Show&Tell

简体中文

基于Mindspore实现“Deep High-Resolution Representation Learning for Visual Recognition"

必要前提

• Mindspore=2.0.0
• HRNet

使用方法

克隆该仓库:

git clone https://github.com/Calylyli/Mindsporehrnet.git

1. Cityscapes Dataset

• 准备数据集 (cityscapes).

├─ cityscapes                           # cityscapes数据集根目录
│   ├─ gtFine                           # 标签文件
│   │   ├─ train                        # 训练标签文件
│   │   │   └─ [city folders]
│   │   │       └─ [label images]
│   │   └─ val                          # 推理标签文件
│   │       └─ [city folders]
│   │           └─ [label images]
│   ├─ leftImg8bit                      # 图像文件
│   │   ├─ train                        # 训练图像文件
│   │   │   └─ [city folders]
│   │   │       └─ [images]
│   │   └─ val                          # 推理图像文件
│   │       └─ [city folders]
│   │           └─ [images]
│   ├─ train.lst                        # 训练样本列表
│   └─ val.lst                          # 推理样本列表

2. 训练

• 运行如下代码 :

python train.py

2. 导出MindIR

• 运行如下代码 :

python export.py --device_id [DEVICE_ID] --checkpoint_file [CKPT_PATH] --file_name [FILE_NAME] --file_format MINDIR --device_target Ascend --dataset [DATASET]

3. 评估

• 运行如下代码 :

sh scripts/run_eval.sh [DEVICE_ID] [DATASET_PATH] [CHECKPOINT_PATH]

4. 推理

• 运行如下代码 :

sh scripts/ascend310_inference.sh [MINDIR_PATH] [DATA_PATH] [DEVICE_ID]

References

• Link: Deep High-Resolution Representation Learning for Visual Recognition