forked from PaddlePaddle/PaddleSlim
-
Notifications
You must be signed in to change notification settings - Fork 0
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
update quantization benchmark docs (PaddlePaddle#1497)
- Loading branch information
Showing
1 changed file
with
73 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,73 @@ | ||
| # 模型量化模型库 | ||
|
|
||
| ## 图像分类 | ||
|
|
||
| | 模型 | 示例 | Top-1 Acc<sup>FP32</sup> | Top-1 Acc<sup>INT8</sup> | 模型体积<sup>FP32</sup> | 模型体积<sup>INT8</sup> | TRT<sup>FP32</sup> | TRT<sup>FP16</sup> | TRT<sup>INT8</sup> | CPU<sup>FP32</sup> | CPU<sup>INT8</sup> | ARM-CPU<sup>FP32</sup>| ARM-CPU<sup>INT8</sup> | | ||
| |:---:|:---:|:----:|:---:|:-----:|:-----:|:-----:|:-----:|:-----:|:----:|:---:|:-----:|:-----:| | ||
| | ResNet50-vd | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/image_classification) | 79.08 | 78.17 | 98MB | 25MB | 7.0ms | 2.4ms | 1.6ms | 77.7ms | 24.7ms | - | - | | ||
| | MobileNetV3_large_x1_0 | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/image_classification) | 75.32 | 74.04 | 22MB | 5.6MB | 2.3ms | 1.4ms | 1.1ms | 11.3ms | 7.7ms | 16.6ms | 9.8ms | | ||
| | PPLCNetV2 | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/image_classification) | 76.63 | 75.87 | 26MB | 6.7MB | 2.1ms | 1.2ms | 0.8ms | 16.3ms | 7.7ms | 36.5ms | 15.8ms | | ||
| | PPHGNet_tiny | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/image_classification) | 79.59 | 78.95 | 57MB | 15MB | 6.4ms | 2.6ms | 1.7ms | 75.3ms | 32.8ms | - | - | | ||
| | EfficientNetB0 | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/image_classification) | 77.02 | 74.64 | 21MB | 5.7MB | 3.6ms | 2.1ms | 1.7ms | 29.5ms | - | - | - | | ||
|
|
||
| 测试环境： | ||
| - GPU: Tesla T4; cuda11.1/cudnn8.1.1/trt8.4.0.6; | ||
| - CPU: Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz | ||
| - ARM CPU 测试环境：SDM865(4xA77+4xA55) | ||
| - PaddlePaddle 2.4.0 | ||
| - CPU测试开启MKLDNN，单线程 | ||
| - 以上数据使用[PaddleTest](https://github.com/PaddlePaddle/PaddleTest/tree/develop/inference/python_api_test/test_int8_model)的Python脚本测试得到 | ||
|
|
||
|
|
||
| ## 目标检测 | ||
|
|
||
| | 模型 | 示例 | mAP(0.5:0.95)<sup>FP32</sup> | mAP(0.5:0.95)<sup>INT8</sup> | 模型体积<sup>FP32</sup> | 模型体积<sup>INT8</sup> | TRT<sup>FP32</sup> | TRT<sup>FP16</sup> | TRT<sup>INT8</sup> | CPU<sup>FP32</sup> | CPU<sup>INT8</sup> | ARM-CPU<sup>FP32</sup>| ARM-CPU<sup>INT8</sup> | | ||
| |:---:|:---:|:----:|:---:|:-----:|:-----:|:-----:|:-----:|:-----:|:----:|:---:|:-----:|:-----:| | ||
| | PP-YOLOE-l | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/detection) | 50.9 | 50.6 | 98MB | 25MB | - | - | - | 1526ms | 1081ms | - | - | | ||
| | PP-PicoDet | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/detection) | 50.9 | 50.6 | 98MB | 25MB | - | 6.8ms | 6.1ms | 56.9ms | 39.0ms | - | - | | ||
| | YOLOv5s | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/pytorch_yolo_series) | 37.4 | 36.9 | 28.1MB | 7.4MB | 14.5ms | 8.5ms | 7.4ms | 310.5ms | 265.7ms | - | - | | ||
| | YOLOv6s | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/pytorch_yolo_series) | 42.4 | 41.3 | 65.9MB | 16.8MB | 17.4ms | 7.4ms | 5.3ms | 387.3ms | 157.2ms | - | - | | ||
| | YOLOv7 | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/pytorch_yolo_series) | 51.1 | 50.8 | 141.4MB | 35.8MB | 48.2 | 17.0ms | 12.4ms | 1268.6ms | 864.3ms | - | - | | ||
|
|
||
| 测试环境： | ||
