A saved GAN trained model for image generation does not generate the same accurate images when GPU is reset
When I train the flower image generation example, everything seems to go well as long as the GPU memory keeps the parameters loaded. I obtain images of easily recognizable flowers, as shown in the example. However, if I save the complete training workspace using the ‘save’ command (for example, save(‘GANWorkspacefile.mat’)), which also includes netG, then clear the GPU memory (reset), and subsequently load the previous workspace (load(‘GANWorkspacefile.mat’)), the images generated with ‘predict’ end up blurry—no flowers at all—resembling the ones generated at the beginning of training. The same issue occurs when I transfer the saved workspace and load it on another machine with the same version of MATLAB (R2022b). It seems that something is missing when loading the workspace variables that prevents generating the images in the same way as they are generated just at the end of training. I would appreciate it if someone has any idea of what I’m doing wrong could comment on it.
Thank you.When I train the flower image generation example, everything seems to go well as long as the GPU memory keeps the parameters loaded. I obtain images of easily recognizable flowers, as shown in the example. However, if I save the complete training workspace using the ‘save’ command (for example, save(‘GANWorkspacefile.mat’)), which also includes netG, then clear the GPU memory (reset), and subsequently load the previous workspace (load(‘GANWorkspacefile.mat’)), the images generated with ‘predict’ end up blurry—no flowers at all—resembling the ones generated at the beginning of training. The same issue occurs when I transfer the saved workspace and load it on another machine with the same version of MATLAB (R2022b). It seems that something is missing when loading the workspace variables that prevents generating the images in the same way as they are generated just at the end of training. I would appreciate it if someone has any idea of what I’m doing wrong could comment on it.
Thank you. When I train the flower image generation example, everything seems to go well as long as the GPU memory keeps the parameters loaded. I obtain images of easily recognizable flowers, as shown in the example. However, if I save the complete training workspace using the ‘save’ command (for example, save(‘GANWorkspacefile.mat’)), which also includes netG, then clear the GPU memory (reset), and subsequently load the previous workspace (load(‘GANWorkspacefile.mat’)), the images generated with ‘predict’ end up blurry—no flowers at all—resembling the ones generated at the beginning of training. The same issue occurs when I transfer the saved workspace and load it on another machine with the same version of MATLAB (R2022b). It seems that something is missing when loading the workspace variables that prevents generating the images in the same way as they are generated just at the end of training. I would appreciate it if someone has any idea of what I’m doing wrong could comment on it.
Thank you. deep learning, gan, matlab MATLAB Answers — New Questions