. 数据预处理](#4)
	@@ -20,6 +21,8 @@
20	- [精度对比](#62)	21	- [精度对比](#62)
21	- [7. 性能对比](#7)	22	- [7. 性能对比](#7)
22	- [npu性能数据](#71)	23	- [npu性能数据](#71)
		24	+ - [T4性能数据](#72)
		25	+ - [性能对比](#73)
23		26
24	## <a name="1">1. 模型概述</a>	27	## <a name="1">1. 模型概述</a>
25		28
	@@ -40,17 +43,30 @@ commit_id: 2864528f8b83f52c3df76f7c3804aa468b91e5cf
40	### <a name="21">2.1 深度学习框架</a>	43	### <a name="21">2.1 深度学习框架</a>
41		44
42	```	45	```
43	-pytorch == 1.5.0	46	+pytorch == 1.8.2+cpu
44	-torchvision == 0.6.0	47	+torchvision == 0.9.2+cpu
45	onnx == 1.9.0	48	onnx == 1.9.0
46	```	49	```
47		50
48	### <a name="22">2.2 python第三方库</a>	51	### <a name="22">2.2 python第三方库</a>
49		52
50	```	53	```
51	-numpy == 1.19.2	54	+numpy == 1.21.6
52	-Pillow == 8.2.0	55	+Pillow == 9.1.0
53	-opencv-python == 4.5.2	56	+opencv-python == 4.5.5.64
		57	+sympy == 1.10.1
		58	+decorator == 5.1.1
		59	+
		60	+```
		61	+
		62	+### <a name="23">2.3 环境导入命令</a>
		63	+
		64	+```
		65	+pip install sympy
		66	+pip install decorator
		67	+pip install onnx==1.9.0
		68	+pip install torch==1.8.2+cpu torchvision==0.9.2+cpu torchaudio==0.8.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
		69	+pip install opencv-python
54	```	70	```
55		71
56	> 说明：	72	> 说明：
	@@ -76,14 +92,16 @@ opencv-python == 4.5.2
76	```bash	92	```bash
77	git clone https://github.com/implus/PytorchInsight.git	93	git clone https://github.com/implus/PytorchInsight.git
78	```	94	```
		95	+ 仓库代码克隆到Sknet50文件夹下
		96	+ 仓库代码包含pth权重文件
79		97
80	3. 使用 sknet2onnx.py 转换pth为onnx文件，在命令行运行如下指令：	98	3. 使用 sknet2onnx.py 转换pth为onnx文件，在命令行运行如下指令：
81		99
82	```bash	100	```bash
83	- python3.7 sknet2onnx.py --pth sk_resnet50.pth.tar --onnx sknet50_bs1	101	+ python sknet2onnx.py --pth sk_resnet50.pth.tar --onnx sk_resnet50.onnx
84	```	102	```
85		103
86	- sk_resnet50.pth.tar文件为步骤1中下载的预训练权重文件，该条指令将在运行处生成一个sknet50_bs1文件，此文件即为目标onnx文件	104	+ sk_resnet50.pth.tar文件为步骤1中下载的预训练权重文件，该条指令将在运行处生成一个sknet50.onnx文件，此文件即为目标onnx文件
87		105
88	模型转换要点：	106	模型转换要点：
89		107
	@@ -96,33 +114,30 @@ opencv-python == 4.5.2
96	1. 设置 atc 工作所需要的环境变量	114	1. 设置 atc 工作所需要的环境变量
97		115
98	```bash	116	```bash
99	- export install_path=/usr/local/Ascend/ascend-toolkit/latest	117	+ source set_env.sh
		Liang Zhen2022年5月18日 `/usr/local/Ascend/ascend-toolkit/` 路径下已经存在 `set_env.sh` 文件了，直接source环境变量就好
100	- export PATH=/usr/local/python3.7.5/bin:${install_path}/atc/ccec_compiler/bin:${install_path}/atc/bin:$PATH
101	- export PYTHONPATH=${install_path}/atc/python/site-packages:$PYTHONPATH
102	- export LD_LIBRARY_PATH=${install_path}/atc/lib64:${install_path}/acllib/lib64:$LD_LIBRARY_PATH
103	- export ASCEND_OPP_PATH=${install_path}/opp
104	```	118	```
105		119
106	2. 使用atc工具将onnx模型转换为om模型，命令参考	120	2. 使用atc工具将onnx模型转换为om模型，命令参考
107		121
108	```bash	122	```bash
109	- atc --framework=5 --model=sknet50.onnx --output=sknet50_bs1 --input_format=NCHW --input_shape="image:1,3,224,224" --log=debug --soc_version=Ascend310	123	+ 1batch命令:
		124	+ atc --framework=5 --model=sk_resnet50.onnx --output=sk_resnet50_bs1_310p --input_format=NCHW --input_shape="image:1,3,224,224" --log=debug --soc_version=Ascend710
110	```	125	```
111	-	126	+
112	- 此命令将在运行路径下生成一个sknet50_1bs.om文件，此文件即为目标om模型文件	127	+ 1batch命令将在运行路径下生成一个Sk_resnet50_bs1_310p.om文件，此文件即为1batch的om模型文件，其他batch同理
113		128
114	## <a name="4">4. 数据预处理</a>	129	## <a name="4">4. 数据预处理</a>
115		130
