text2live / Text2LIVE-main /configs /video_config.yaml
SupermanxKiaski's picture
Upload 356 files
16d007c
raw
history blame
No virus
1.93 kB
seed: -1
device: cuda
use_wandb: False
wandb_project: project_name # enter your project name here
wandb_entity: project_entity # enter your entity here
results_folder: results # if not use wandb, this is the folder where the results will be saved
resx: 768
resy: 432
example_config: "car-turn_winter.yaml"
save_model_starting_epoch: 900
multiply_foreground_alpha: True
flip_p: 0.5 # probability of applying flip before cnn
n_aug: 6 # set to -1 to disable augmentation before CLIP
clip_affine_transform_fill: 0 # 0 for black, 1 for white
clip_model_name: "ViT-B/32" # ViT-B/16 | ViT-B/32 | ViT-L/14
text_criterion: spherical # spherical | cosine | scaled_cosine (*1.2)
bootstrap_text: ""
bootstrap_scheduler: none
bootstrap_epoch: -1 # epoch to stop penalizing sparsity
use_negative_bootstrap: False # whether to use negative relevance
lambda_bootstrap_min: 0
bootstrap_negative_text: [] # negative alpha - will ignore this
bootstrap_negative_map_threshold: 0.6 # penalizing only locations with high values in relevancy
bootstrapping_min_cover: 1
relevancy_num_layers: 10
lambda_screen: 1
lambda_sparsity: 0.1 # lambda_sparsity * ( lambda_alpha_l0 * L0_loss + lambda_alpha_l1 * L1_loss )
lambda_alpha_l0: 0.005
lambda_alpha_l1: 0.01
lambda_structure: 3
lambda_bootstrap: 10
lambda_clip: 1 # lambda_clip * ( lambda_comp_clip * L_comp + lambda_layer_clip * L_layer )
lambda_composition: 1
n_epochs: 3000
gamma: 0.999
min_lr: 0.00001
lr: 0.0025
optimizer: madgrad # [adam | radam | rmsprop | sgd]
# the following is relevant only for dip_backbones backbone
skip_n33d: 128
skip_n33u: 128
skip_n11: 4
num_scales: 7
log_images_freq: 500
center_frame_distance: 2
input_entire_atlas: True
entire_atlas_every: 75
return_atlas_alpha: False
grid_atlas_resolution: 2000
align_corners: False
crops_min_cover: 0.95 # 0.95 for foreground, 0.8 for background
masks_border_expansion: 30
mask_alpha_threshold: 0.95