text2live / configs /image_config.yaml
SupermanxKiaski's picture
Upload 351 files
3b40f46
seed: -1
use_wandb: False
wandb_project: project_name # enter your project name here
wandb_entity: project_entity # enter your entity here
device: cuda
log_images_freq: 50
# dataset configuration
results_folder: "results"
resize_input: 512
d_divisible_crops: 1
source_image_every: 75
crops_min_cover: 0.85
# input augmentations
flip_p: 0.5
jitter_p: 0.1
scale_min: 0.8
scale_max: 1.2
# text configuration
#bootstrap_text: horse # this prompt is not composed with any augmentations at the moment
bootstrap_scheduler: linear # linear | exponential | none
text_criterion: spherical # spherical | cosine
# loss configuration
bootstrap_epoch: -1
lambda_bootstrap: 10
lambda_bootstrap_min: 0
bootstrapping_min_cover: 1
lambda_structure: 2
lambda_screen: 1
lambda_sparsity: 6 # lambda_sparsity * ( lambda_alpha_l0 * L0_loss + lambda_alpha_l1 * L1_loss )
lambda_alpha_l0: 0.01
lambda_alpha_l1: 0.01
lambda_composition: 1
use_negative_bootstrap: False
# CLIP handling configuration
n_aug: 16 # number of augmentations to be applied before CLIP
clip_model_name: "ViT-B/32" # ViT-B/16 | ViT-B/32 | ViT-L/14
relevancy_num_layers: 10
# CLIP augmentations
clip_affine_transform_fill: 1
# training configuration
n_epochs: 1000
scheduler_policy: exponential # [exponential| linear | step | plateau | cosine | none]
gamma: 0.99
min_lr: 0.00001
lr: 0.0025
optimizer: madgrad # [madgrad | adam | radam | rmsprop | sgd]
# CNN backbone configuration
skip_n33d: 128
skip_n33u: 128
skip_n11: 4
num_scales: 7