116	### <a name="41">4.1 数据集获取</a>	131	### <a name="41">4.1 数据集获取</a>
117		132
118	-该模型使用[ImageNet官网](http://www.image-net.org/)的5万张验证集进行测试，图片与标签分别存放在/opt/npu/imagenet/val与/opt/npu/imagenet/val_label.txt。	133	+该模型使用[ImageNet官网](http://www.image-net.org/)的5万张验证集进行测试，图片与标签分别存放在/opt/npu/imageNet/val与/opt/npu/imageNet/val_label_LSL.txt。
119		134
120	### <a name="42">4.2 数据集预处理</a>	135	### <a name="42">4.2 数据集预处理</a>
121		136
		Liang Zhen2022年5月11日使用shell命令for循环
122	使用 sknet_preprocess.py 脚本进行数据预处理，脚本执行命令：	137	使用 sknet_preprocess.py 脚本进行数据预处理，脚本执行命令：
123		138
124	```bash	139	```bash
125	-python3.7 sknet_preprocess.py -s /opt/npu/imagenet/val -d ./prep_data	140	+python sknet_preprocess.py -s /opt/npu/imageNet/val -d ./prep_data
126	```	141	```
127		142
128	### <a name="43">4.3 生成数据集信息文件</a>	143	### <a name="43">4.3 生成数据集信息文件</a>
	@@ -132,7 +147,7 @@ python3.7 sknet_preprocess.py -s /opt/npu/imagenet/val -d ./prep_data
132	2. 执行生成数据集信息脚本，生成数据集信息文件	147	2. 执行生成数据集信息脚本，生成数据集信息文件
133		148
134	```bash	149	```bash
135	- python3.7 get_info.py bin ./prep_data ./sknet_prep_bin.info 224 224	150	+ python get_info.py bin ./prep_data ./sknet_prep_bin_lsl.info 224 224
136	```	151	```
137		152
138	第一个参数为模型输入的类型，第二个参数为生成的bin文件路径，第三个为输出的info文件，后面为宽高信息	153	第一个参数为模型输入的类型，第二个参数为生成的bin文件路径，第三个为输出的info文件，后面为宽高信息
	@@ -141,15 +156,18 @@ python3.7 sknet_preprocess.py -s /opt/npu/imagenet/val -d ./prep_data
141		156
142	### <a name="51">5.1 benchmark工具概述</a>	157	### <a name="51">5.1 benchmark工具概述</a>
143		158
144	-benchmark工具为华为自研的模型推理工具，支持多种模型的离线推理，能够迅速统计出模型在Ascend310上的性能，支持真实数据和纯推理两种模式，配合后处理脚本，可以实现诸多模型的端到端过程，获取工具及使用方法可以参考CANN V100R020C10 推理benchmark工具用户指南 01	159	+benchmark工具为华为自研的模型推理工具，支持多种模型的离线推理，能够迅速统计出模型在Ascend310p上的性能，支持真实数据和纯推理两种模式，配合后处理脚本，可以实现诸多模型的端到端过程，获取工具及使用方法可以参考CANN V100R020C10 推理benchmark工具用户指南 01
145		160
146	### <a name="52">5.2 离线推理</a>	161	### <a name="52">5.2 离线推理</a>
147		162
148	```bash	163	```bash
149	-./benchmark.x86_64 -model_type=vision -device_id=0 -batch_size=1 -om_path=sknet50_bs1.om -input_text_path=sknet_prep_bin.info -input_width=224 -input_height=224 -output_binary=False -useDvpp=False	164	+chmod +x benchmark.x86_64
150	-```	165	+1batch:
		166	+./benchmark.x86_64 -model_type=vision -device_id=0 -batch_size=1 -om_path=sk_resnet50_bs1_310p.om -input_text_path=sknet_prep_bin_lsl.info -input_width=224 -input_height=224 -output_binary=False -useDvpp=False
151		167
152	-输出结果默认保存在当前目录result/dumpOutput_device{0}，模型只有一个名为class的输出，shape为bs * 1000，数据类型为FP32，对应1000个分类的预测结果，每个输入对应的输出对应一个_x.bin文件。	168	+```
		169	+首先要添加benchmark.x86_64可执行权限再进行推理
		170	+输出结果默认保存在当前目录result/dumpOutput_device{0}，模型只有一个名为class的输出，shape为bs * 1000，数据类型为FP32，对应1000个分类的预测结果，每个输入对应的输出对应一个_x.bin文件。上面为1batch，其他batch同理
153		171
154	## <a name="6">6. 精度对比</a>	172	## <a name="6">6. 精度对比</a>
155		173
	@@ -158,13 +176,13 @@ benchmark工具为华为自研的模型推理工具，支持多种模型的离
158	后处理统计TopN精度，调用imagenet_acc_eval.py脚本推理结果与label比对，可以获得Accuracy Top5数据，结果保存在result.json中：	176	后处理统计TopN精度，调用imagenet_acc_eval.py脚本推理结果与label比对，可以获得Accuracy Top5数据，结果保存在result.json中：
159		177
160	```bash	178	```bash
161	-python3.7 vision_metric_ImageNet.py result/dumpOutput_device0/ ../data/sknet/val_label.txt ./ result.json	179	+python vision_metric_ImageNet.py result/dumpOutput_device0/ /opt/npu/imageNet/val_label_LSL.txt ./ result_bs1.json
162	```	180	```
163		181
