8prodigy_full版本介绍

以8张512x1024的图作为训练集[全身图]，主要训练参数如下

------------------

network_dim = 32

network_alpha = 16

optimizer_type = 'prodigy'

learning_rate = 1

optimizer_args = ["d0=1e-5", "d_coef=0.4"]

lr_scheduler = 'cosine_with_restarts'

lr_scheduler_num_cycles = 4

lr_warmup_steps = 100

batch_size=4

------------------

目标Loss：0.07-0.1

最终Loss：0.107

------------------

========================== 分割线 ==========================