python-ldl 0.0.3

Label distribution learning (LDL) and label enhancement (LE) toolkit implemented in python.

PyLDL

Label distribution learning (LDL) and label enhancement (LE) toolkit implemented in python, including:

• LDL algorithms:
  • (Geng, Yin, and Zhou 2013)[TPAMI]: CPNN$^1$.
  • (Geng and Hou 2015)[IJCAI]: LDSVR.
  • ⭐(Geng 2016)[TKDE]: SA_BFGS, SA_IIS, AA_KNN, AA_BP, PT_Bayes, and PT_SVM.
  • (Yang, Sun, and Sun 2017)[AAAI]: BCPNN and ACPNN.
  • (Xu and Zhou 2017)[IJCAI]: IncomLDL$^2$.
  • (Shen et al. 2017)[NeurIPS]: LDLF.
  • (Wang and Geng 2019)[IJCAI]: LDL4C$^3$.
  • (Shen et al. 2020)[南京理工大学学报 (Chinese)]: AdaBoostLDL.
  • (González et al. 2021a)[Inf. Sci.]: SSG_LDL$^4$.
  • (González et al. 2021b)[Inf. Fusion]: DF_LDL.
  • (Wang and Geng 2021a)[IJCAI]: LDL_HR$^3$.
  • (Wang and Geng 2021b)[ICML]: LDLM$^3$.
  • (Jia et al. 2021)[TKDE]: LDL_SCL.
  • (Jia et al. 2023a)[TKDE]: LDL_LRR.
  • (Jia et al. 2023b)[TNNLS]: LDL_DPA.
  • (Wen et al. 2023)[ICCV]: CAD$^1$, QFD2$^1$, and CJS$^1$.
• LE algorithms:
  • (Xu, Liu, and Geng 2019)[TKDE]: FCM, KM, LP, ML, and GLLE.
  • (Xu et al. 2020)[ICML]: LEVI.
  • (Zheng, Zhu, and Tang 2023)[CVPR]: LIBLE.
• LDL metrics: chebyshev, clark, canberra, kl_divergence, cosine, intersection, etc.
• Structured LDL datasets: Human_Gene, Movie, Natural_Scene, s-BU_3DFE, s-JAFFE, Yeast, etc.
• LDL applications:
  • Facial emotion recognition (supported datasets: JAFFE)
  • (Shirani et al. 2019)[ACL]: Emphasis selection (supported datasets: SemEval2020; pre-trained GloVe embeddings can be downloaded here).
  • (Wu et al. 2019)[ICCV]: Lesion counting (supported datasets: ACNE04).
  • (Chen et al. 2020)[CVPR]: Facial emotion recognition with auxiliary label space graphs (supported datasets: CK+; OpenFace can be downloaded here, and the required models can be downloaded here).

$^1$ Technically, these methods are only suitable for totally ordered labels.

$^2$ These are algorithms for incomplete LDL, so you should use pyldl.utils.random_missing to generate the missing label distribution matrix and the corresponding mask matrix in the experiments.

$^3$ These are LDL classifiers, so you should use predict_proba to get label distributions and predict to get predicted labels.

$^4$ These are oversampling algorithms for LDL, therefore you should use fit_transform to generate synthetic samples.

Installation

PyLDL is now available on PyPI. Use the following command to install.

pip install python-ldl

To install the newest version, you can clone this repo and run the setup.py file.

python setup.py install

Usage

Here is an example of using PyLDL.

from pyldl.utils import load_dataset
from pyldl.algorithms import SA_BFGS
from pyldl.metrics import score

from sklearn.model_selection import train_test_split

dataset_name = 'SJAFFE'
X, y = load_dataset(dataset_name)
X_train, X_test, y_train, y_test = train_test_split(X, y)

model = SA_BFGS()
model.fit(X_train, y_train)

y_pred = model.predict(X_test)
print(score(y_test, y_pred))

For those who would like to use the original implementation:

1. Install MATLAB.
2. Install MATLAB engine for python.
3. Download LDL Package here.
4. Get the package directory of PyLDL (...\Lib\site-packages\pyldl).
5. Place the LDLPackage_v1.2 folder into the matlab_algorithms folder.

Now, you can load the original implementation of the method, e.g.:

from pyldl.matlab_algorithms import SA_IIS

You can visualize the performance of any model on the artificial dataset (Geng 2016) with the pyldl.utils.plot_artificial function, e.g.:

from pyldl.algorithms import LDSVR, SA_BFGS, SA_IIS, AA_KNN, PT_Bayes, GLLE, LIBLE
from pyldl.utils import plot_artificial

methods = ['LDSVR', 'SA_BFGS', 'SA_IIS', 'AA_KNN', 'PT_Bayes', 'GLLE', 'LIBLE']

plot_artificial(model=None, figname='GT')
for i in methods:
    plot_artificial(model=eval(f'{i}()'), figname=i)

The output images are as follows.

(Ground Truth)	LDSVR

SA_BFGS	SA_IIS

AA_KNN	PT_Bayes

GLLE	LIBLE

Enjoy! :)

Experiments

For each algorithm, a ten-fold cross validation is performed, repeated 10 times with s-JAFFE dataset and the average metrics are recorded. Therefore, the results do not fully describe the performance of the model.

Results of ours are as follows.

Algorithm	Cheby.(↓)	Clark(↓)	Can.(↓)	K-L(↓)	Cos.(↑)	Int.(↑)
SA-BFGS	.092 ± .010	.361 ± .029	.735 ± .060	.051 ± .009	.954 ± .009	.878 ± .011
SA-IIS	.100 ± .009	.361 ± .023	.746 ± .050	.051 ± .008	.952 ± .007	.873 ± .009
AA-kNN	.098 ± .011	.349 ± .029	.716 ± .062	.053 ± .010	.950 ± .009	.877 ± .011
AA-BP	.120 ± .012	.426 ± .025	.889 ± .057	.073 ± .010	.931 ± .010	.848 ± .011
PT-Bayes	.116 ± .011	.425 ± .031	.874 ± .064	.073 ± .012	.932 ± .011	.850 ± .012
PT-SVM	.117 ± .012	.422 ± .027	.875 ± .057	.072 ± .011	.932 ± .011	.850 ± .011

Results of the original MATLAB implementation (Geng 2016) are as follows.

Algorithm	Cheby.(↓)	Clark(↓)	Can.(↓)	K-L(↓)	Cos.(↑)	Int.(↑)
SA-BFGS	.107 ± .015	.399 ± .044	.820 ± .103	.064 ± .016	.940 ± .015	.860 ± .019
SA-IIS	.117 ± .015	.419 ± .034	.875 ± .086	.070 ± .012	.934 ± .012	.851 ± .016
AA-kNN	.114 ± .017	.410 ± .050	.843 ± .113	.071 ± .023	.934 ± .018	.855 ± .021
AA-BP	.130 ± .017	.510 ± .054	1.05 ± .124	.113 ± .030	.908 ± .019	.824 ± .022
PT-Bayes	.121 ± .016	.430 ± .035	.904 ± .086	.074 ± .014	.930 ± .016	.846 ± .016
PT-SVM	.127 ± .017	.457 ± .039	.935 ± .074	.086 ± .016	.920 ± .014	.839 ± .015

Requirements

matplotlib>=3.6.1
numpy>=1.22.3
qpsolvers>=4.0.0
quadprog>=0.1.11
scikit-fuzzy>=0.4.2
scikit-learn>=1.0.2
scipy>=1.8.0
tensorflow>=2.8.0
tensorflow-probability>=0.16.0