164	第一个为benchmark输出目录，第二个为数据集配套标签，第三个是生成文件的保存目录，第四个是生成的文件名。查看输出结果：	182	第一个为benchmark输出目录，第二个为数据集配套标签，第三个是生成文件的保存目录，第四个是生成的文件名。查看输出结果：
165		183
166	```json	184	```json
167	-{"title": "Overall statistical evaluation", "value": [{"key": "Number of images", "value": "50000"}, {"key": "Number of classes", "value": "1000"}, {"key": "Top1 accuracy", "value": "77.54%"}, {"key": "Top2 accuracy", "value": "87.12%"}, {"key": "Top3 accuracy", "value": "90.73%"}, {"key": "Top4 accuracy", "value": "92.55%"}, {"key": "Top5 accuracy", "value": "93.71%"}]}	185	+{"title": "Overall statistical evaluation", "value": [{"key": "Number of images", "value": "50000"}, {"key": "Number of classes", "value": "1000"}, {"key": "Top1 accuracy", "value": "77.54%"}, {"key": "Top2 accuracy", "value": "87.12%"}, {"key": "Top3 accuracy", "value": "90.73%"}, {"key": "Top4 accuracy", "value": "92.55%"}, {"key": "Top5 accuracy", "value": "93.70%"}]}
168	```	186	```
169		187
170	经过对bs1与bs16的om测试，本模型batch1的精度与batch16的精度没有差别，精度数据均如上	188	经过对bs1与bs16的om测试，本模型batch1的精度与batch16的精度没有差别，精度数据均如上
		Liang Zhen2022年5月11日使用shell命令for循环
	@@ -173,8 +191,8 @@ python3.7 vision_metric_ImageNet.py result/dumpOutput_device0/ ../data/sknet/val
173		191
174	\| \| TOP1 \| TOP5 \|	192	\| \| TOP1 \| TOP5 \|
175	\| :----------------: \| :------: \| :------: \|	193	\| :----------------: \| :------: \| :------: \|
176	-\| 原github仓库精度 \| 77.5380% \| 93.7000% \|	194	+\| 310精度 \| 77.54% \| 93.70% \|
177	-\| om模型离线推理精度 \| 77.54% \| 93.71% \|	195	+\| 310p精度 \| 77.54% \| 93.70% \|
178		196
179	将得到的om离线模型推理TopN精度与该模型github代码仓上公布的精度对比，精度下降在1%范围之内，故精度达标。	197	将得到的om离线模型推理TopN精度与该模型github代码仓上公布的精度对比，精度下降在1%范围之内，故精度达标。
180		198
	@@ -187,83 +205,79 @@ benchmark工具在整个数据集上推理时也会统计性能数据，但是
187	benchmark工具作纯推理时使用的命令参考如下：	205	benchmark工具作纯推理时使用的命令参考如下：
188		206
189	```bash	207	```bash
190	-./benchmark.x86_64 -round=20 -om_path=sknet50_bs1.om -batch_size=1	208	+./benchmark.x86_64 -round=20 -om_path=sknet50_1bs.om -batch_size=1
191	```	209	```
192		210
193	1. batch1 性能	211	1. batch1 性能
194		212
195	使用benchmark工具在整个数据集上推理时获得的性能数据：	213	使用benchmark工具在整个数据集上推理时获得的性能数据：
196		214
197	- ```	215	+ Interface throughputRate: 812.681 即是batch1 310p单卡吞吐率
198	- [e2e] throughputRate: 143.402, latency: 348669
199	- [data read] throughputRate: 152.003, moduleLatency: 6.57881
200	- [preprocess] throughputRate: 151.416, moduleLatency: 6.60433
201	- [infer] throughputRate: 143.733, Interface throughputRate: 210.306, moduleLatency: 6.16176
202	- [post] throughputRate: 143.732, moduleLatency: 6.95737
203	- ```
204	-
205	- Interface throughputRate: 210.306 * 4 = 841.224 即是batch1 310单卡吞吐率
206		216
207	2. batch4 性能	217	2. batch4 性能
208		218
209	- ```	219	+ Interface throughputRate: 2062.16 即是batch4 310p单卡吞吐率
210	- [INFO] ave_throughputRate: 315.424samples/s, ave_latency: 3.30141ms
211	- ```
212	-
213	- Interface throughputRate: 315.424 * 4 = 1261.696 即是batch4 310单卡吞吐率
214		220
215	3. batch8 性能	221	3. batch8 性能
216		222
217	- ```	223	+ Interface throughputRate: 2174.18 即是batch8 310p单卡吞吐率
218	- [INFO] ave_throughputRate: 365.813samples/s, ave_latency: 2.76526ms
219	- ```
220	-
221	- Interface throughputRate: 365.813 * 4 = 1463.252 即是batch8 310单卡吞吐率
222		224
223	4. batch16 性能	225	4. batch16 性能
224		226
225	- ```	227	+ Interface throughputRate: 2041.62 即是batch16 310p单卡吞吐率
226	- [e2e] throughputRate: 196.399, latency: 254584
