search

SerialLain3170 / adeleine

Automatic line art colorization using various types of hint or without hint
Other
355 stars 47 forks source link

The attn_type in param.yaml of reference_adain #39

Closed humuhumu39 closed 1 year ago

humuhumu39 commented 1 year ago

Is the attn_type:"sa" in model.generator wrong option for training adain. I think it shoud be attn_type:"adain". Am I wrong?

humuhumu39 commented 1 year ago

oh, and how much GPU memory do you have?

I'm using RTX 3060 with 6GB memory. When I'm training some models I always need to set batchsize very small, like adain model I can only train in batchsize:2, train_size:256 and I even can NOT train scft model in batchsize:1, train_size:256.

It's sad, because I want to try scft model. ;(

SerialLain3170 commented 1 year ago

I am very sorry for the late response.

Is the attn_type:"sa" in model.generator wrong option for training adain. I think it shoud be attn_type:"adain". Am I wrong?

You are right, and I am sorry for the confusion.

oh, and how much GPU memory do you have?

I trained my model with GPU whose memory is 24GB. If you would like to train SCFT model with RTX3060, I recommend that you decrease the train_size by changing 256 into 128.

humuhumu39 commented 1 year ago

Thank you for replying!

I haven't trained adain model yet. But the SCFT model, when I decrease the train_size to 128 that would cause a error. It seems because instance_norm needs the Width and Height greater than 1, so eventually I upgraded my GPU to RTX3070 with 8GB memory, and train in batchsize: 1, train_size: 256. Now I'm waiting the result, appreciate your help again. :)