Pretrained models

[TylerCarberry]

Do you have any CycleGAN pretrained models available that use pytorch?

[junyanz]

We are training some models now, and we will upload them soon.

[TylerCarberry]

Any update on this?

[petergerten]

I would also be very interested? Anybody has a pretrained model?

[jingweiz]

Hey,
do you have any training curves? I tried your code and the results do not look too good ... would be great if you can provide the training curves for a comparison, thanks a lot!

[AndreJFBico]

Any update?

[rednithin]

Where can I find a pre trained model for edges2shoes and edges2handbags?

[isaacmg]

Anyone have the monet2photo pre-trained model?

[hugozanini]

Did someone train the network for segmentation?

[junyanz]

We have a few pre-trained models available. More are coming.
pix2pix: edges2shoes, sat2map, and facades_label2photo
CycleGAN: horse2zebra and zebra2horse

Please check out the download_scripts, instructions(pix2pix) and instructions(CyclceGAN) for more details.

[linkerzhang]

Hi @junyanz I'm trying to load the pretrained cyclegan (horse2zebra) model as below code template,

model = XXXModel()
model.load_state_dict(torch.load('latest_net_G.pth'))

dumb question please. which XXXModel should I use in your code base please? I'm assuming that it should be a child class of torch.nn.Module.

or could you save and share the whole model for reference please? :)

Thank you very much!

Forget about it. Figured out from the test_model.py. :)

[al3xsh]

Hi,
I wondered if there was any progress on adding pretrainined models for photo 2 monet/cezanne/vangogh. I have trained cyclegan myself with the default parameters, but wasn't 100% happy with the results and would like to see the results of your models for comparison!
Regards,
Alex

[junyanz]

we don't have these models available. You can improve the results with additional identity loss.

[wnls]

Hi,
Since there's no pretrained model for sat2map, I'm using default params to train sat2map by running
python train.py --dataroot ./datasets/maps --name maps_cyclegan --model cycle_gan --no_dropout
as suggested in README.

The generated images don't seem to have the quality it should at epoch 100+. Here's a generated image at epoch 151. Am I missing something? Thank you!

(epoch151_fake_A.png, epoch151_fake_B.png)

[chandakQ]

Could you please add a pre-trained model for apple2orange.

Regards,
Mukesh Chandak

[junyanz]

We will add more models soon. @taesung89

[taesungp]

Added! Please use the script pretrained_models/download_cyclegan_model.sh to download the pretrained models.

[joeatton]

Do you have the pretrained pix2pix model available for colourisation

[junyanz]

We don't' have one for pix2pix. Maybe you can check out this repo.

[willhunger]

@wnls I caught the same problem, how did you slove it?

[DmitryUnanyants]

Hi @junyanz I'm trying to load the pretrained cyclegan (horse2zebra) model as below code template,

model = XXXModel()
model.load_state_dict(torch.load('latest_net_G.pth'))

dumb question please. which XXXModel should I use in your code base please? I'm assuming that it should be a child class of torch.nn.Module.

or could you save and share the whole model for reference please? :)

Thank you very much!

Forget about it. Figured out from the test_model.py. :)

Can you write please how did you solve this problem?

[ZHE-SAPI]

Hi, @junyanz , when I click on the following URL https://people.eecs.berkeley.edu/~taesung_park/pytorch-CycleGAN-and-pix2pix/models/$FILE.pth, it shows "Resource not found".
Could you please upload the pretrained models and related datasets again? Thanks!

[taesungp]

Hello,

we've updated the URLs to the pretrained models. The models are now at http://efrosgans.eecs.berkeley.edu/cyclegan/pretrained_models/.

[sameerjvn]

Hi @junyanz @taesungp, I would like to compare the performance of the CycleGANs with UNET generator v/s ResNet generator. I had a few questions:

1. Do you have training details of your pre-trained models? (How many epochs, loss curves, etc.)
2. Since the CycleGAN class is using the ResNet 9-blocks generator by default, I am assuming that the pre-trained models are for this generator. In such case, Do you have pre-trained models using the UNET generator? (256 or 128)