vid2vid + training + test code debugging (debug + train + test) (iii) training in their own data sets

 Man of few words said on the code.

# train on 256_g1

$ python train.py --name <任务名>_256_g1 --input_nc 3 --loadSize 256 --n_downsample_G 2 --num_D 1 --max_frames_per_gpu 4 --n_frames_total 15 --dataroot datasets/<我的数据集>/ --save_epoch_freq 10 --gpu_ids=2

# train on 512_g1

$ python train.py --name <任务名>_512_g1 --input_nc 3 --loadSize 512 --n_scales_spatial 2  --n_downsample_G 2 --max_frames_per_gpu 1 --n_frames_total 4 --niter_step 2 --niter_fix_global 8 --niter_decay 5 --load_pretrain checkpoints/<任务名>_256_g1 --dataroot datasets/<我的数据集>/ --save_epoch_freq 10 --gpu_ids 2

# train on 1024_g1

$ python train.py --name <任务名>_1024_g1 --input_nc 3 --loadSize 896 --n_scales_spatial 3 --n_frames_D 2 --n_downsample_G 2 --num_D 3 --max_frames_per_gpu 1 --n_frames_total 4 --niter_step 2 --niter_fix_global 8 --niter_decay 5 --load_pretrain checkpoints/<任务名>_512_g1 --lr 0.0001 --dataroot datasets/<我的数据集>/ --save_epoch_freq 8 --gpu_ids 2

# test

# test on 512_g1
$ python test.py --name <任务名>_512_g1 --input_nc 3 --loadSize 512 --n_scales_spatial 2 --n_downsample_G 2  --gpu_ids=3 --dataroot datasets/<我的数据集>/ --no_first_img

# test on 1024_g1
$ python test.py --name <任务名>_1024_g1 --input_nc 3 --loadSize 1024 --n_scales_spatial 3 --n_downsample_G 2  --gpu_ids 2 --dataroot datasets/<我的数据集>/ --no_first_img

Remarks:

1) We are on a single GPU to do the training;

2) command to be executed sequentially;

Preparation 3) sets a reference data format street;

4) Training 1024 when, if the card is 11G, the second update parameters of training time may be running out of memory, but as long as the set "save_epoch_freq" smaller so that it can retain at least once (but can not guarantee the results of the training);

Within 5) <xxx> represents this part requires custom, no angle brackets after finished.

""" Dataset 结构图 """
"""
datasets/my_dataset_name
    ├── train_A
    │   ├── seq_000000 
    │   ├── seq_000001
    │   └── ...
    │   
    ├── train_B
    │   ├── seq_000000 
    │   ├── seq_000001
    │   └── ...
    │
    ├── test_A
    │   ├── seq_000000 
    │   ├── seq_000001
    │   └── ...
    │   
    └── test_B
        ├── seq_000000 
        ├── seq_000001
        └── ...
"""

 So far, the study of vid2vid come to an end, the next phase of the study can compare its experiment!

Zhu Junan!