快速开始

备注

阅读本篇前，请确保已按照 安装教程 准备好昇腾环境及 timm ！

本文档帮助昇腾开发者快速使用 timm × 昇腾 进行训练和推理。

导入 torch-npu

首先在入口脚本（如本文档中的 train.py， validate.py ， inference.py）导入 torch 后，导入 torch-npu：

1import torch
2import torch-npu

单卡/分布式训练

以 ImageNet-1000 数据集的训练为例，使用以下脚本启动单卡/多卡 NPU 上基于 timm 的图像分类模型训练：

备注

请根据您的 NPU 环境指定 NPU 卡数量 num_npus 和模型名称/路径 model ，并替换数据集路径 path/to/dataset/ImageNet-1000

 1num_npus=1
 2./distributed_train.sh $num_npus path/to/dataset/ImageNet-1000 \
 3    --device npu \
 4    --model seresnet34 \
 5    --sched cosine \
 6    --epochs 150 \
 7    --warmup-epochs 5 \
 8    --lr 0.4 \
 9    --reprob 0.5 \
10    --remode pixel \
11    --batch-size 256 \
12    --amp -j 4

模型验证

备注

请根据实际情况替换验证集数据路径 path/to/data 、模型路径 path/to/model

1python validate.py path/to/data --device npu --model path/to/model --batch-size 64 --pretrained

正常输出验证过程日志及最终验证结果 result 说明验证成功，如下为一种示例（根据模型及数据集不同，日志会有区别）：

Validating in float32. AMP not enabled.
Loading pretrained weights from Hugging Face hub (timm/tiny_vit_21m_512.dist_in22k_ft_in1k)
[timm/tiny_vit_21m_512.dist_in22k_ft_in1k] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
Model ./model_ckpts/tiny_vit_21m_512 created, param count: 21268120
Data processing configuration for current model + dataset:
        input_size: (3, 512, 512)
        interpolation: bicubic
        mean: (0.485, 0.456, 0.406)
        std: (0.229, 0.224, 0.225)
        crop_pct: 1.0
        crop_mode: squash
Test: [   0/157]  Time: 7.083s (7.083s,    9.04/s)  Loss:  0.4765 (0.4765)  Acc@1:  93.750 ( 93.750)  Acc@5:  96.875 ( 96.875)
Test: [  10/157]  Time: 0.400s (1.008s,   63.50/s)  Loss:  0.6594 (0.4929)  Acc@1:  78.125 ( 87.926)  Acc@5:  98.438 ( 98.011)
Test: [  20/157]  Time: 0.399s (0.719s,   89.04/s)  Loss:  0.1891 (0.4682)  Acc@1:  96.875 ( 89.435)  Acc@5: 100.000 ( 98.289)

... ...

* Acc@1 86.040 (13.960) Acc@5 97.750 (2.250)
--result
{
    "model": "./model_ckpts/tiny_vit_21m_512",
    "top1": 86.04,
    "top1_err": 13.96,
    "top5": 97.75,
    "top5_err": 2.25,
    "param_count": 21.27,
    "img_size": 512,
    "crop_pct": 1.0,
    "interpolation": "bicubic"
}

模型推理

备注

请根据实际情况替换验证集数据路径 path/to/data 和模型权重路径 path/to/checkpoint/model_best.pth.tar

1python inference.py ../open_clip/data/ImageNet-1000/val/ \
2    --device npu \
3    --batch-size 64 \
4    --model ./model_ckpts/tiny_vit_21m_512 \
5    --label-type detail \
6    --topk 5

正常输出验证过程日志及最终验证结果 result 说明验证成功，如下为一种示例（根据模型及数据集不同，日志会有区别）：

Running inference in float32. AMP not enabled.
Loading pretrained weights from Hugging Face hub (timm/tiny_vit_21m_512.dist_in22k_ft_in1k)
[timm/tiny_vit_21m_512.dist_in22k_ft_in1k] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
Model ./model_ckpts/tiny_vit_21m_512 created, param count: 21268120
Predict: [0/157] Time 6.418 (6.418)
Predict: [10/157] Time 0.394 (0.942)
Predict: [20/157] Time 0.427 (0.708)

... ...

"ILSVRC2012_val_00005844.JPEG":{
    "label":[
        "stinkhorn, carrion fungus: any of various ill-smelling brown-capped fungi of the order Phallales",
        "earthstar: any fungus of the family Geastraceae; in form suggesting a puffball whose outer peridium splits into the shape of a star",
        "coral fungus: any of numerous fungi of the family Clavariaceae often brightly colored that grow in often intricately branched clusters like coral",
        "mushroom: fleshy body of any of numerous edible fungi",
        "gyromitra: any fungus of the genus Gyromitra"
    ],
    "prob":[
        0.878154695,
        0.0030552391,
        0.0012754521,
        0.0010740706,
        0.000946458
    ]
},

... ...