hyperparams for training on 1200 labels

[fengtingl]

Hi, I enjoyed the paper and was able to run the training example.
I set up the CelebA dataset with 1200 identities provided in legit_indices.npy. However, I am not sure about the hyper-parameter setting:

1. The learning rate is 0.0001 in train.py but 0.00001 in the paper(section 4.1)
2. The iterations of the critic, generator, and siamese are 5, 1, 1 in train.py, and 1, 3, 1 in run_training.py.
3. filter number, batch size, etc

Could you give me the detailed hyper-parameters for training on the dataset with 1200 identities? Thanks!

The current hyper-parameters I'm using, which yield an unsatisfactory result:
'TRAIN_PARAMS': {
'ARCH_NUM': 'unet_flex',
'ARCH_SIAM': 'resnet_siam',
'EPOCH_START': 0,
'EPOCHS_NUM': 120,
'LEARNING_RATE': 0.00001,
'FILTER_NUM': 32,
'ITER_CRITIC': 1,
'ITER_GENERATOR': 3,
'ITER_SIAMESE': 1,
'GAN_TYPE': 'lsgan',
'FLAG_SIAM_MASK': False,
},
'DATA_PARAMS':{
'LABEL_NUM': 1200,
'WORKERS_NUM': 4,
'BATCH_SIZE': 32,
'IMG_SIZE': 128,
'FLAG_DATA_AUGM': True,
},
'OUTPUT_PARAMS': {
'SAVE_EPOCH': 1,
'SAVE_CHECKPOINT': 60,
'LOG_ITER': 2,
'COMMENT': "Something here",
'EXP_TRY': 'check',
}

[ldz666666]

I have the similar problem , have you find the proper params yet?

[fengtingl]

Not yet.
For the best result so far, I use a Proxy-NCA pre-trained network, which achieves a recall@1 rate of around 81%, as the identity discriminator. Then I extend the epochs to 300. Batch size and learning rate are set to be 16 and 1e-5, respectively. The visual quality is not satisfactory, with an FID rate of around 5.5, though. Keep me posted if you have better results.

[skye95git]

Hello, I'm also trying to train with 1200 identities，but I have a few problems：

1. There are 1563 pieces of data in legit_indices.npy, how do you get the 1,200 identities the author refers to?
2. How do you generate new images using a trained model?
3. I run the training example, but the result is not satisfactory. Is it the lack of training data?
   

[fengtingl]

1. I tried two initializations, with the first 1200 identities and a random sampled 1200 identities. The results are not very much distinguishable.
2. I simply change the forward conditional information into a random one-hot vector.
3. You could try the following options, but I'm not sure if there could be more improvements with more training data.
   (1) since the training of a GAN is not stable, try multiple initializations.
   (2) use larger training epochs.

Hello, I'm also trying to train with 1200 identities，but I have a few problems：

1. There are 1563 pieces of data in legit_indices.npy, how do you get the 1,200 identities the author refers to?
2. How do you generate new images using a trained model?
3. I run the training example, but the result is not satisfactory. Is it the lack of training data?
   

[fengtingl]

Thanks for your reply. I use batch size 16 and learning rate 1e-5 here. Could you provide a pre-trained model?

…

________________________________ From: Ismail Elezi <notifications@github.com> Sent: Wednesday, December 16, 2020 22:38 To: dvl-tum/ciagan <ciagan@noreply.github.com> Cc: fengtingl <fengtingl@hotmail.com>; Author <author@noreply.github.com> Subject: Re: [dvl-tum/ciagan] hyperparams for training on 1200 labels (#5) What batch size you are using? We used batch size of 8 and 16 for the results in the paper. We'll try to put the configuration and an example for the generation before the holiday, but slightly busy with current work. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#5 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AIQR347PS4KLQVDYA43LGMLSVDA7NANCNFSM4SRYQYTA>.

[skye95git]

1. I tried two initializations, with the first 1200 identities and a random sampled 1200 identities. The results are not very much distinguishable.
2. I simply change the forward conditional information into a random one-hot vector.
3. You could try the following options, but I'm not sure if there could be more improvements with more training data.
   (1) since the training of a GAN is not stable, try multiple initializations.
   (2) use larger training epochs.

Hello, I'm also trying to train with 1200 identities，but I have a few problems：

1. There are 1563 pieces of data in legit_indices.npy, how do you get the 1,200 identities the author refers to?
2. How do you generate new images using a trained model?
3. I run the training example, but the result is not satisfactory. Is it the lack of training data?
   

Thanks for your reply. I don't understand how to change the forward conditional information into a random one-hot vector. I have tried many times to load the trained model to generate new images, but all reported errors. Could you provide an example or share your code? Thanks.