预训练的模型没法直接用来测试

[xinntao]

预训练的模型没法直接用来测试

[qiuyajun]

@xinntao

[qiuyajun]

还有一个问题想请教您一下,就是假设训练用的HR图片的大小是124x124的,但数据集的图片都是1024x2048,这个时候应该怎么用1024x2048的图片的产生124x124,是用随机裁剪的方式吗,还是其他什么方式,您认为哪一种方式好一点. @xinntao

[xinntao]

Hi @qiuyajun
I have not encountered your problems. I saw that you use python2.7, and I am not sure whether it causes the problem. I've tested in Python 3 and PyTorch >= 0.4.0.

You can print more information to debug.
For example, printing the model parameters dict after constructing the model.

state_dict = model.state_dict()
for key, param in state_dict.items():
    print(key)

[xinntao]

@qiuyajun
For the data, I think random cropping is good. For faster IO, you can first crop to sub-images and use lmdb format. Some discussions can be found here.

[qiuyajun]

好的,谢谢您啦@xinntao