how to optimize the inference time about barbershop HOT 9 OPEN

Thanks to your advice, I could reduce the processing time to 1:40, around 2 minutes (Only align and blend), which is awesome. Is there any way to speed up more?
Nvidia Rtx 3070
Windows 11

from barbershop.

Dear @ZPdesu

Do you have any numbers on inference time for different hardware setups?

I ran the following command

python main.py --im_path1 90.png --im_path2 15.png --im_path3 117.png --sign realistic --smooth 5

the embedding process runs with ~1.6it/s on a Tesla K80 and takes about 8 minutes for default options for one image. How would one speed up inference times in general because in your paper you wrote that the overall process takes about 2 minutes [1] and I wondered If you could share some details on your hardware setup.

[1]: For each
composite image, we solve equation (6) and then (13) to generate a
composite image in an average time of two minutes

from barbershop.

It has already been implemented. For example, you can first use functions ii2s.invert_images_in_W() and ii2s.invert_images_in_FS() to invert all of your images . When you do the inference, you can comment these two lines in main.py.
Also, please git pull the newest repo for improvements and new features.

from barbershop.

It has already been implemented. For example, you can first use functions ii2s.invert_images_in_W() and ii2s.invert_images_in_FS() to invert all of your images . When you do the inference, you can comment these two lines in main.py. Also, please git pull the newest repo for improvements and new features.

First,thank you for your early reply.Second,there may be a problem with my statement that has misunderstood you(it`s my fault).For example,I want only transfer the style and appearance of image2(python image1 image2 image2 ) to the new another
image，and save the informations of the ref image(image2) of the three processes(embedding、alignment、blendding) in advanced.when I transfer the style and appearance of image2 to new user head image,can the transfer process whitout the
"pbar" for loop process of alignment,blendding,so the whole inference process time can decrease hugely when transfer new user image use the saved informations of the preprocessed ref image.Can this idea implement?

from barbershop.

It has already been implemented. For example, you can first use functions ii2s.invert_images_in_W() and ii2s.invert_images_in_FS() to invert all of your images . When you do the inference, you can comment these two lines in main.py. Also, please git pull the newest repo for improvements and new features.

First,thank you for your early reply.Second,there may be a problem with my statement that has misunderstood you(it`s my fault).For example,I want only transfer the style and appearance of image2(python image1 image2 image2 ) to the new another image，and save the informations of the ref image(image2) of the three processes(embedding、alignment、blendding) in advanced.when I transfer the style and appearance of image2 to new user head image,can the transfer process whitout the "pbar" for loop process of alignment,blendding,so the whole inference process time can decrease hugely when transfer new user image use the saved informations of the preprocessed ref image.Can this idea implement?

For now, only the embedding information can be reused.

from barbershop.

Dear @ZPdesu

Do you have any numbers on inference time for different hardware setups?

I ran the following command

python main.py --im_path1 90.png --im_path2 15.png --im_path3 117.png --sign realistic --smooth 5

the embedding process runs with ~1.6it/s on a Tesla K80 and takes about 8 minutes for default options for one image. How would one speed up inference times in general because in your paper you wrote that the overall process takes about 2 minutes [1] and I wondered If you could share some details on your hardware setup.

[1]: For each composite image, we solve equation (6) and then (13) to generate a composite image in an average time of two minutes

The total time is to add up all the parts for each image. The alignment and blending part does not require that much time. II2S embedding takes less than 3 minutes for one image on my RTX3090 machine. You can reuse the embedding results for different editing combinations.

from barbershop.

It has already been implemented. For example, you can first use functions ii2s.invert_images_in_W() and ii2s.invert_images_in_FS() to invert all of your images . When you do the inference, you can comment these two lines in main.py. Also, please git pull the newest repo for improvements and new features.

First,thank you for your early reply.Second,there may be a problem with my statement that has misunderstood you(it`s my fault).For example,I want only transfer the style and appearance of image2(python image1 image2 image2 ) to the new another image，and save the informations of the ref image(image2) of the three processes(embedding、alignment、blendding) in advanced.when I transfer the style and appearance of image2 to new user head image,can the transfer process whitout the "pbar" for loop process of alignment,blendding,so the whole inference process time can decrease hugely when transfer new user image use the saved informations of the preprocessed ref image.Can this idea implement?

For now, only the embedding information can be reused.

same issue

from barbershop.

Anyone know how to store embedding of input face for re-use latter?

from barbershop.

Thanks to your advice, I could reduce the processing time to 1:40, around 2 minutes (Only align and blend), which is awesome. Is there any way to speed up more?
Nvidia Rtx 3070
Windows 11

How were you able to do it? Can you elaborate on the steps please?

from barbershop.