Ma Zerun

Here is the [colab tutorial](https://colab.research.google.com/github/open-mmlab/mmclassification/blob/master/docs/en/tutorials/MMClassification_tools.ipynb) Please notice the `Prepare data` section and the `Create a new config file` section

Looks it's a legacy problem. Welcome to change it to the `OptimizerHook` in mmcv, and we will test it.

Sorry but I think the problem is not use `pip` or `python3 -m pip` Looks like you have installed an extra pip for your user. And the path of this...

Good survey, we should split it to several tasks in our plan.

Well, looks the batch size in unit test is too large. I have created a pull request and you can test if it can solve your problem.

> 可以看一下经过处理后的ILSVRC 2012的meta文件夹内容吗，我没有使用datasetprepare中提到的处理ImageNet脚本进行处理 训练集上的 loss 有下降吗？

> 下降了 ``` imagenet/ ├── meta │   ├── test.txt │   ├── train.txt │   └── val.txt ├── train │   ├── n01440764 │   │   ├── n01440764_10026.JPEG │   │   ├── n01440764_10027.JPEG │   │  ...

Hello, what's the specific difference between PyTorch output and ONNX output?

> Ok, we will check the detailed difference and increase the tolerance of `np,allclose`. Thank you.

@ZwwWayne Any updates on this issue?