Training problem help about neuface HOT 10 CLOSED

Could you provide more details about your training configuration and results, such as the config file, the number of GPUs, and the rendered images?

from neuface.

Could you provide more details about your training configuration and results, such as the config file, the number of GPUs, and the rendered images?

Thank you for your reply. The config file is almost the same as the file "train_neuface.yaml", except that param 'num_gpus' is set to be 1, because I have only one nvidia 3090D chip. Two images in the tensorboard, are shown below:

The evaluation results (id=1, exp=2, image:045)are shown below:

from neuface.

If you are using a single GPU to train NeuFace, it is recommended to increase the num_pixels in the config file. Note that the PSNR shown in TensorBoard is not accurate due to the calibration code. For an accurate PSNR evaluation, please run eval.py. Additionally, as observed in TensorBoard, the render result appears much better than the evaluation. This might be because the wrong ckpt was used when running eval.py.

from neuface.

If you are using a single GPU to train NeuFace, it is recommended to increase the num_pixels in the config file. Note that the PSNR shown in TensorBoard is not accurate due to the calibration code. For an accurate PSNR evaluation, please run eval.py. Additionally, as observed in TensorBoard, the render result appears much better than the evaluation. This might be because the wrong ckpt was used when running eval.py.

Hi, when I used 4 GPU(nvidia 4090) to train NueFace, the PSNR is 25.18, still too lower. I increase the 'num_pixels' to 6000 in the config file, and get PSRN about 26. the result is below:
PSNR:26

from neuface.

Could you please provide the PSNR in detail for each test image?

from neuface.

Could you please provide the PSNR in detail for each test image?

the PSNR of images 47.png, 49.png, 50.png（large rotation）, are small, below 17, others are above 24. And only image of 52.png , is above 33.

in facescape_multi_dataset.py, ply_path is the TU model and not the cut mesh?
I place TU model in the ply_path as suggested.

from neuface.

'ply_path' refers to the original TU model, which provides the position of nose.

from neuface.

'ply_path' refers to the original TU model, which provides the position of nose.

Hi, I have trained a model in 4 V800 GPUs, and get PSNR 33.28 in train set, 26.87 in eval set, with num_pixels: 20000 and sdf_threshold:5e-5.

from neuface.

Can you send me your results via email? ([email protected])

from neuface.

Can you send me your results via email? ([email protected])

Have send it to you. If not received, pls let me know.

from neuface.