multitask network based on PFLD
if there is any error in the codes , just remove the line in which the errors are
refer to gen_data.py
as for my dataset and annotations
I have four kinds of dataset and for every dataset there is a especial indicates_str in gen_data.py
every line in the indicate_str represent an attribute and the first follwing number indicates whether the attribute is recorded in the annotation file . The second following number indicates how many values the attribute has or how long the array representing the attributs is.
For example, if datatype == 'imdb' ,then
indicates_str = '''lmk 0 212
race_gt 0 4
race_st 1 1
angles_by_gt 0 3
angles_by_st 1 3
expressions 1 10
attr6_gt 0 5
attr6_st 1 5
age_gt 1 1
age_st 1 1
gender_gt 1 1
gender_st 1 1
young 0 1
mask 1 1
open_eye_gt 0 1
open_eye_st 1 1
mouth_open_slightly 0 1
mouth_open_widely 0 1
sunglasses_gt 0 1
sunglasses_st 1 1
forty_attr 0 33'''
That is :
the imdb dataset does not have attributes lmk, race_gt , angles_by_gt or etc.
But it has race_st, angles_by_st, expressions and etc.
We can infer by the numbers in the third col that lmk should be an array with length of 212 and race_gt should be an array with lenth of 4 and so on.
refer to train_model_v4_dynamic_weights.py and train-scripts/train-v4-dynamic_weights.sh
change the parameters in train-scripts/train-v4-dynamic_weights.sh as you need and run this script