Experiment setting

style_shoes_label_folder: 原本模型架構
style_shoes_label_folder2: 將原本兩個gen改成一個
w/o flowing
- style_shoes_label_folder_v2 : 一個 gen + cycle loss*10 (MASTER) // 20/07/21 [112]
- style_shoes_label_folder_v11 : 一個 gen + w/o latent loss // 20/07/15 [112]
- style_shoes_label_folder_v12 : 一個 gen + w/o reconstruction loss 20/07/15 [241]
with flowing
- style_shoes_label_folder_v9 : iteration 500000 + beta distribution(half) + flowing加上content loss (Good) [241]

randn --> style encoder

style_shoes_label_folder_v3 : style code 沒有使用 randn取 (MASTER_v2)
- model = MASTER_v2
style_shoes_label_folder_v4 [241]: style code 沒有使用 randn取 + cycle loss *10 (MASTER_v2)
- model = MASTER_v2 + cycle loss
style_shoes_label_folder_v4 [112]: add flowing on one generator with randn noise
- model = MASTER + flowing
style_shoes_label_folder_v5 :
- model = MASTER_v2 + flowing
style_shoes_label_folder_v6 : 在decode的時候加入domain資訊 + iteration 改成500000 + beta distribution
- model = MASTER + flowing
style_shoes_label_folder_v7 : 將 iteration 改成500000 和 domainess sample on uniform distribution
style_shoes_label_folder_v8 : 在decode的時候加入domain資訊 + iteration 1000000 + beta distribution
- model = MASTER + flowing
style_shoes_label_folder_v9 : iteration 500000 + beta distribution(half) + flowing加上content loss (Good) [在241上]
style_shoes_label_folder_v10 : 跟v9一樣+ lerp + + 更改過後flowing loss
style_shoes_label_folder_v13 : 重現v9

nohup python master_train.py --config configs/style_shoes_label_folder_v2.yaml --trainer MASTER --gpu 0 &> v2_log.txt &
nohup python master_train.py --config configs/style_shoes_label_folder_v7.yaml --trainer MASTER --gpu 1 --port 8000 &> v7_log.txt &
python master_test.py --trainer MASTER --config configs/style_shoes_label_folder_v4.yaml --checkpoint outputs/style_shoes_label_folder_v4/checkpoints/gen_00500000.pt
windows
- start/min python master_train.py --config configs/style_shoes_label_folder_v8.yaml --trainer MASTER --gpu 1 --port 8097