DL-Art-School/recipes/esrgan/rrdb_process_video.yml

69 lines
1.9 KiB
YAML
Raw Normal View History

2020-12-20 18:50:31 +00:00
name: video_process
suffix: ~ # add suffix to saved images
model: extensibletrainer
scale: 4
gpu_ids: [0]
fp16: true
minivid_crf: 12 # Defines the 'crf' output video quality parameter fed to FFMPEG
frames_per_mini_vid: 360 # How many frames to process before generating a small video segment. Used to reduce number of images you must store to convert an entire video.
minivid_start_no: 360
recurrent_mode: false
dataset:
n_workers: 1
name: myvideo
video_file: <your path> # <-- Path to your video file here. any format supported by ffmpeg works.
frame_rate: 30 # Set to the frame rate of your video.
start_at_seconds: 0 # Set this if you want to start somewhere other than the beginning of the video.
end_at_seconds: 5000 # Set to the time you want to stop at.
batch_size: 1 # Set to the number of frames to convert at once. Larger batches provide a modest performance increase.
vertical_splits: 1 # Used for 3d binocular videos. Leave at 1.
force_multiple: 1
#### network structures
networks:
generator:
type: generator
which_model_G: RRDBNet
in_nc: 3
out_nc: 3
initial_stride: 1
nf: 64
nb: 23
scale: 4
blocks_per_checkpoint: 3
#### path
path:
pretrain_model_generator: <your path> # <-- Set your generator path here.
steps:
generator:
training: generator
generator: generator
# Optimizer params. Not used, but currently required to initialize ExtensibleTrainer, even in eval mode.
lr: !!float 5e-6
weight_decay: 0
beta1: 0.9
beta2: 0.99
injectors:
gen_inj:
type: generator
generator: generator
in: lq
out: gen
# Train section is required, even though we are just evaluating.
train:
niter: 500000
warmup_iter: -1
mega_batch_factor: 1
val_freq: 500
default_lr_scheme: MultiStepLR
gen_lr_steps: [20000, 40000, 80000, 100000, 140000, 180000]
lr_gamma: 0.5
eval:
output_state: gen