PaddlePaddle Denglin Model Zoo
- 兼容性适配:目前登临科技与百度飞桨深度学习框架已完成三级兼容性适配认证,支持当下主流模型应用场景,覆盖了计算机视觉、智能语音、自然语言处理、推荐、图神经网络和强化学习等领域,支持当下主流模型数量100+;
- 一键启动:通过兼容飞桨推理接口,用户通过指定enable_dlnne()接口一键启动模型,并部署在登临GPU上执行;
- 性能评估:开启enable_profile()接口即可评估模型性能;
- 支持拓展:用户可自行准备飞桨预训练inference模型,通过登临GPU实现加速推理;
- 其他特性:有关enable_dlnne()接口的是详细使用方法可参考Paddle-dlNNE;
Models |
Evaluate Datasets |
Input shape |
Hmean(paddle) |
Hmean(Denglin GPU) |
Latency(ms)(Denglin GPU,BS=1) |
Inference Model |
det_mv3_db_v2.0 |
ICDAR2015 |
1x3x736x1280 |
0.7512 |
0.75092 |
96.365 |
inference_model |
det_r50_vd_db_v2.0 |
ICDAR2015 |
1x3x736x1280 |
0.8238 |
0.82368 |
318.926 |
inference_model |
det_mv3_east_v2.0 |
ICDAR2015 |
1x3x704x1280 |
0.7865 |
0.78680 |
74.671 |
inference_model |
det_r50_vd_east_v2.0 |
ICDAR2015 |
1x3x704x1280 |
0.8488 |
0.84903 |
408.758 |
inference_model |
det_r50_vd_sast_icdar15_v2.0 |
ICDAR2015 |
1x3x896x1536 |
0.8742 |
0.87415 |
1772.236 |
inference_model |
det_mv3_pse_v2.0 |
ICDAR2015 |
1x3x736x1312 |
0.7589 |
0.75894 |
304.274 |
inference_model |
det_r50_vd_pse_v2.0 |
ICDAR2015 |
1x3x736x1312 |
0.8255 |
0.82538 |
674.206 |
inference_model |
rec_svtr_tiny_none_ctc_en |
IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE |
1x3x64x256 |
0.9013(Avg_10,acc) |
0.90105 (acc) |
6.564 |
inference_model |
Models |
Evaluate Datasets |
Input shape |
mAP(paddle) |
mAP(Denglin GPU) |
Latency(ms)(Denglin GPU,BS=1) |
Inference Model |
picodet_lcnet_1_5x_416_coco |
coco |
1x3x416x416 |
0.363 |
0.363 |
133.012 |
inference_model |
picodet_s_320_coco |
coco |
1x3x320x320 |
0.271 |
0.271 |
66.497 |
inference_model |
ppyolo_mbv3_large_coco |
coco |
1x3x320x320 |
0.232 |
0.240 |
27.512 |
inference_model |
ppyolo_r50vd_dcn_1x_coco |
coco |
1x3x608x608 |
0.448 |
0.447 |
444.563 |
inference_model |
ppyolo_tiny_650e_coco |
coco |
1x3x320x320 |
0.206 |
0.207 |
29.661 |
inference_model |
ppyoloe_crn_s_300e_coco |
coco |
1x3x640x640 |
0.430 |
0.430 |
115.896 |
inference_model |
ppyolov2_r50vd_dcn_365e_ |
coco |
1x3x640x640 |
0.491 |
0.491 |
630.109 |
inference_model |
ttfnet_darknet53_1x_coco |
coco |
1x3x512x512 |
0.335 |
0.336 |
413.021 |
inference_model |
yolov3_darknet53_270e_coco |
coco |
1x3x608x608 |
0.391 |
0.391 |
279.647 |
inference_model |
yolov3_mobilenet_v1_270e_coco |
coco |
1x3x608x608 |
0.294 |
0.294 |
136.460 |
inference_model |
yolox_s_300e_coco |
coco |
1x3x640x640 |
0.404 |
0.404 |
142.276 |
inference_model |
Models |
Evaluate Datasets |
Input shape |
mIoU(paddle) |
mIoU(Denglin GPU) |
Latency(ms)(Denglin GPU,BS=1) |
Inference Model |
BiSeNetV1 |
Cityscapes |
1x3x1024x2048 |
0.7519 |
0.75191 |
23283.000 |
inference_model |
BiSeNetv2 |
Cityscapes |
1x3x1024x2048 |
0.7319 |
0.73169 |
594.874 |
inference_model |
CCNet |
Cityscapes |
1x3x1025x2049 |
0.8095 |
0.80951 |
6435.860 |
inference_model |
DDRNet_23(DDRNet) |
Cityscapes |
1x3x1024x2048 |
0.7985 |
0.79847 |
729.794 |
inference_model |
