Training a face model with train_g1_256.sh

[petergerten]

Did anyone successfully trained a model with train_g1_256.sh?
The readme says that single GPU models were to well tested.
For some reasons the training finishes after only a few hours and trying to test I get:

ISSUE: Pretrained network G0 has fewer layers; The following are not initialized:
['model_down_img', 'model_down_seg', 'model_final_flow', 'model_final_img', 'model_final_w', 'model_res_flow', 'model_res_img', 'model_up_flow', 'model_up_img']

[tcwang0509]

Did you make sure you downloaded the entire dataset? It requires external download and is not supported by the download script.

[petergerten]

@tcwang0509 maybe I have some files missing. But if they are not downloaded by the download script - where do I get them?

[tcwang0509]

For face videos you need to download faceForensics,
as specified in https://github.com/NVIDIA/vid2vid#dataset.

[petergerten]

@tcwang0509 I thought that I would only need the faceForensics dataset if I want to reproduce the results. Am I wrong? I used different training data (and getting the above error).

[jakeelwes]

Did this get solved? I'm also having problems training with my own data -
It looks to be training very quickly with no errors (15 secs per epoch) but won't output anything in logs or web dirs (--dataset_mode pose worked fine for me on my own data).

My dataset is ~10,000 .jpg files in train_img/00001/ and corresponding .txt files in train_keypoints/00001/ and looks identical to the sample example data from python scripts/download_datasets.py. Running the training script on this same data also gives the same result. Is there something I'm missing about the structure of faceForensics dataset?

python train.py --name test \
--dataroot datasets/test/ --dataset_mode face \
--n_scales_spatial 2 --num_D 3 \
--input_nc 15 --loadSize 512 --ngf 64 \
--tf_log --display_freq 20