StarGAN-JTN training CelebA and RaFD facial expression synthesis

[tsly123]

Hi, Thank you for sharing your work @yunjey

I am training StarGAN-JNT CelebA and RaFD for facial expression synthesis on CelebA. The training hasn't done yet (110k/200k iters) but the facial expression synthesis task seems not to converge. From the image below, while the attribute outputs are good, at least you can see the attribute in the output, but the facial expression synthesis outputs (last 7 columns) are not much different. This could be due to the size of RaFD dataset (4,824 samples as in the paper, 7 labels without contempt).

python main.py --mode=train --dataset Both --image_size 256 --c_dim 5 --c2_dim 7 \
                 --sample_dir stargan_both/samples --log_dir stargan_both/logs \
                 --model_save_dir stargan_both/models --result_dir stargan_both/results

What is the solution to this, more training? How to do facial expression synthesis on celebA as in the paper? Could you share the StarGAN-JNT pre-trained model to my email 176064 at jnu dot ac dot kr?

I run several versions of StarGAN-JNT and ogange one is 110k/200k iter which gives the above result.

Thank you very much. tsly

[ujjawalcse]

Hi @tsly123 ! Have you got the pretrained model you asked for? Can you share me also if you got it?

[Mouna-Dhaouadi]

Hi @tsly123

In case you have got the pretrained mode you; can you share it ? : mounaadhaouadi@gmail.com Thanks you

[alarazin]

I have the same problem @tsly123. Could you solve it? Thanks

[ShisuiUzumaki]

@tsly123 If you got the pre-trained model, could you share it at bsef17m051@pucit.edu.pk

[SonNNguyen]

@tsly123 Im working on my school project could you share the pre-trained model for RaFD, my email is sonnn.m21ict@st.usth.edu.vn . Thank you very much!