进迭时空 Model zoo
进迭时空 AI model zoo 收录了受官方加速支持的各类模型，仓库地址位于https://archive.spacemit.com/spacemit-ai/ModelZoo/
Model zoo列表
|--dataset
|--gguf
|--llm
|--classification.tar.gz
|--detection.tar.gz
|--nlp.tar.gz
|--segmentation.tar.gz
模型仓库中包括了图像分类，目标检测，自然语言处理，图像分割以及大模型等仓库，其中收录的量化模型使用进迭时空推理工具推理可以获得良好的加速性能。
模型列表
图像分类
| Model Name | Fp32 top1/% | Quant top1/% | 1线程推理耗时@1.6GHz/ms | 2线程@1.6GHz/ms | 4线程@1.6GHz/ms |
|---|---|---|---|---|---|
| resnet18 | 69.64 | 69.64 | 38.6 | 31.12 | 20.18 |
| resnet50 | 75.46 | 76.00 | 97.3237 | 74.34 | 49.15 |
| seresnet50 | 79.26 | 79.00 | 113.707 | 102.63 | 69.31 |
| resnet50-v1.5 | 75.9 | 74.96 | 97.8166 | 74.32 | 49.28 |
| resnext50 | 80.64 | 79.28 | 149.08 | 123.09 | 95.5 |
| mobilenet_v1 | 71.64 | 71.42 | 32.51 | 21.79 | 13.47 |
| mobilenet_v2 | 71.32 | 71.66 | 36.61 | 24.91 | 15.81 |
| mobilenet_v3_large | 73.26 | 73.20 | 36.815 | 40.916 | 31.624 |
| efficientnet_v1_b0 | 76.88 | 74.7 | 81.5 | 61.6 | 38.85 |
| efficientnet_v1_b1 | 76.7 | 76.1 | 114.01 | 87.02 | 56.66 |
| efficientnet_v2_s | 80.3 | 79.68 | 149.06 | 116.06 | 77.92 |
| mnasnet0_5 | 67.68 | 67.5 | 38.87 | 28.09 | 23.21 |
| mnasnet1_0 | 73.38 | 72.94 | 57.76 | 40.815 | 27.36 |
| v100_gpu64_6ms | 74.58 | 74.82 | 59.3398 | 47.6031 | 31.5937 |
| shufflenet_v2_x1_0 | 68.72 | 68.6 | 39.4998 | 32.5273 | 27.4406 |
| inception_v1 | 66.04 | 65.1 | 122.197 | 86.1181 | 63.6081 |
| inception_resnet_v2 | 80.18 | 80.04 | 309.361 | 227.131 | 152.768 |
| inception_v3 | 76.74 | 76.32 | 125.457 | 87.0467 | 56.5531 |
目标检测
| Model Name | Fp32 mAP/% | Quant mAP/% | 1线程推理耗时@1.6GHz/ms | 2线程推理耗时@1.6GHz/ms | 4线程推理耗时@1.6GHz/ms |
|---|---|---|---|---|---|
| yolov6_n_dyn(320x320) | 39.5 | 38.5 | 325.55 | 317.767 | 309.926 |
| yolov6_n(640x640) | 50.7 | 47.2 | 420.043 | 365.711 | 335.315 |
| yolov6_s_dyn(320x320) | 45.5 | 45.3 | 344.415 | 313.461 | 286.881 |
| yolov5_n_dyn(320x320) | 35.1 | 34.1 | 130.924 | 100.532 | 76.286 |
| yolov5_n(640x640) | 44.0 | 42.9 | 490.434 | 359.839 | 259.232 |
| yolov5_s_dyn(320x320) | 42.0 | 40.4 | 180.013 | 138.459 | 99.3787 |
| yolov3_mobilenetv2(320x320) | 33.7 | 32.4 | 451.454 | 430.331 | 390.081 |
| yolov3_darknet53_dyn(320x320) | 39.8 | 37.8 | 595.954 | 500.68 | 369.697 |
| yolox_s(640x640) | 50.7 | 49.9 | 452.011 | 344.289 | 222.526 |
姿态估计
| Model Name | 1线程推理耗时@1.6GHz/ms | 2线程推理耗时@1.6GHz/ms | 4线程推理耗时@1.6GHz/ms |
|---|---|---|---|
| handpose_x_resnet50 | 115.993 | 78.3508 | 53.2492 |
| rtmdet-nano_hand | 76.8098 | 56.0213 | 40.642 |
| rtmdet-nano | 76.9662 | 56.614 | 41.2351 |
| rtmpose-s_simcc | 116.418 | 94.8116 | 63.014 |
| rtmpose-t | 39.133 | 30.3657 | 23.3327 |
| rtmpose-s | 49.702 | 39.332 | 30.0369 |