To specify the parameters and training strategies of the model, yaml file format
is being used.
estimator: # basic MLP layer configuration
class: 'ensemble'
num_networks : 5
network:
estimator_network:
- fc1 : {class : Linear, in_features : 8, out_features : 50}
- projection : {class : LinearVarianceNetworkHead, in_features : 50, out_features : 1}
predictor_network:
- fc1 : {class : Linear, in_features : 8, out_features : 50}
- projection : {class : Linear, in_features : 50, out_features : 1}
optimizer:
class : 'QHAdam'
lr : 0.01
dataset:
class: 'xls'
path: "regression_datasets/Concrete_Data.xls"
batch_size : 512
cv_split_num: 10
test_ratio: 0.10
transforms:
x :
- {class : Standardize}
y :
- {class : Standardize}
train:
train_type : epoch
num_iter : 40
weight_type : both
logger:
type: 'wandb'
project: 'uncertainty-estimation'
entity: 'kbora'
name: 'Toy Dataset Complex Weighted'Example config file for Concrete dataset.
Similar to mmdetection, we allow anyone to define & use their own classes in any of the blocks. Simply, define your own class under the folder that your object belongs to and add corresponding class to the *REGISTRY dictionaries defined under each sub-modules init.py file.
.
└── regression-uncertainty
├── LICENSE
├── README.md
├── configs
│ ├── toy_dataset.yaml
│ └── xls_dataset.yaml
├── datasets
│ ├── __init__.py
│ ├── toydata.py
│ ├── toyfunc.py
│ └── xlsdata.py
├── estimations
│ ├── __init__.py
│ └── ensemble.py
├── figures
│ ├── non-weight.png
│ ├── regression-uncertainty.png
│ ├── weighted-ru.png
│ └── weighted.png
├── tools
│ └── train.py
├── utils
│ ├── __init__.py
│ ├── device.py
│ └── logger.py
This section covers some of the well-known paper for imbalanced learning and uncertainity estimation for regression tasks.
Imbalanced regression are the collection of method that try to increase model performances for the areas where model in unsure due to the lack of training data.
- Balanced MSE for Imbalanced Visual Regression | github
- Density‑based weighting for imbalanced regression |
- Delving into Deep Imbalanced Regression |
Uncertainity of the models and the data can be estimated with various methods which are usually classified as (i) bayesian and (ii) non-bayesian methods.