From 71a258796dba25845e8fc023b5c67da54765ab40 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=AD=99=E5=BD=A6=E5=9B=BD?= Date: Thu, 27 Jul 2023 07:10:24 +0000 Subject: [PATCH] update ACL_TensorFlow/built-in/cv/Facenet_for_ACL/README.md. MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Signed-off-by: 孙彦国 --- .../built-in/cv/Facenet_for_ACL/README.md | 36 +++++++++++++------ 1 file changed, 25 insertions(+), 11 deletions(-) diff --git a/ACL_TensorFlow/built-in/cv/Facenet_for_ACL/README.md b/ACL_TensorFlow/built-in/cv/Facenet_for_ACL/README.md index 3df8eb450..068560664 100644 --- a/ACL_TensorFlow/built-in/cv/Facenet_for_ACL/README.md +++ b/ACL_TensorFlow/built-in/cv/Facenet_for_ACL/README.md @@ -52,7 +52,7 @@ cd Modelzoo-TensorFlow/ACL_TensorFlow/built-in/cv/Facenet_for_ACL pb模型样例(20180402): ``` - atc --framework=3 --model=./model/facenet_tf.pb --output=./model/facenet --soc_version=Ascend310P3 --insert_op_conf=./facenet_tensorflow.cfg --input_format=NHWC --input_shape=input:64,160,160,3 + atc --framework=3 --model=./facenet_tf.pb --output=./facenet --soc_version=Ascend310P3 --insert_op_conf=./facenet_tensorflow.cfg --input_format=NHWC --input_shape=input:64,160,160,3 ``` ### 4.量化 @@ -79,6 +79,30 @@ mv ./quant/facenet_quantized.pb ./ atc --framework=3 --model=./facenet_quantized.pb --output=./facenet_quant --soc_version=Ascend310P3 --insert_op_conf=./facenet_tensorflow.cfg --input_format=NHWC --input_shape=input:64,160,160,3 +### 5.模型精度性能 + +1.执自行安装ais_bench工具 + +2.执行 +原模型： +python3 -m ais_bench --model ./facenet.om --input datasets_bin/data_image_bin --output ./output --device 0 +量化模型： +python3 -m ais_bench --model ./facenet_quant.om --input datasets_bin/data_image_bin --output ./output --device 0 + +3. 精度验证 + +python3 post2.py ../datasets ../output/2023_05_11-10_55_20 ../datasets_bin/data_label_bin --lfw_batch_size 1 --distance_metric 1 --use_flipped_images --subtract_mean + + +| model | mode | ***data*** | Embeddings Accuracy | +| :---------------:| :---------------: | :---------: | :---------: | +| pb(20180402)| offline Inference | 12000 images | 99.550% | +| pb(20180408)| offline Inference | 12000 images | 99.133% | +| pb(20180408量化)| offline Inference | 12000 images | 99.06% | + + + + - 编译程序 ``` @@ -104,13 +128,3 @@ atc --framework=3 --model=./facenet_quantized.pb --output=./facenet_quant --soc | pb(20180402)| offline Inference | 12000 images | 99.550% | | pb(20180408)| offline Inference | 12000 images | 99.133% | -### 6.量化模型精度性能 - -1.执行推理，自行安装ais_bench工具 - -2.执行 python3 -m ais_bench --model ./facenet_quant.om --input datasets_bin/data_image_bin --output ./output --device 0 - -3. 精度验证 - -python3 post2.py ../datasets ../output/2023_05_11-10_55_20 ../datasets_bin/data_label_bin --lfw_batch_size 1 --distance_metric 1 --use_flipped_images --subtract_mean - -- Gitee