A python library focuses on constructing deep probabilistic models on GPU. Pydpm provides efficient distribution sampling functions and has included lots of implemented probabilistic models.
This library can be installed under both Windows and Linux systems.
$ pip install pydpm
Under Windows system, we recommed to install Visual Studio 2019 and latest CUDA Toolkit. The combination of VS2019(with MSVC v142) and CUDA 11.5 has been tested in pydpm2.0.
Model list
Model list is as following:
| Probabilistic Model Name | Abbreviation | Paper Link |
|---|---|---|
| Latent Dirichlet Allocation | LDA | Link |
| Poisson Factor Analysis | PFA | Link |
| Poisson Gamma Belief Network | PGBN | Link |
| Convolutional Poisson Factor Analysis | CPFA | Link |
| Convolutional Poisson Gamma Belief Network | CPGBN | Link |
| Poisson Gamma Dynamical Systems | PGDS | Link |
| Deep Poisson Gamma Dynamical Systems | DPGDS | Link |
| Dirichlet Belief Networks | DirBN | Link |
| Deep Poisson Factor Analysis | DPFA | Link |
| Word Embeddings Deep Topic Model | WEDTM | Link |
More probabilistic models will be further included in pydpm/_model/...
Demo
Create a PGBN model:
from pydpm._model import PGBN
# create the model and deploy it on gpu or cpu
model = PGBN([128, 64, 32], device='gpu')
model.initial(train_data)
train_local_params = model.train(100, train_data)
train_local_params = model.test(100, train_data)
test_local_params = model.test(100, test_data)
# evaluate the model with classification accuracy
# the demo accuracy can achieve 0.8549
results = ACC(train_local_params.Theta[0], test_local_params.Theta[0], train_label, test_label, 'SVM')
# save the model after training
model.save()More model demos can be found in pydpm/examples/...
Source data can be found in Link
Function list
The parameters of partial distribution functions are as following:
| Function | Parameters List |
|---|---|
| Normal | mean, std, times |
| Multinomial | count, prob, times |
| Poisson | lambda, times |
| Gamma | shape, scale, times |
| Beta | alpha, beta, times |
| F | n1, n2, times |
| StudentT | n, times |
| Dirichlet | alpha, times |
| Crt | point, p, times |
| Weibull | shape, scale, times |
| Chisquare | n, times |
| Geometric | p, times |
| ... | ... |
Example
from pydpm._sampler import Basic_Sampler
sampler = Basic_Sampler('gpu')
a = sampler.gamma(np.ones(100)*5, 1, times=10)
b = sampler.gamma(np.ones([100, 100])*5, 1, times=10)More sampler demos can be found in pydpm/_sampler/...
Compare
Compare the sampling speed of distribution functions with numpy:
The compared code can be found in pydpm/example/Sampler_Speed_Demo.py
Compare the sampling speed of distribution functions with tensorflow and torch:
The compared code can be found in pydpm/example/Sampler_Speed_Demo.py
License: Apache License Version 2.0
Contact: Chaojie Wang xd_silly@163.com, Wei Zhao 13279389260@163.com, Jiawen Wu wjw19960807@163.com
Copyright (c), 2020, Chaojie Wang, Wei Zhao, Jiawen Wu, Xinyang Liu, Jie Ren, Yewen Li, Hao Zhang, Bo Chen and Mingyuan Zhou