1 Star 0 Fork 0

herrvan/S2VD-copy

加入 Gitee
与超过 1200万 开发者一起发现、参与优秀开源项目,私有仓库也完全免费 :)
免费加入
文件
克隆/下载
options_derain.json 2.07 KB
一键复制 编辑 原始数据 按行查看 历史
zsyOAOA 提交于 2021-03-19 13:36 . Upload training and testing code
{
# training settings
"epochs": 45,
"max_iter_EM": 2,
"print_freq": 5,
"resume": "None",
"delta": 0.03, # step size for E-step
"langevin_steps": 5, # number of steps of langevin
"epsilon2": 1e-6,
"rho": 0.5, # TV regularized penalty for unlabled data
"tv_weight": [1.0, 1.0, 2.0], # hyper-parameter gamma of paper
"gpu_id": [1],
"lr_D": 2e-4, # Learning rate for the deraining network
"lr_GRain": 1e-4, # Learning rate for the emission model
"lr_GState": 1e-3, # Learning rate for the transition model
"weight_decay_D": 1e-4, # Weight decay for the deraining network
"weight_decay_GRain": 0, # Weight decay for the emission model
"weight_decay_GState": 0,# Weight decay for the transition model
"factor_lr": 0.5, # decay factor for learning rate
"patch_size": 64, # patch size
"model_dir": "models_save",
"log_dir": "logs_save",
"milestones": [30, 45], # Epochs to decay learning rate
"pretrain_derain": 5, # Number of epochs to pretrain the generator
# network settings
"latent_size": 64,
"state_size": 128,
"motion_size": 64,
"feature_state": 128,
"feature_rain_G": 64, # number of feature for rain generator
"feature_derain_D": 32, # number of feature for derain net
"n_resblocks": 8, # residual blocks in derain net
"max_grad_norm_D": 1e4, # grad norm threshold for derain net
"truncate": 20,
"truncate_test": 16,
# dataset .mat file
"train_path": "/ssd1t/NTURain/train", # folder to save the training data of labeled synthetic data: *.mat
"train_path_semi": "/ssd1t/NTURain/train_semi", # folder to save the training data of unlabeled real data: *.mat
"test_path": "testsets/synthetic_NTURain/a1_Rain", # testing images of NTURain synthetic data: *.jpg
"test_path_semi": "testsets/real_NTURain/ra1_Rain", # testing images of NTURain real data:*.jpg
"latent_dir_name": "latent_eps6_rho05_tv112" , # floder to save the latent and state variable
# random seed
"seed": 1000
}
马建仓 AI 助手
尝试更多
代码解读
代码找茬
代码优化
Python
1
https://gitee.com/herrvan/s2-vd-copy.git
git@gitee.com:herrvan/s2-vd-copy.git
herrvan
s2-vd-copy
S2VD-copy
master

搜索帮助