227	- [data read] throughputRate: 208.891, moduleLatency: 4.78718
228	- [preprocess] throughputRate: 207.779, moduleLatency: 4.81281
229	- [infer] throughputRate: 197.514, Interface throughputRate: 392.072, modul
230	- [post] throughputRate: 12.3443, moduleLatency: 81.0088
231	- ```
232	-
233	- Interface throughputRate: 392.072 * 4 = 1568.288 即是batch16 310单卡吞吐率
234		228
235	5. batch32 性能	229	5. batch32 性能
236		230
237	- ```	231	+ Interface throughputRate: 1879.74 即是batch32 310p单卡吞吐率
238	- [INFO] ave_throughputRate: 376.691samples/s, ave_latency: 2.66319ms
239	- ```
240		232
241	- Interface throughputRate: 376.691 * 4 = 1506.764 即是batch32 310单卡吞吐率	233	+6. batch64 性能
242		234
243	-性能优化	235	+ Interface throughputRate: 1780.68 即是batch64 310p单卡吞吐率
244		236
245	-> 从profiling数据的op_statistic_0_1.csv看出影响性能的是transpose算子，从onnx结构图看出该算子用于实现softmax任意轴，由pytorch导出时引入，然而softmax在onnx中现已支持任意轴，故可直接删除该算子提升性能，删除代码参考如下：	237	+### <a name="72">7.2 T4性能数据</a>
246		238
247	-```python	239	+在装有T4卡的服务器上测试gpu性能，测试过程请确保卡没有运行其他任务，TensorRT版本：7.2.3.4，cuda版本：11.0，cudnn版本：8.2
248	-model = onnx.load(args.onnx+'.onnx')	240	+
249	-graph = model.graph	241	+```bash
250	-node = graph.node	242	+trtexec --onnx=sk_resnet50.onnx --fp16 --shapes=image:1x3x224x224 --threads
		Liang Zhen2022年5月11日各batch详细的性能数据结果文件保存归档obs即可，readme中不用体现 Liang Zhen2022年5月11日各batch详细的性能数据结果文件保存归档obs即可，readme中不用体现 readme中只呈现性能结果就好
251	-softmax_node_index = []
252	-del_group = []
253	-for i in range(len(node)):
254	- if node[i].op_type == 'Softmax':
255	- del_group.append((node[i-1], node[i], node[i+1], i))
256	-for g in del_group:
257	- new_input = g[0].input
258	- new_output = g[2].output
259	- new_name = g[1].name
260	- new_index = g[3]
261	- new_node = onnx.helper.make_node("Softmax", new_input, new_output, new_name, axis=1)
262	- for n in g[:-1]:
263	- graph.node.remove(n)
264	- graph.node.insert(new_index, new_node)
265	-onnx.save(model, args.onnx+'.onnx')
266	```	243	```
267		244
		245	+gpu T4是4个device并行执行的结果，mean是时延（tensorrt的时延是batch个数据的推理时间），即吞吐率的倒数乘以batch。其中--fp16是算子精度，目前算子精度只测--fp16的。
268		246
		247	+1. batch1 性能
269		248
		249	+ batch1 t4单卡吞吐率：1000/(2.16149/1)=462.64fps
		250	+
		251	+2. batch4 性能
		252	+
		253	+ batch4 t4单卡吞吐率：1000/(4.43787/4)=895.11fps
		254	+
		255	+3. batch8 性能
		256	+
		257	+ batch8 t4单卡吞吐率：1000/(7.84424/8)=1019.86fps
		258	+
		259	+4. batch16 性能
		260	+
		261	+ batch16 t4单卡吞吐率：1000/(15.3644/16)=1041.37fps
		262	+
		263	+5. batch32 性能
		264	+
		265	+ batch32 t4单卡吞吐率：1000/(26.6422/32)=1201.10fps
		266	+
		267	+6. batch64 性能
		268	+
		269	+ batch64 t4单卡吞吐率：1000/(50.5169/64)=1266.90fps
		270	+
		271	+### <a name="73">7.3 性能对比</a>
		272	+
		273	+\| \| 310 \| 310p \| T4 \| 310p/310 \| 310p/T4 \|
		274	+\| :-----\| :------: \| :------: \| :------: \| :------: \| :------: \|
		275	+\| bs1 \| 723.5 \| 812.681 \| 462.64 \| 1.123263303 \| 1.756616376 \|
		276	+\| bs4 \| 1177.36 \| 2062.16 \| 895.11 \| 1.751511857 \| 2.303806236 \|
		277	+\| bs8 \| 1325.04 \| 2174.18 \| 1019.86 \| 1.640841031 \| 2.131841625 \|
		278	+\| bs16 \| 1357.24 \| 2041.62 \| 1041.37 \| 1.504243907 \| 1.960513554 \|
		279	+\| bs32 \| 1329.876 \| 1879.74 \| 1201.1 \| 1.413470128 \| 1.565015403 \|
		280	+\| bs64 \| 1165.084 \| 1780.68 \| 1266.9 \| 1.528370487 \| 1.405541085 \|
		281	+\| 最优bs \| 1357.24 \| 2174.18 \| 1266.9 \| 1.601912705 \| 1.716141763 \|
		282	+
		283	+310P的每个batch的性能都需要达到310的1倍以上，性能达标。

