drryanhuang / mmclassification

OpenMMLab Image Classification Toolbox and Benchmark

基础作业1: 花卉分类

配置文件为:

configs\convnext\convnext-tiny_32xb128_in1k_hw.py

有改动的数据配置文件为:

configs\_base_\datasets\imagenet_bs32_hw1_.py

我的数据 train.txt 和 val.txt 中有前缀:

hw1/flower_dataset\val\daisy\14816364517_2423021484_m.jpg 0
hw1/flower_dataset\val\daisy\7416083788_fcb4c4f27e_n.jpg 0
hw1/flower_dataset\val\daisy\18684594849_7dd3634f5e_n.jpg 0
hw1/flower_dataset\val\daisy\6136947177_47ff445eb4_n.jpg 0

于是我在 mmcls/datasets/pipelines/loading.py 添加了 ChangePathInWindows 类:

@PIPELINES.register_module()
class ChangePathInWindows(object):

    def __init__(self,
                 remove="hw1/flower_dataset/"):
        self.remove = remove


    def __call__(self, results):
        results['img_info']['filename'] = results['img_info']['filename'].replace("\\", '/')
        results['img_info']['filename'] = results['img_info']['filename'].replace(self.remove, "")
        return results

    def __repr__(self):
        repr_str = (f'{self.__class__.__name__}('
                    'Change "\\" => "/", ',
                    f"remove '{self.remove}')")
        return repr_str

训练 log 和 模型文件在 https://github.com/DrRyanHuang/mmclassification/releases/tag/HW1