如何复现mmgeneration 中的 cyclegan 模型结果

[lovejing0306]

你好，我如下配置文件，在 mmgen 中尝试复现你们给出的开源模型结果，但失败了。

我的训练配置信息如下：
配置文件使用是：https://github.com/open-mmlab/mmgeneration/blob/master/configs/cyclegan/cyclegan_lsgan_resnet_in_horse2zebra_b1x1_270k.py

显卡型号：1080 ti
gpu 个数：4

下面是模型训练完之后使用 best 模型在 FID 和 IS 的评估结果

下面这个是官方给出的开源模型评估结果

从指标上看，自己训练的和官方开源的差异挺大。

请问你们是在什么显卡上用多少个 gpu 训练出了上面的结果？

[plyfager]

Hey @lovejing0306 thanks for your interest in our project.
You can try to train cyclegan with 1 GPU. The name field b1x1_270k means batch_size equals to 1, number of gpu equals to 1.

[lovejing0306]

Hey @lovejing0306 thanks for your interest in our project. You can try to train cyclegan with 1 GPU. The name field b1x1_270k means batch_size equals to 1, number of gpu equals to 1.

ok, thanks. I will try it later .

[lovejing0306]

Hey @lovejing0306 thanks for your interest in our project. You can try to train cyclegan with 1 GPU. The name field b1x1_270k means batch_size equals to 1, number of gpu equals to 1.

I try to train cyclegan with 1 GPU and batch size 1 , but the result is different from yours.

The best fid is iter_30000, but your best fid is 266800. Can you share your train log ?

[plyfager]

I suggest you reduce the eval interval to 100.

mmgeneration/configs/cyclegan/cyclegan_lsgan_resnet_in_horse2zebra_b1x1_270k.py

Line 164 in b5ed84d

interval=10000,

Since the test dataset is very small, the fid results can vary a lot due to different initialization weights. I think the training difference can be ignored.

[lovejing0306]

Since the test dataset is very sma

Thanks for your advice, I will set the interval=100 and try it.