What's the configuration of pretrained model?

[FantasyJXF]

Thanks for your great work, it shows SOTA performance among the image2image GAN models. The attention mechanism is really good idea.

Here I have a question about your pretrained model:

1. What's the configuration of the training parameter? Did you use light?
2. It's there anyway to accelerate the trainning process？
3. How to observe the model when it is trainning? Since it takes lifelong to train 100 epoches.

I downloaded your selfie2anime dataset and replaced some low quality trainA images, and trained the model with Pytorch code, without light and use image size 128, with GPU Tesla P40, it took me 7 days to compelete the trainning, but the performance of my model was not so good as yours.

Looking forward to your reply.

[FantasyJXF]

The pretrained model file was saved 10 monthes ago, and the code in this repo is still updated recently, so my question is how could I get the similar performance as you did with the CODE and the DATASET you provide.

The models I trained out are really frastrating.

Thanks again.

[alloin]

light false gan_type lsgan n_res 4 n_dis 6 batch_size 1 adv_weight 1 adv_weight 10 cycle_weight 10 cam_weight 1000 smoothing true

If you want to observe, print_freq will print samples in the samples folder every x steps

I'm currently training a light version of this on a gtx 1080 ti (on epoch 60, been training for over 2 weeks now) with an extended dataset and results are looking better than the pretrained model already.

[FantasyJXF]

@alloin The config you use looks like the same with the original code. I've trained the model use the 3400 pic dataset Author provided, and I use V100 to train, it tooks 10 minutes to produce one sample, after trained 100 epoches, the result still couldn't be compared with the Author's.

As you said, you didn't change any thing of the model architecture and just expand the dataset to train? So I'd like to ask how did you create your dataset? How many pictures you used in trainA?

Thanks again.

[alloin]

@alloin The config you use looks like the same with the original code. I've trained the model use the 3400 pic dataset Author provided, and I use V100 to train, it tooks 10 minutes to produce one sample, after trained 100 epoches, the result still couldn't be compared with the Author's.

As you said, you didn't change any thing of the model architecture and just expand the dataset to train? So I'd like to ask how did you create your dataset? How many pictures you used in trainA?

Thanks again.

I use about 60.000 pictures in trainA and 62.000 in trainB

[FantasyJXF]

You mean you use 60,000(sixty thousands) pictures in each training directory to train? That's really lots of pictures for GAN. It should be very very slow to train.

1. Did you choose the selfie picture from dataset the author use?
2. Why don't you conduct the transfer Learning from the pretrained model?
3. If you don't mind, could you please show some test results of your model?

Thanks

[alloin]

You mean you use 60,000(sixty thousands) pictures in each training directory to train? That's really lots of pictures for GAN. It should be very very slow to train.

1. Did you choose the selfie picture from dataset the author use? 2. Why don't you conduct the transfer Learning from the pretrained model? 3. If you don't mind, could you please show some test results of your model? Thanks

sixty thousands in each directory yes, training takes ages indeed, been training for weeks on a gtx 1080 ti now. I can't transfer the pretrained model because my gtx 1080 ti runs out of memory on that one, hence I use Light training and had to start over. My data comes from a variety of sources, like https://www.crcv.ucf.edu/data/Selfie/ and https://www.gwern.net/Danbooru2018 and processed with some python scripts.

Here's a video comparing the pretrained model (100 epoch) and my model (at 40 epoch in this video) https://www.youtube.com/watch?v=Dx7CfykIyzU

Another video: https://www.youtube.com/watch?v=lgkKz7KQF_E

My own model is at epoch 65 as we speak, will probably test it soon again.

[FantasyJXF]

It's amazing that you test the model on video, I was used to think the model is for selfie to anime image transformation.

Here I got 2 questions:

1. As you said before, you used light=False to train your model, and then you said hence I use Light training and had to start over. I wonder did you use light on the model in the video?
2. I watched the video you release, as far as I can say, the performance of the model is not so good to see, how did you evaluate the model, KID?

I really think some results in the paper was fascinating, but can't be repeated. What a pity.

[alloin]

1. As you said before, you used light=False to train your model, and then you said hence I use Light training and had to start over. I wonder did you use light on the model in the video?

Yes, both video's are my own trained 'light' model

2. I watched the video you release, as far as I can say, the performance of the model is not so good to see, how did you evaluate the model, KID?

I evaluate with my eyes, also asked a dozen of people and all found that the overall result is better on my model. There's a big difference in both models though, mine uses more pictures 'chest to head', while the original model is 'heads only'.

I really think some results in the paper was fascinating, but can't be repeated. What a pity.

I would experience more if I had access to a better GPU, takes over a month to train on my GTX 1080 ti or on google colab.

[FantasyJXF]

I do envy your patient, I tried to select thousands of head pictures to retrained the model with all my GPUs, it took me weeks to train, and I couldn't do anything during training.

May be it's more challenging to do the half body image transformation.

Did you test your model with pictures ?

[xyxxmb]

1. As you said before, you used light=False to train your model, and then you said hence I use Light training and had to start over. I wonder did you use light on the model in the video?

Yes, both video's are my own trained 'light' model

1. I watched the video you release, as far as I can say, the performance of the model is not so good to see, how did you evaluate the model, KID?

I evaluate with my eyes, also asked a dozen of people and all found that the overall result is better on my model. There's a big difference in both models though, mine uses more pictures 'chest to head', while the original model is 'heads only'.

I really think some results in the paper was fascinating, but can't be repeated. What a pity.

I would experience more if I had access to a better GPU, takes over a month to train on my GTX 1080 ti or on google colab.

Could you share your pre-trained model? I want to test it with myself model. Contact me with my mail: echobingo@foxmail.com. Thank you.

[tankfly2014]

1. As you said before, you used light=False to train your model, and then you said hence I use Light training and had to start over. I wonder did you use light on the model in the video?

Yes, both video's are my own trained 'light' model

1. I watched the video you release, as far as I can say, the performance of the model is not so good to see, how did you evaluate the model, KID?

I evaluate with my eyes, also asked a dozen of people and all found that the overall result is better on my model. There's a big difference in both models though, mine uses more pictures 'chest to head', while the original model is 'heads only'.

I really think some results in the paper was fascinating, but can't be repeated. What a pity.

I would experience more if I had access to a better GPU, takes over a month to train on my GTX 1080 ti or on google colab.

Could you share your pre-trained model? I want to test it with myself model. Contact me with my mail: echobingo@foxmail.com. Thank you.

me too. explorer0902(#)gmail.com