Sākums Izpētīt Palīdzība
Reģistrēties Pierakstīties
zhengchun
/
PaddleX
1
0
Atdalīts 0
Faili Problēmas 0 Izmaiņu pieprasījumi 0 Vikivietne
Koks: 3892d429df
Atzari Tagi
PaddleX
/
docs
/
support_list
/
model_list_npu_en.md

model_list_npu_en.md 7.9 KB
Vēsture Neapstrādāts

简体中文 | English

PaddleX Model List (Huawei Ascend NPU)

PaddleX incorporates multiple pipelines, each containing several modules, and each module encompasses various models. You can select the appropriate models based on the benchmark data below. If you prioritize model accuracy, choose models with higher accuracy. If you prioritize model size, select models with smaller storage requirements.

Image Classification Module

| Model Name | Top-1 Accuracy (%) | Model Size (M) | |-|-|-| | CLIP_vit_base_patch16_224 | 85.36 | 306.5 M | | CLIP_vit_large_patch14_224 | 88.1 | 1.04 G | | ConvNeXt_base_224 | 83.84 | 313.9 M | | ConvNeXt_base_384 | 84.90 | 313.9 M | | ConvNeXt_large_224 | 84.26 | 700.7 M | | ConvNeXt_large_384 | 85.27 | 700.7 M | | ConvNeXt_small | 83.13 | 178.0 M | | ConvNeXt_tiny | 82.03 | 101.4 M | | MobileNetV1_x0_75 | 68.8 | 9.3 M | | MobileNetV1_x1_0 | 71.0 | 15.2 M | | MobileNetV2_x0_5 | 65.0 | 7.1 M | | MobileNetV2_x0_25 | 53.2 | 5.5 M | | MobileNetV2_x1_0 | 72.2 | 12.6 M | | MobileNetV2_x1_5 | 74.1 | 25.0 M | | MobileNetV2_x2_0 | 75.2 | 41.2 M | | MobileNetV3_large_x0_5 | 69.2 | 9.6 M | | MobileNetV3_large_x0_35 | 64.3 | 7.5 M | | MobileNetV3_large_x0_75 | 73.1 | 14.0 M | | MobileNetV3_large_x1_0 | 75.3 | 19.5 M | | MobileNetV3_large_x1_25 | 76.4 | 26.5 M | | MobileNetV3_small_x0_5 | 59.2 | 6.8 M | | MobileNetV3_small_x0_35 | 53.0 | 6.0 M | | MobileNetV3_small_x0_75 | 66.0 | 8.5 M | | MobileNetV3_small_x1_0 | 68.2 | 10.5 M | | MobileNetV3_small_x1_25 | 70.7 | 13.0 M | | PP-HGNet_base | 85.0 | 249.4 M | | PP-HGNet_small | 81.51 | 86.5 M | | PP-HGNet_tiny | 79.83 | 52.4 M | | PP-HGNetV2-B0 | 77.77 | 21.4 M | | PP-HGNetV2-B1 | 79.18 | 22.6 M | | PP-HGNetV2-B2 | 81.74 | 39.9 M | | PP-HGNetV2-B3 | 82.98 | 57.9 M | | PP-HGNetV2-B4 | 83.57 | 70.4 M | | PP-HGNetV2-B5 | 84.75 | 140.8 M | | PP-HGNetV2-B6 | 86.30 | 268.4 M | |PP-LCNet_x0_5|63.14|6.7 M| |PP-LCNet_x0_25|51.86|5.5 M| |PP-LCNet_x0_35|58.09|5.9 M| |PP-LCNet_x0_75|68.18|8.4 M| |PP-LCNet_x1_0|71.32|10.5 M| |PP-LCNet_x1_5|73.71|16.0 M| |PP-LCNet_x2_0|75.18|23.2 M| |PP-LCNet_x2_5|76.60|32.1 M| |PP-LCNetV2_base|77.05|23.7 M| |ResNet18_vd|72.3|41.5 M| |ResNet18|71.0|41.5 M| |ResNet34_vd|76.0|77.3 M| |ResNet34|74.6|77.3 M| |ResNet50_vd|79.1|90.8 M| |ResNet50|76.5|90.8 M| |ResNet101_vd|80.2|158.4 M| |ResNet101|77.6|158.7 M| |ResNet152_vd|80.6|214.3 M| |ResNet152|78.3|214.2 M| |ResNet200_vd|80.9|266.0 M| |SwinTransformer_base_patch4_window7_224|83.37|310.5 M| |SwinTransformer_small_patch4_window7_224|83.21|175.6 M| |SwinTransformer_tiny_patch4_window7_224|81.10|100.1 M|

Note: The above accuracy metrics refer to Top-1 Accuracy on the ImageNet-1k validation set.

