search

junyanz / CycleGAN

Software that can generate photos from paintings, turn horses into zebras, perform style transfer, and more.
Other
12.22k stars 1.94k forks source link

How many epochs were required to create a model "horse2zebra"? #111

Closed bumsun closed 4 years ago

bumsun commented 5 years ago

How many epochs were required to create a model "horse2zebra"?

junyanz commented 5 years ago

200 epochs in total. We keep the same learning rate for the first 100 epochs and linearly decay the rate to zero over the next 100 epochs. See the appendix of our paper for more details.

3DTOPO commented 5 years ago

Is there a way to achieve that with the options? I see the lr_policy flag is set to linear by default. But how do you keep it steady for the first 100 epochs?

Do you run 10,000 iterations per epoch?

junyanz commented 5 years ago

We go through the entire dataset per epoch. If there are 1000 images, the model will be trained for 1000 iterations per epoch. Regarding the flag 'linear' (the name is a little bit confusing). In our implementation, we keep the same learning rate for the first 100 epochs and linearly decay the rate to zero over the next 100 epochs.

surajsubramanian commented 4 years ago

After how many epochs will we start to get reasonable images, will the images be just random noise for say the first 5 epochs

junyanz commented 4 years ago

around 50 epochs (for horse2zebra). It depends on the dataset.

surajsubramanian commented 4 years ago

Thanks for replying I am working on a project to convert a normal image of a person into an image in which the person is smiling. I have 1200 images of people not smiling and another 1500 images of people smiling.

Also will I be able to get good results if I use the Google colab version of your code by just changing the dataset folder.

Again thank you so much :)

youjin-c commented 1 year ago

Hello, @SurajSubramanian
Are they not smiling and smiling images matching? I mean, do I need to prepare a set of smiling and not smiling faces of the same person? Or just random smiling and not smiling faces work?

surajsubramanian commented 1 year ago

@youjinChung You don't need to have paired/matching images, that's the point of CycleGAN. i.e. Random images from both sets should work.

youjin-c commented 1 year ago

@SurajSubramanian That is so cool. Thanks for the explanation, Suraj. Let me try it right away.