DeepLabv3p_resnet50_cityscapes |
Cityscapes |
1x3x1024x2048 |
0.8036 |
0.8036 |
3712.280 |
inference_model |
ENet |
Cityscapes |
1x3x1024x2048 |
0.6742 |
0.67420 |
801.838 |
inference_model |
FCN_HRNet_W18 |
飞桨内部人像数据集 |
1x3x1024x2048 |
0.787 |
0.78969 |
1580.298 |
inference_model |
GloRe |
Cityscapes |
1x3x1024x2048 |
0.7826 |
0.78256 |
31732.400 |
inference_model |
HRNetW48Contrast |
Cityscapes |
1x3x1024x2048 |
0.8230 |
0.82398 |
3544.080 |
inference_model |
OCRNet_HRNetW18 |
Cityscapes |
1x3x1024x2048 |
0.8067 |
0.80702 |
3801.400 |
inference_model |
PFPNNet |
Cityscapes |
1x3x1024x2048 |
0.7907 |
0.79072 |
28974.200 |
inference_model |
STDC_STDC1 |
Cityscapes |
1x3x1024x2048 |
0.7474 |
0.74739 |
904.822 |
inference_model |
UPERNet |
ADE20K |
1x3x1024x2048 |
0.7958 |
0.79581 |
8477.040 |
inference_model |
Models |
Evaluate Datasets |
Metrics(paddle) |
Metrics(Denglin GPU) |
Latency(ms)(Denglin GPU,BS=1) |
Inference Model |
DSSM |
BQ |
0.93(正序率) |
0.92875(正序率) |
2.805 |
inference_model |
match-pyramid |
Letor07 |
0.39(map) |
0.39296map) |
0.895 |
inference_model |
NCF |
movielens |
0.58(HR@10) 、0.33(NDCG@10) |
0.58543(HR@10) 、 0.33538(NDCG@10) |
0.699 |
inference_model |
DLRM |
criteo |
Auc:0.79 + |
0.80120 |
6.016 |
inference_model |
DeepFM |
Criteo |
Auc:0.78 |
0.794357 |
1.357 |
inference_model |
Models |
Evaluate Datasets |
Metrics |
Reward(CPU) |
Reward(Denglin GPU) |
Latency(ms)(Denglin GPU,BS=1) |
Inference Model |
DQN_variant |
Atari games |
Reward |
3.66667 |
3.66667 |
7.171 |
inference_model |
PPO Atari |
games |
Reward |
-21.0 |
-21.0 |
1.587 |
inference_model |
DQN |
CartPole-v0 |
Reward |
19.0 |
19.0 |
0.350 |
inference_model |
MADDPG |
gym |
Reward |
-75.19758 |
-75.19758 |
0.768 |
inference_model |
Models |
Evaluate Datasets |
Acc(paddle) |
Acc(Denglin GPU) |
Acc(CPU) |
Latency(ms)(Denglin GPU,BS=1) |
Inference Model |
gin |
MUTAG |
-- |
0.78947 |
0.78947 |
10.529 |
inference_model |
GraphSage |
reddit |
-- |
0.74706 |
0.74706 |
581.025 |
inference_model |
gat |
cora |
0.83 |
0.83333 |
-- |
103.456 |
inference_model |
gcn |
CORA |
0.81 |
0.81000 |
-- |
109.458 |
inference_model |
Models |
Evaluate Datasets |
Metrics |
Acc(paddle) |
Acc(Denglin GPU) |
Acc(CPU) |
Latency(ms)(Denglin GPU,BS=1) |
Inference Model |
hifigan |
AISHELL-3 |
mel_loss |
0.1068 |
-- |
0.10699 |
104.712 |
inference_model |
Tacotron2 |
CSMSC |
eval/loss |
-- |
1.928438 |
1.928438 |
-- |
inference_model |
Speedyspeech |
CSMSC |
eval/loss |
-- |
0.879209 |
0.879209 |
146.011 |
inference_model |
以图像分类为例简要介绍模型使用方法,其他模型场景详细用法请参考飞桨官方模型库:
1.模型准备: 通过链接下载登临飞桨ImageNet1K图像分类模型,例如 MobileNetV3.pdmodel 、 MobileNetV3.pdiparams
2.数据准备: 输入图像应符合NCHW Format , Shape 为 [1,3,224,224]
python3 tools/deploy/predict.py \
--model_file ${MODEL_PATH}/MobileNetV3.pdmodel \
--params_file ${MODEL_PATH}/MobileNetV3.pdiparams \
--input_data ${INPUT_DATA}
4.获取最终推理结果,如图像类别、Bouding Box可视化、OCR检测结果等,可参考飞桨模型库相关代码