Train a DepthNet to predict cloth depth map

[lyk412]

Thanks for your great and solid work!
I noticed that the ICON pipeline includes a Normal model that predicts cloth normal map. Now I'm trying to train a Depth model to predict cloth depth map with the help of Normal's training code, but the results are bad. The train and val losses are on the rise. What do I need to do to fix this? Thanks again!

The visualized training results from tensorboard are as follows.

the training configs are:

name: depth
ckpt_dir: "./data/ckpt/"
resume_path: "./data/ckpt/depth.ckpt"
results_path: "./results"

dataset:
root: "./data/"
rotation_num: 36
train_bsize: 1.0
val_bsize: 1.0
test_bsize: 1.0
types: ["thuman2"]
scales: [100.0]

net:
in_depth: (('image',3), ('T_depth_F',3), ('T_depth_B',3))

lr_N: 1e-4
weight_decay: 0.0
momentum: 0.0
batch_size: 8
num_threads: 8
gpus: [0]
test_gpus: [0]

fast_dev: 0
resume: False
test_mode: False
num_sanity_val_steps: 1

momentum: 0.0
optim: Adam

overfit: False
num_epoch: 20
freq_show_train: 0.1
freq_show_val: 0.01
freq_plot: 0.01
freq_eval: 0.1
schedule: [18]