config.py

'''This file configures the training procedure because handling arguments in every single function is so exhaustive for
research purposes. Don't try this code if you are a software engineer.'''

# device settings
device = 'cuda'  # or 'cpu'

# data settings
dataset_path = "data/images"  # parent directory of datasets
class_name = "dummy_data"  # dataset subdirectory
modelname = "dummy_test"  # export evaluations/logs with this name
pre_extracted = True  # were feature preextracted with extract_features?

img_size = (768, 768)  # image size of highest scale, others are //2, //4
assert img_size[0] % 128 == 0 and img_size[1] % 128 == 0, "image width/height should be a multiple of 128"

img_dims = [3] + list(img_size)

# transformation settings
norm_mean, norm_std = [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]

# network hyperparameters
n_scales = 3  # number of scales at which features are extracted, img_size is the highest - others are //2, //4,...
clamp = 3  # clamping parameter
max_grad_norm = 1e0  # clamp gradients to this norm
n_coupling_blocks = 4  # higher = more flexible = more unstable
fc_internal = 1024  # * 4 # number of neurons in hidden layers of s-t-networks
lr_init = 2e-4  # inital learning rate
use_gamma = True

extractor = "effnetB5"  # feature dataset name (which was used in 'extract_features.py' as 'export_name')
n_feat = {"effnetB5": 304}[extractor]  # dependend from feature extractor
map_size = (img_size[0] // 32, img_size[1] // 32)

# dataloader parameters
batch_size = 16 
kernel_sizes = [3] * (n_coupling_blocks - 1) + [5]

# total epochs = meta_epochs * sub_epochs
# evaluation after <sub_epochs> epochs
meta_epochs = 4  # total epochs = meta_epochs * sub_epochs
sub_epochs = 60  # evaluate after this number of epochs

# output settings
verbose = True
hide_tqdm_bar = True
save_model = True