PaddleX/docs/support_list/model_list_dcu.en.md

---

comments: true

PaddleX Model List (Hygon DCU)

PaddleX incorporates multiple pipelines, each containing several modules, and each module encompasses various models. The specific models to use can be selected based on the benchmark data below. If you prioritize model accuracy, choose models with higher accuracy. If you prioritize model storage size, select models with smaller storage sizes.

Image Classification Module

Model Name	Top-1 Accuracy (%)	Model Storage Size (MB)	Model Download Link
ResNet18	71.0	41.5	Inference Model/Training Model
ResNet34	74.6	77.3	Inference Model/Training Model
ResNet50	76.5	90.8	Inference Model/Training Model
ResNet101	77.6	158.7	Inference Model/Training Model
ResNet152	78.3	214.2	Inference Model/Training Model
PP-LCNet_x1_0	71.32	10.5	Inference Model/Training Model

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

Image Multi-label Classification Module

Model Name	mAP (%)	Model Storage Size (MB)	Model Download Link
CLIP_vit_base_patch16_448_ML	89.15	325.6	Inference Model/Training Model
PP-HGNetV2-B0_ML	80.98	7.15 / 1.77	Inference Model/Training Model	PP-HGNetV2-B4_ML	21.35 / 8.19	88.5	Inference Model/Training Model
PP-HGNetV2-B6_ML	91.06	286.5	Inference Model/Training Model

Note: The above accuracy metrics are for the multi-label classification task mAP of COCO2017.

Image Feature Module

Model Name	recall@1（%）	Model Storage Size (MB)	Model Download Link
PP-ShiTuV2_rec_CLIP_vit_base	88.69	306.6	Inference Model/Training Model

Note: The above accuracy metrics are for AliProducts recall@1。

Object Detection Module

Model Name	mAP (%)	Model Storage Size (MB)	Model Download Link
PicoDet-L	42.6	20.9	Inference Model/Training Model
PicoDet-M	37.5	16.8	Inference Model/Training Model
PicoDet-S	29.1	4.4	Inference Model/Training Model
PicoDet-XS	26.2	5.7	Inference Model/Training Model
PP-YOLOE_plus-L	52.9	185.3	Inference Model/Training Model
PP-YOLOE_plus-M	49.8	83.2	Inference Model/Training Model
PP-YOLOE_plus-S	43.7	28.3	Inference Model/Training Model
PP-YOLOE_plus-X	54.7	349.4	Inference Model/Training Model
RT-DETR-R18	46.5	70.7	Inference Model/Training Model
FCOS-ResNet50	39.6	124.2	Inference Model/Training Model
YOLOX-N	26.1	3.4	Inference Model/Training Model
FasterRCNN-ResNet34-FPN	37.8	137.5	Inference Model/Training Model
YOLOv3-DarkNet53	39.1	219.7	Inference ModelInference Model/Training Model
Cascade-FasterRCNN-ResNet50-FPN	41.1	245.4	Inference Model/Training Model

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

Small Object Detection Module

Model Name	mAP（%）	Model Storage Size (MB)	Model Download Link
PP-YOLOE_plus_SOD-S	25.1	77.3	Inference Model/Training Model
PP-YOLOE_plus_SOD-L	31.9	325.0	Inference Model/Training Model
PP-YOLOE_plus_SOD-largesize-L	42.7	340.5	Inference Model/Training Model

Note: The above accuracy metrics are for VisDrone-DET validation set mAP(0.5:0.95)。

Semantic Segmentation Module

Model Name	mIoU (%)	Model Storage Size (MB)	Model Download Link
Deeplabv3_Plus-R50	80.36	94.9	Inference Model/Training Model
Deeplabv3_Plus-R101	81.10	162.5	Inference Model/Training Model
PP-LiteSeg-T	77.04	31	Inference Model/Training Model

Note: The above accuracy metrics are mIoU on the Cityscapes dataset.

Abnormality Detection Module

Model Name	Avg（%）	Model Storage Size (MB)	Model Download Link
STFPM	96.2	21.5	Inference Model/Training Model

Note: The above accuracy metrics are evaluated on the MVTec AD dataset using the average anomaly score.

Face Detection Module

Model Name	AP (%) Easy/Medium/Hard	Model Storage Size (MB)	Model Download Link
PicoDet_LCNet_x2_5_face	93.7/90.7/68.1	28.9	Inference Model/Training Model

Note: The above accuracy metrics are evaluated on the WIDER-FACE validation set with an input size of 640*640.

Text Detection Module

Model Name	Detection Hmean (%)	Model Storage Size (MB)	Model Download Link
PP-OCRv4_mobile_det	77.79	4.2	Inference Model/Training Model
PP-OCRv4_server_det	82.69	100.1	Inference Model/Training Model

Note: The evaluation set for the above accuracy metrics is PaddleOCR's self-built Chinese dataset, covering street scenes, web images, documents, handwriting, and more scenarios, with 500 images for detection.

Text Recognition Module

Model Name	Recognition Avg Accuracy (%)	Model Storage Size (MB)	Model Download Link
PP-OCRv4_mobile_rec	78.20	10.6	Inference Model/Training Model
PP-OCRv4_server_rec	79.20	71.2	Inference Model/Training Model

Note: The evaluation set for the above accuracy metrics is PaddleOCR's self-built Chinese dataset, covering street scenes, web images, documents, handwriting, and more scenarios, with 11,000 images for text recognition.

Time Series Forecasting Module

Model Name	mse	mae	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (MB)	Model Download Link
DLinear	0.382	0.394	0.34 / 0.12	0.64 / 0.06	0.072	Inference Model/Training Model
NLinear	0.386	0.392	0.27 / 0.10	0.49 / 0.08	0.04	Inference Model/Training Model
RLinear	0.385	0.392	0.39 / 0.18	0.82 / 0.08	0.04	Inference Model/Training Model

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