DACL_PyTorch/contrib/cv/classfication/Sknet50/env.sh+0-8

	@@ -1,8 +0,0 @@
1	-#! /bin/bash
2	-
3	-export install_path=/usr/local/Ascend/ascend-toolkit/latest
4	-export PATH=/usr/local/python3.7.5/bin:${install_path}/atc/ccec_compiler/bin:${install_path}/atc/bin:$PATH
5	-export PYTHONPATH=${install_path}/atc/python/site-packages:$PYTHONPATH
6	-export LD_LIBRARY_PATH=${install_path}/atc/lib64:${install_path}/acllib/lib64:$LD_LIBRARY_PATH
7	-export ASCEND_OPP_PATH=${install_path}/opp
8	-export ASCEND_AICPU_PATH=/usr/local/Ascend/ascend-toolkit/latest

AACL_PyTorch/contrib/cv/classfication/Sknet50/modelzoo_level.txt+4-0

@@ -0,0 +1,4 @@
	1	+FuncStatus:OK
	2	+PrecisionStatus:OK
	3	+AutoTune:OK
	4	+PerfStatus:POK

MACL_PyTorch/contrib/cv/classfication/Sknet50/requirements.txt+7-5

	@@ -1,6 +1,8 @@
1	-torch == 1.5.0
2	-torchvision == 0.6.0
3	onnx == 1.9.0	1	onnx == 1.9.0
4	-numpy == 1.19.2	2	+pytorch == 1.8.2+cpu
5	-Pillow == 8.2.0	3	+torchvision == 0.9.2+cpu
6	-opencv-python == 4.5.2	4	+numpy == 1.21.6
		5	+Pillow == 9.1.0
		6	+opencv-python == 4.5.5.64
		7	+sympy == 1.10.1
		8	+decorator == 5.1.1

