Questions about using the given dataset for training（such as miniImagenet） and using my own dataset for testing!

[MasterWinston]

Hello, as described in the title, how do I handle my own dataset after training with the miniImagenet dataset you have given ? For example,my dataset has 4 classes and each class has about 100 images,
Do I need to process my dataset into the following format?

MyDataset_folder/
├── images/
│ ├── images_1.jpg
│ ├── ...
│ └── images_n.jpg
├── train.csv *
├── test.csv *
└── val.csv *

Or do I need to put all my datasets into the miniImagenet dataset and create a new test.csv to replace the original test.csv in the miniImagenet dataset?

Or other methods?

Look forward to your reply, thank you!

[wZuck]

Yes, you need to process your dataset in such a format and set the data_root to the folder. Notice that the format of your *.csv should be the same as miniimagenet. When training/testing on your dataset, you should make sure your dataset can handle your setting.(for example, if there are only 4 classes in your dataset, you should set way_num <=4)

Another solution is to modify the core/data/dataset.py to read your dataset.

[MasterWinston]

其实我是想试试先用miniImagenet训练得到一个不错的model，然后再用这个model在自己的数据集上测试。

另外我刚跑run_test.py代码报了如下错误：
Traceback (most recent call last):
File "E:/Desktop/fewshot/LibFewShot/run_test.py", line 27, in
main(0, config)
File "E:/Desktop/fewshot/LibFewShot/run_test.py", line 22, in main
test.test_loop()
File "E:\Desktop\fewshot\LibFewShot\core\test.py", line 61, in test_loop
_, accuracies = self._validate(epoch_idx)
File "E:\Desktop\fewshot\LibFewShot\core\test.py", line 104, in _validate
prefetcher = data_prefetcher(self.test_loader)
File "E:\Desktop\fewshot\LibFewShot\core\utils\utils.py", line 303, in init
self.preload()
File "E:\Desktop\fewshot\LibFewShot\core\utils\utils.py", line 313, in preload
self.next_data = [data.cuda(non_blocking=True) for data in self.next_data]
File "E:\Desktop\fewshot\LibFewShot\core\utils\utils.py", line 313, in
self.next_data = [data.cuda(non_blocking=True) for data in self.next_data]
AttributeError: 'list' object has no attribute 'cuda'

请问这是什么原因呢？
ps：run_trainer.py已成功跑完。

[wZuck]

请重新pull一下代码，这个问题前几天修复了一下

[MasterWinston]

您好，我又有新问题了：）。
我想知道MAML方法外层循环中的θ参数更新时对应的代码位置。我能找到的只有MAML.py里的set_forward_adaptation函数，它似乎是在每个episode后对内层参数的更新，得到多个θ' 。我并没有找到根据多个θ'来对θ进行更新的代码块。

Hope your reply！
Thanks！

[yangcedrus]

MAML的过程内循环更新θ'，外循环通过θ'更新θ。

其中内循环通过手动更新的方式进行的原因是如果内循环使用pytorch的optimizer进行更新，pytorch不可避免地会进行梯度释放等操作，难以处理。但是外循环θ的更新，可以直接使用pytorch的optimizer进行更新，这一步和其他的方法并没有什么不同，因此封装在了trainer.py中。

因此maml.py中并没有外循环更新的代码。