search

google / stereo-magnification

Code accompanying the SIGGRAPH 2018 paper "Stereo Magnification: Learning View Synthesis using Multiplane Images"
https://people.eecs.berkeley.edu/~tinghuiz/projects/mpi/
Apache License 2.0
390 stars 87 forks source link

Fail to train VGG loss #30

Open yzxing87 opened 4 years ago

yzxing87 commented 4 years ago

Hi,

I can sucessfully train the pixel loss but failed at VGG loss (output images are black). The command I used to train VGG loss is

python train.py --cameras_glob=RealEstate10K/sample_dataset/camera_params/train/*txt \ --image_dir=RealEstate10K/sample_dataset/train --experiment_name=debug_vgg_loss --which_loss=vgg --batch_size=4 \ --vgg_model_file=imagenet-vgg-verydeep-19.mat --learning_rate=5e-5

The loss seems to descent normally. During the training process, most output images visualized by tensorboard are black while some are normal. I have tried to adjust different learning rates (1e-4, 5e-5, etc.) but it is still not working. Is there any special guideline to train the vgg loss? Thanks.

Richerhooders commented 1 year ago

I have the same problem as you. Have you solved it?