AACL_PyTorch/contrib/cv/classfication/Sknet50/set_env.sh+8-0

@@ -0,0 +1,8 @@
	1	+export LD_LIBRARY_PATH=/usr/local/Ascend_5_0_4/ascend-toolkit/latest/lib64:/usr/local/Ascend_5_0_4/ascend-toolkit/latest/compiler/lib64/plugin/opskernel:/usr/local/Ascend_5_0_4/ascend-toolkit/latest/compiler/lib64/plugin/nnengine:$LD_LIBRARY_PATH
	Liang Zhen2022年5月18日 `/usr/local/Ascend_5_0_4/ascend-toolkit` 已经存在 `set_env.sh` 文件了，没必要上传这个文件。原有test下面的 `env.sh` 也要删除
	2	+export PYTHONPATH=/usr/local/Ascend_5_0_4/ascend-toolkit/latest/python/site-packages:/usr/local/Ascend_5_0_4/ascend-toolkit/latest/opp/op_impl/built-in/ai_core/tbe:$PYTHONPATH
	3	+export PATH=/usr/local/Ascend_5_0_4/ascend-toolkit/latest/bin:/usr/local/Ascend_5_0_4/ascend-toolkit/latest/compiler/ccec_compiler/bin:$PATH
	4	+export ASCEND_AICPU_PATH=/usr/local/Ascend_5_0_4/ascend-toolkit/latest
	5	+export ASCEND_OPP_PATH=/usr/local/Ascend_5_0_4/ascend-toolkit/latest/opp
	6	+export TOOLCHAIN_HOME=/usr/local/Ascend_5_0_4/ascend-toolkit/latest/toolkit
	7	+export ASCEND_HOME_PATH=/usr/local/Ascend_5_0_4/ascend-toolkit/latest:$ASCEND_HOME_PATH
	8	+

MACL_PyTorch/contrib/cv/classfication/Sknet50/sknet2onnx.py+0-20

	@@ -44,26 +44,6 @@ def main():
44	dynamic_axes = {'image': {0: '-1'}, 'class': {0: '-1'}}	44	dynamic_axes = {'image': {0: '-1'}, 'class': {0: '-1'}}
45	dummy_input = torch.randn(1, 3, 224, 224)	45	dummy_input = torch.randn(1, 3, 224, 224)
46	torch.onnx.export(model.module, dummy_input, args.onnx, input_names = input_names, dynamic_axes = dynamic_axes, output_names = output_names, opset_version=11, verbose=True)	46	torch.onnx.export(model.module, dummy_input, args.onnx, input_names = input_names, dynamic_axes = dynamic_axes, output_names = output_names, opset_version=11, verbose=True)
		Liang Zhen2022年5月11日更建议shell脚本中是用 `--onnx sk_resnet50.onnx`，然后该python脚本不用修改
47	-
48	- # delete transpose operator to improve performance
49	- model = onnx.load(args.onnx)
50	- graph = model.graph
51	- node = graph.node
52	- softmax_node_index = []
53	- del_group = []
54	- for i in range(len(node)):
55	- if node[i].op_type == 'Softmax':
56	- del_group.append((node[i-1], node[i], node[i+1], i))
57	- for g in del_group:
58	- new_input = g[0].input
59	- new_output = g[2].output
60	- new_name = g[1].name
61	- new_index = g[3]
62	- new_node = onnx.helper.make_node("Softmax", new_input, new_output, new_name, axis=1)
63	- for n in g[:-1]:
64	- graph.node.remove(n)
65	- graph.node.insert(new_index, new_node)
66	- onnx.save(model, args.onnx)
67		47
68	if __name__ == '__main__':	48	if __name__ == '__main__':
69	main()	49	main()

MACL_PyTorch/contrib/cv/classfication/Sknet50/test/README.md+1-1

	@@ -23,4 +23,4 @@ git clone https://github.com/implus/PytorchInsight
23		23
24	8.310上执行，执行时确保device空闲	24	8.310上执行，执行时确保device空闲
25	bash test/pth2om.sh	25	bash test/pth2om.sh
26	-bash test/eval_acc_perf.sh --datasets_path=/root/datasets	26	+bash test/eval_acc_perf.sh --datasets_path=/opt/npu