Object Detection Module

| Model Name | mAP (%) | Model Size (M) | |-|-|-| |CenterNet-DLA-34|37.6|75.4 M| |CenterNet-ResNet50|38.9|319.7 M| |DETR-R50|42.3|159.3 M| |FasterRCNN-ResNet34-FPN|37.8|137.5 M| |FasterRCNN-ResNet50-FPN|38.4|148.1 M| |FasterRCNN-ResNet50-vd-FPN|39.5|148.1 M| |FasterRCNN-ResNet50-vd-SSLDv2-FPN|41.4|148.1 M| |FasterRCNN-ResNet101-FPN|41.4|216.3 M| |FCOS-ResNet50|39.6|124.2 M| |PicoDet-L|42.6|20.9 M| |PicoDet-M|37.5|16.8 M| |PicoDet-S|29.1|4.4 M | |PicoDet-XS|26.2|5.7M | |PP-YOLOE_plus-L|52.9|185.3 M| |PP-YOLOE_plus-M|49.8|83.2 M| |PP-YOLOE_plus-S|43.7|28.3 M| |PP-YOLOE_plus-X|54.7|349.4 M| |RT-DETR-H|56.3|435.8 M| |RT-DETR-L|53.0|113.7 M| |RT-DETR-R18|46.5|70.7 M| |RT-DETR-R50|53.1|149.1 M| |RT-DETR-X|54.8|232.9 M| |YOLOv3-DarkNet53|39.1|219.7 M| |YOLOv3-MobileNetV3|31.4|83.8 M| |YOLOv3-ResNet50_vd_DCN|40.6|163.0 M|

Note: The above accuracy metrics are for COCO2017 validation set mAP(0.5:0.95).

Semantic Segmentation Module

| Model Name | mIoU (%) | Model Size (M) | |-|-|-| | Deeplabv3_Plus-R50 | 80.36 | 94.9 M | | Deeplabv3_Plus-R101 | 81.10 | 162.5 M | | Deeplabv3-R50 | 79.90 | 138.3 M | | Deeplabv3-R101 | 80.85 | 205.9 M | | OCRNet_HRNet-W48 | 82.15 | 249.8 M | | PP-LiteSeg-T | 73.10 | 28.5 M |

Note: The above accuracy metrics are for Cityscapes dataset mIoU.

Instance Segmentation Module

| Model Name | Mask AP | Model Size (M) | |-|-|-| | Mask-RT-DETR-H | 50.6 | 449.9 M | | Mask-RT-DETR-L | 45.7 | 113.6 M | | Mask-RT-DETR-M | 42.7 | 66.6 M | | Cascade-MaskRCNN-ResNet50-FPN | 36.3 | 254.8 M | | Cascade-MaskRCNN-ResNet50-vd-SSLDv2-FPN | 39.1 | 254.7 M | | PP-YOLOE_seg-S | 32.5 | 31.5 M |

Note: The above accuracy metrics are for COCO2017 validation set Mask AP(0.5:0.95).

Text Detection Module

| Model Name | Detection Hmean (%) | Model Size (M) | |-|-|-| | PP-OCRv4_mobile_det | 77.79 | 4.2 M | | PP-OCRv4_server_det | 82.69 | 100.1 M |

Note: The above accuracy metrics are evaluated on PaddleOCR's self-built Chinese dataset, covering street scenes, web images, documents, and handwritten scenarios, with 500 images for detection.

Text Recognition Module

| Model Name | Recognition Avg Accuracy (%) | Model Size (M) | |-|-|-| | PP-OCRv4_mobile_rec | 78.20 | 10.6 M | | PP-OCRv4_server_rec | 79.20 | 71.2 M |

Note: The above accuracy metrics are evaluated on PaddleOCR's self-built Chinese dataset, covering street scenes, web images, documents, and handwritten scenarios, with 11,000 images for text recognition.

| Model Name | Recognition Avg Accuracy (%) | Model Size (M) | |-|-|-| | ch_SVTRv2_rec | 68.81 | 73.9 M |

Note: The above accuracy metrics are evaluated on the PaddleOCR Algorithm Model Challenge - Task 1: OCR End-to-End Recognition A-Rank.

| Model Name | Recognition Avg Accuracy (%) | Model Size (M) | |-|-|-| | ch_RepSVTR_rec | 65.07 | 22.1 M |

Note: The above accuracy metrics are evaluated on the PaddleOCR Algorithm Model Challenge - Task 1: OCR End-to-End Recognition B-Rank.

Table Structure Recognition Module

| Model Name | Accuracy (%) | Model Size (M) | |-|-|-| | SLANet | 76.31 | 6.9 M |

Note: The above accuracy metrics are measured on the PubtabNet English table recognition dataset.

Layout Analysis Module

Model Name mAP (%) Model Size (M)
PicoDet_layout_1x 86.8 7.4M

Note: The evaluation set for the above accuracy metrics is PaddleOCR's self-built layout analysis dataset, containing 10,000 images.

Time Series Forecasting Module

Model Name MSE MAE Model Size (M)
DLinear 0.382 0.394 72K
NLinear 0.386 0.392 40K
Nonstationary 0.600 0.515 55.5 M
PatchTST 0.385 0.397 2.0M
RLinear 0.384 0.392 40K
TiDE 0.405 0.412 31.7M
TimesNet 0.417 0.431 4.9M

Note: The above accuracy metrics are measured on the ETTH1 dataset (evaluation results on the test set test.csv).

Time Series Anomaly Detection Module

Model Name Precision Recall F1-Score Model Size (M)
AutoEncoder_ad 99.36 84.36 91.25 52K
DLinear_ad 98.98 93.96 96.41 112K
Nonstationary_ad 98.55 88.95 93.51 1.8M
PatchTST_ad 98.78 90.70 94.57 320K
TimesNet_ad 98.37 94.80 96.56 1.3M

Note: The above accuracy metrics are measured on the PSM dataset.

Time Series Classification Module

Model Name Acc (%) Model Size (M)
TimesNet_cls 87.5 792K

Note: The above accuracy metrics are measured on the UWaveGestureLibrary: Training, Evaluation datasets.