支持模型

目前,Paddle Lite 已严格验证 52 个模型的精度和性能。对视觉类模型做到了充分的支持,覆盖分类、检测和定位,也包含了特色的 OCR 模型的支持。对 NLP 模型也做到了广泛支持,包含翻译、语义表达等等。

除了已严格验证的模型,Paddle Lite 对其他 CV 和 NLP 模型也可以做到大概率支持。

类别 类别细分 模型 支持平台
CV 分类 MobileNetV1 ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, MediatekAPU, KunlunxinXPU, HuaweiAscendNPU, VerisiliconTIMVX, AndroidNNAPI
CV 分类 MobileNetV2 ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, KunlunxinXPU, HuaweiAscendNPU
CV 分类 MobileNetV3_large ARM, X86, GPU(OPENCL,METAL), HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 MobileNetV3_small ARM, X86, GPU(OPENCL,METAL), HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 DPN68 ARM, X86, HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 AlexNet ARM, X86, HuaweiAscendNPU
CV 分类 DarkNet53 ARM, X86, HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 DenseNet121 ARM, X86, HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 EfficientNetB0 ARM, X86, GPU(OPENCL), KunlunxinXPU, HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 GhostNet_x1_3 ARM,HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 HRNet_W18_C ARM, X86,HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 RegNetX_4GF ARM, X86
CV 分类 Xception41 ARM, X86
CV 分类 ResNet18 ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU, VerisiliconTIMVX, AndroidNNAPI
CV 分类 ResNet50 ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU, VerisiliconTIMVX, AndroidNNAPI, IntelOpenVINO
CV 分类 ResNet101 ARM, X86, HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU
CV 分类 ResNeXt50 ARM, X86, HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 MnasNet ARM, HuaweiKirinNPU, HuaweiAscendNPU
CV 分类 SqueezeNet ARM, HuaweiKirinNPU, KunlunxinXPU, HuaweiAscendNPU
CV 分类 ShuffleNet ARM,HuaweiAscendNPU
CV 分类 ShufflenetV2 ARM, KunlunxinXPU, HuaweiAscendNPU
CV 分类 InceptionV3 ARM, X86, HuaweiAscendNPU, HuaweiKirinNPU
CV 分类 InceptionV4 ARM, X86, HuaweiKirinNPU, KunlunxinXPU, HuaweiAscendNPU
CV 分类 VGG16 ARM, X86, GPU(OPENCL), KunlunxinXPU, HuaweiAscendNPU
CV 分类 VGG19 ARM, X86, GPU(OPENCL,METAL), KunlunxinXPU, HuaweiAscendNPU
CV 分类 GoogleNet ARM, X86, KunlunxinXPU, HuaweiAscendNPU, HuaweiKirinNPU
CV 检测 SSD-MobileNetV1 ARM, HuaweiKirinNPU, HuaweiAscendNPU, VerisiliconTIMVX, AndroidNNAPI
CV 检测 SSD-MobileNetV3-large ARM, X86, GPU(OPENCL,METAL),HuaweiAscendNPU*
CV 检测 SSD-VGG16 ARM, X86, HuaweiAscendNPU*
CV 检测 YOLOv3-DarkNet53 ARM, X86, HuaweiAscendNPU*
CV 检测 YOLOv3-MobileNetV1 ARM, X86, HuaweiAscendNPU*, HuaweiKirinNPU
CV 检测 YOLOv3-MobileNetV3 ARM, X86, HuaweiAscendNPU*, HuaweiKirinNPU
CV 检测 yolov3_r50vd_dcn ARM, HuaweiKirinNPU, HuaweiAscendNPU, HuaweiKirinNPU
CV 检测 YOLOv4 ARM, X86, HuaweiAscendNPU*
CV 检测 Faster RCNN ARM
CV 检测 Mask RCNN* ARM
CV 检测 ppyolo_2x ARM,HuaweiAscendNPU*
CV 检测 solov2_r50_fpn_1x ARM
CV OCR ch_ppocr_mobile_v2.0_cls_infer ARM, X86, GPU(OPENCL),HuaweiAscendNPU
CV OCR ch_ppocr_mobile_v2.0_det_infer ARM, X86, GPU(OPENCL), HuaweiAscendNPU, HuaweiKirinNPU
CV OCR ch_ppocr_mobile_v2.0_rec_infer ARM, X86, GPU(OPENCL), HuaweiAscendNPU, HuaweiKirinNPU
CV OCR ch_ppocr_server_v2.0_rec_infer ARM,HuaweiAscendNPU, HuaweiKirinNPU
CV OCR CRNN ARM,HuaweiAscendNPU
CV OCR DB ARM, GPU(OPENCL),HuaweiAscendNPU
CV OCR OCR-Attention ARM
CV REG inference_dnn ARM, GPU(OPENCL)
CV 分割 Deeplabv3 ARM, GPU(OPENCL), HuaweiAscendNPU
CV 分割 UNet ARM, GPU(OPENCL), HuaweiAscendNPU
CV 分割 bisenet ARM, GPU(OPENCL),HuaweiAscendNPU
CV 分割 fastscnn ARM, GPU(OPENCL)
CV 分割 bisenet_v2 ARM, GPU(OPENCL),HuaweiAscendNPU
CV 关键点 HigherHRNet ARM,X86,HuaweiAscendNPU,HuaweiKirinNPU
CV 关键点 HRNet ARM,X86,HuaweiAscendNPU,HuaweiKirinNPU
CV 人脸 FaceDetection ARM
CV 人脸 FaceBoxes* ARM, GPU(OPENCL), HuaweiAscendNPU
CV 人脸 BlazeFace* ARM,HuaweiAscendNPU
CV 人脸 MTCNN ARM, GPU(OPENCL)
NLP 机器翻译 Transformer* ARM,HuaweiKirinNPU,HuaweiAscendNPU
NLP 机器翻译 BERT KunlunxinXPU,HuaweiAscendNPU
NLP 语义表示 ERNIE KunlunxinXPU,HuaweiAscendNPU
NLP 语义理解 ERNIE-TINY ARM,KunlunxinXPU,HuaweiAscendNPU
GAN 风格转换 CycleGAN* HuaweiKirinNPU
GAN 超分辨率 ESRGAN ARM,X86,HuaweiAscendNPU,HuaweiKirinNPU

注意:

  1. 模型列表中 * 代表该模型链接来自 PaddlePaddle/models ,否则为推理模型的下载链接

  2. 支持平台列表中 HuaweiKirinNPU* 代表 ARM + HuaweiKirinNPU 异构计算,否则为 HuaweiKirinNPU 计算

  3. 支持平台列表中 HuaweiAscendNPU* 代表 X86 或 ARM+HuaweiAscendNPU 异构计算,否则为 HuaweiAscendNPU 计算

  4. 寻找更多的可支持的模型还可以转至 PaddlePaddle/models , PaddleHub , PaddleOCR , PaddleDetection