| - GPU: Tesla T4; cuda11.1/cudnn8.1.1/trt8.4.0.6; | ||
| - CPU: Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz | ||
| - ARM CPU 测试环境：SDM865(4xA77+4xA55) | ||
| - PaddlePaddle 2.4.0 | ||
| - CPU测试开启MKLDNN，单线程 | ||
| - 以上数据使用[PaddleTest](https://github.com/PaddlePaddle/PaddleTest/tree/develop/inference/python_api_test/test_int8_model)的Python脚本测试得到，在COCO val2017验证集上测试 | ||
|
|
||
| ## 语义分割 | ||
|
|
||
| | 模型 | 示例 | mIoU<sup>FP32</sup> | mIoU<sup>INT8</sup> | 模型体积<sup>FP32</sup> | 模型体积<sup>INT8</sup> | TRT<sup>FP32</sup> | TRT<sup>FP16</sup> | TRT<sup>INT8</sup> | CPU<sup>FP32</sup> | CPU<sup>INT8</sup> | ARM-CPU<sup>FP32</sup>| ARM-CPU<sup>INT8</sup> | | ||
| |:---:|:---:|:----:|:---:|:-----:|:-----:|:-----:|:-----:|:-----:|:----:|:---:|:-----:|:-----:| | ||
| | PP-HumanSeg-Lite | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/semantic_segmentation) | 96.0 | 95.94 | 0.54MB | 0.2MB | 2.7ms | 2.1ms | 1.9ms | 63.8ms | 59.6ms | 56.36ms | 49.65ms | | ||
| | PP-Liteseg | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/semantic_segmentation) | 77.04 | 76.99 | 31MB | 7.8MB | 46.4ms | 30.4ms | 27.1ms | 1289ms | 785ms | - | - | | ||
| | HRNet | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/semantic_segmentation) | 78.97 | 78.46 | 37MB | 9.7MB | 172.9ms | 83.3ms | 67.6ms | 4348ms | 2358ms | - | - | | ||
| | UNet | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/semantic_segmentation) | 65.0 | 64.28 | 52MB | 13MB | 555.4ms | 115.0ms | 82.2ms | 18687ms | 7768ms | - | - | | ||
| | Deeplabv3-ResNet50 | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/semantic_segmentation) | 79.91 | 78.78 | 150MB | 38MB | 983.2ms | 106.2ms | 82.1ms | 22515ms | 5436ms | - | - | | ||
|
|
||
|
|
||
| 测试环境： | ||
| - GPU: Tesla T4; cuda11.1/cudnn8.1.1/trt8.4.0.6; | ||
| - CPU: Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz | ||
| - ARM CPU 测试环境：SDM865(4xA77+4xA55) | ||
| - PaddlePaddle 2.4.0 | ||
| - CPU测试开启MKLDNN，单线程；ARM CPU单线程测试 | ||
| - 以上数据使用[PaddleTest](https://github.com/PaddlePaddle/PaddleTest/tree/develop/inference/python_api_test/test_int8_model)的Python脚本测试得到 | ||
|
|
||
| ## NLP | ||
|
|
||
| | 模型 | 示例 | Acc<sup>FP32</sup> | Acc<sup>INT8</sup> | 模型体积<sup>FP32</sup> | 模型体积<sup>INT8</sup> | TRT<sup>FP32</sup> | TRT<sup>FP16</sup> | TRT<sup>INT8</sup> | CPU<sup>FP32</sup> | CPU<sup>INT8</sup> | ARM-CPU<sup>FP32</sup>| ARM-CPU<sup>INT8</sup> | | ||
| |:---:|:---:|:----:|:---:|:-----:|:-----:|:-----:|:-----:|:-----:|:----:|:---:|:-----:|:-----:| | ||
| | ERNIE 3.0-Medium | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/nlp) | 75.37 | 74.4 | 288MB | 155MB | 73.84ms | 11.38ms | 4.43ms | 1519ms | 591ms | - | - | | ||
| | PP-MiniLM | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/nlp) | 74.03 | 73.3 | 228MB | 106MB | 2.7ms | 73.72ms | 11.24ms | 1454ms | 702ms | - | - | | ||
| | BERT | [示例](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/pytorch_huggingface) | 60.3 | 58.69 | 414MB | 149MB | 13.93ms | 2.85ms | 2.13ms | 251.9ms | 79.8ms | - | - | | ||
|
|
||
| 测试环境： | ||
| - GPU: Tesla T4; cuda11.1/cudnn8.1.1/trt8.4.0.6; | ||
| - CPU: Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz | ||
| - ARM CPU 测试环境：SDM865(4xA77+4xA55) | ||
| - PaddlePaddle 2.4.0 | ||
| - CPU测试开启MKLDNN，单线程 | ||
| - 以上数据使用[PaddleTest](https://github.com/PaddlePaddle/PaddleTest/tree/develop/inference/python_api_test/test_int8_model)的Python脚本测试得到 |