MACL_PyTorch/contrib/cv/classfication/Sknet50/test/eval_acc_perf.sh+47-57

	@@ -1,7 +1,6 @@
1	#!/bin/bash	1	#!/bin/bash
2		2
3	-datasets_path="/opt/npu"	3	+datasets_path="/root/datasets/"
4	-
5	for para in $*	4	for para in $*
6	do	5	do
7	if [[ $para == --datasets_path* ]]; then	6	if [[ $para == --datasets_path* ]]; then
	@@ -10,80 +9,71 @@ do
10	done	9	done
11		10
12	arch=`uname -m`	11	arch=`uname -m`
13	-# prepare data
14	-if [ ! -d "prep_data" ]
15	-then
16	- python3.7 sknet_preprocess.py -s ${datasets_path}/imagenet/val -d ./prep_data
17	- if [ $? != 0 ]; then
18	- echo "fail to preprocess data!"
19	- exit -1
20	- fi
21	-else
22	- echo "prep_data exists, skip."
23	-fi
24		12
25	-# get .info file	13	+echo "====data preprocess===="
26	-if [ ! -f "sknet_prep_bin.info" ]	14	+rm -rf ./prep_data
27	-then	15	+python sknet_preprocess.py -s ${datasets_path}/imageNet/val -d ./prep_data
28	- python3.7 get_info.py bin ./prep_data ./sknet_prep_bin.info 224 224
29	- if [ $? != 0 ]; then
30	- echo "fail to get info file!"
31	- exit -1
32	- fi
33	-else
34	- echo "info file exists, skip."
35	-fi
36	-
37	-# config environment
38	-source env.sh
39	-
40	-# run om
41	-if [ ! -d "result" ]
42	-then
43	- ./benchmark.${arch} -model_type=vision -device_id=0 -batch_size=1 -om_path=sknet50_1bs.om -input_text_path=./sknet_prep_bin.info -input_width=224 -input_height=224 -output_binary=False -useDvpp=False
44	- if [ $? != 0 ]; then
45	- echo "fail to switch 1bs onnx to om!"
46	- exit -1
47	- fi
48	- ./benchmark.${arch} -model_type=vision -device_id=1 -batch_size=16 -om_path=sknet50_16bs.om -input_text_path=./sknet_prep_bin.info -input_width=224 -input_height=224 -output_binary=False -useDvpp=False
49	- if [ $? != 0 ]; then
50	- echo "fail to switch 16bs onnx to om!"
51	- exit -1
52	- fi
53	-else
54	- echo "om running result exists, skip"
55	-fi
56	-
57	-python3.7 vision_metric_ImageNet.py result/dumpOutput_device0/ ${datasets_path}/imagenet/val_label.txt ./ result_1bs.json
58	if [ $? != 0 ]; then	16	if [ $? != 0 ]; then
59	- echo "fail to calculate accurancy of 1bs om!"	17	+ echo "fail!"
60	exit -1	18	exit -1
61	fi	19	fi
62	-python3.7 vision_metric_ImageNet.py result/dumpOutput_device1/ ${datasets_path}/imagenet/val_label.txt ./ result_16bs.json	20	+
		21	+rm -rf sknet_prep_bin_lsl.info
		22	+python get_info.py bin ./prep_data ./sknet_prep_bin_lsl.info 224 224
63	if [ $? != 0 ]; then	23	if [ $? != 0 ]; then
64	- echo "fail to calculate accurancy of 16bs om!"	24	+ echo "fail!"
		25	+ exit -1
		26	+fi
		27	+
		28	+chmod +x benchmark.x86_64
		29	+chmod +x benchmark.aarch64
		30	+source set_env.sh
		Liang Zhen2022年5月11日所有的路径都建议规范成默认标准路径
		31	+if [ $? != 0 ]; then
		32	+ echo "fail!"
		33	+ exit -1
		34	+fi
		35	+
		36	+echo "bs1"
		37	+echo "====performance data===="
		38	+rm -rf result/dumpOutput_device0
		39	+./benchmark.${arch} -model_type=vision -device_id=0 -batch_size=1 -om_path=sk_resnet50_bs1_710.om -input_text_path=sknet_prep_bin_lsl.info -input_width=224 -input_height=224 -output_binary=False -useDvpp=False
		40	+if [ $? != 0 ]; then
		41	+ echo "fail!"
65	exit -1	42	exit -1
66	fi	43	fi
67	echo "====accuracy data===="	44	echo "====accuracy data===="
68	-python3.7 test/parse.py result_1bs.json	45	+rm -rf result_bs1.json
		46	+python vision_metric_ImageNet.py result/dumpOutput_device0/ ${datasets_path}/imageNet/val_label_LSL.txt ./ result_bs1.json
69	if [ $? != 0 ]; then	47	if [ $? != 0 ]; then
70	- echo "fail to show accurancy data of 1bs!"	48	+ echo "fail!"
71	exit -1	49	exit -1
72	fi	50	fi
73	-python3.7 test/parse.py result_16bs.json	51	+
		52	+echo "bs8"
		53	+echo "====performance data===="
		54	+rm -rf result/dumpOutput_device1
		55	+./benchmark.${arch} -model_type=vision -device_id=0 -batch_size=8 -om_path=sk_resnet50_bs8_710.om -input_text_path=sknet_prep_bin_lsl.info -input_width=224 -input_height=224 -output_binary=False -useDvpp=False
74	if [ $? != 0 ]; then	56	if [ $? != 0 ]; then
75	- echo "fail to show accurancy data of 16bs!"	57	+ echo "fail!"
76	exit -1	58	exit -1
77	-fiss	59	+fi
		60	+echo "====accuracy data===="
		61	+rm -rf result_bs8.json
		62	+python vision_metric_ImageNet.py result/dumpOutput_device0/ ${datasets_path}/imageNet/val_label_LSL.txt ./ result_bs8.json
		63	+if [ $? != 0 ]; then
		64	+ echo "fail!"
		65	+ exit -1
		66	+fi
		67	+
78	echo "====performance data===="	68	echo "====performance data===="
79	python test/parse.py result/perf_vision_batchsize_1_device_0.txt	69	python test/parse.py result/perf_vision_batchsize_1_device_0.txt
80	if [ $? != 0 ]; then	70	if [ $? != 0 ]; then
81	- echo "fail to show performance data of 1bs!"	71	+ echo "fail!"
82	exit -1	72	exit -1
83	fi	73	fi
84	-python3.7 test/parse.py result/perf_vision_batchsize_16_device_1.txt	74	+python test/parse.py result/perf_vision_batchsize_8_device_0.txt
85	if [ $? != 0 ]; then	75	if [ $? != 0 ]; then
86	- echo "fail to show performance data of 16bs!"	76	+ echo "fail!"
87	exit -1	77	exit -1
88	fi	78	fi
89	echo "success"	79	echo "success"

AACL_PyTorch/contrib/cv/classfication/Sknet50/test/perf_g.sh+15-0

@@ -0,0 +1,15 @@
	1	+#!/bin/bash
	2	+
	3	+echo "=========bs1========="
	4	+trtexec --onnx=sk_resnet50.onnx --fp16 --shapes=image:1x3x224x224 --threads
	5	+if [ $? != 0 ]; then
	6	+ echo "fail!"
	7	+ exit -1
	8	+fi
	9	+
	10	+echo "=========bs64========="
	11	+trtexec --onnx=sk_resnet50.onnx --fp16 --shapes=image:64x3x224x224 --threads
	12	+if [ $? != 0 ]; then
	13	+ echo "fail!"
	14	+ exit -1
	15	+fi

MACL_PyTorch/contrib/cv/classfication/Sknet50/test/pth2om.sh+30-9

	@@ -1,11 +1,32 @@
1	-rm -rf *.onnx	1	+#! /bin/bash
2	-python3.7 sknet2onnx.py --pth sk_resnet50.pth.tar --onnx sknet50.onnx	2	+
3	-source env.sh	3	+rm -rf sk_resnet50.onnx
4	-rm -rf *.om	4	+rm -rf sk_resnet50_bs1_710.om
5	-atc --framework=5 --model=sknet50.onnx --output=sknet50_1bs --input_format=NCHW --input_shape="image:1,3,224,224" --log=debug --soc_version=Ascend310	5	+rm -rf sk_resnet50_bs8_710.om #batch8最优
6	-atc --framework=5 --model=sknet50.onnx --output=sknet50_16bs --input_format=NCHW --input_shape="image:16,3,224,224" --log=debug --soc_version=Ascend310	6	+rm -rf fusion_result.json
7	-if [ -f "sknet50_1bs.om" ] && [ -f "sknet50_16bs.om" ]; then	7	+
8	- echo "Success changing pth to om."	8	+python sknet2onnx.py --pth sk_resnet50.pth.tar --onnx sk_resnet50
		9	+if [ $? != 0 ]; then
		10	+ echo "fail!"
		11	+ exit -1
		12	+fi
		13	+source set_env.sh
		Liang Zhen2022年5月11日这个路径不规范 `/usr/local/Ascend_5_0_4/ascend-toolkit/set_env.sh`
		14	+if [ $? != 0 ]; then
		15	+ echo "fail!"
		16	+ exit -1
		17	+fi
		18	+atc --framework=5 --model=sk_resnet50.onnx --output=sk_resnet50_bs1_710 --input_format=NCHW --input_shape="image:1,3,224,224" --log=debug --soc_version=Ascend710
		19	+if [ $? != 0 ]; then
		20	+ echo "fail!"
		21	+ exit -1
		22	+fi
		23	+atc --framework=5 --model=sk_resnet50.onnx --output=sk_resnet50_bs8_710 --input_format=NCHW --input_shape="image:8,3,224,224" --log=debug --soc_version=Ascend710
		24	+if [ $? != 0 ]; then
		25	+ echo "fail!"
		26	+ exit -1
		27	+fi
		28	+if [ -f "sk_resnet50_bs1_710.om" ] && [ -f "sk_resnet50_bs8_710.om" ]; then
		29	+ echo "success"
		Liang Zhen2022年5月18日这里建议使用for循环来简化代码
9	else	30	else
10	- echo "Fail!"	31	+ echo "fail!"
11	fi	32	fi