bochengroup / pydpm Goto Github PK

At present, the diffusion model is popular, especially the DDPM, and it should also belong to the depth probability generative model.

沙发。
windows版本啥时候上线？

There are some problems about the sampler. and I suspect that these errors are caused by the compiled files. I want to recompile those cuda file. so which files are needed for recompilation and how can I do that? tks : )

I’m surprise to see that the Diffusion Model has been added to your repository of Bayesian models. But it is a little complicated to understand the code of Diffusion Model, is there any tutorial to guide me to generate images with the trained model. And It would be nice if there was an example.
Thanks

Would you please provide any definations of metric for evaluation in your repository?

Thanks.

I see that other models have some good API, but I can't find a suitable metric to evaluate the quality of the “Generative Adversarial Networks.” Would you please add an evaluation metric for this model, I think it will greatly improve the flexibility of the application.
Thanks.

I got an error when I run the sampler in my ubuntu PC:
'''
/bin/sh: 1: nvcc: not found
File "/PyDPM4.0.1/pydpm/sampler/distribution_sampler_gpu.py", line 99, in init
dll = ctypes.cdll.LoadLibrary(compact_path)
File "/anaconda3/envs/pydpm/lib/python3.6/ctypes/init.py", line 426, in LoadLibrary
return self._dlltype(name)
File "anaconda3/envs/pydpm/lib/python3.6/ctypes/init.py", line 348, in init
self._handle = _dlopen(self._name, mode)
OSError:/PyDPM4.0.1/pydpm/sampler/_compact/distribution_sampler.so: cannot open shared object file: No such file or directory
'''
I'm green to this, how can I deal with it
Thank you

Dear writer, in the PyDPM/pydpm/model/deep_learning_pm/vae.py, attribute sample and forword in class VAE have some bugs (decoder is wrong, where vae_decoder is right) , and now have be checked.
BuFeng Ge

It is great to see u to start to increase the variants of diffusion models. I wonder if you have a plan to include more recent popular diffusion models in your project, just like

Denoising Diffusion Implicit Models. NIPS 2022
Improved Vector Quantized Diffusion Models
. CVPR 2023

These two are widely treated as baselines in today's reseach

Thanks

I get errors during installing the denpendency of pycuda:
ERROR: Could not build wheels for pycuda, which is required to install pyproject.toml-based projects
from pycuda._driver import * # noqa
ImportError: DLL load failed:
Those errors occurred during the installation by pip. I hava changed 2 device to install pydpm, but It doesn't work both : (
The environment of my win PC is as list:
python 3.6
cuda 10.2
win 10

Hi

Thanks for your efforts for deep generative models. However, I find that the datasets used in your demos are not public datasets. Can you replace the datasets in your code with those in torchvision/torchtext etc, which will be more convenient for us to extend the models in your libarary.

Thanks

The email from Hjelkrem Tan, who is a PhD student in University of Oslo:

``
Dear Mr. Chaojie Wang,

I am a PhD student at the University of Oslo, Norway. I would very much like to use your PyDPM library in my research, as I have read several papers from you and your colleagues on topic models. Would you be able to answer some questions I have about the implementations in your GitHub repo?

I cannot find any implementation of SawETM in the PyDPM library. Are you planning to release this with PyDPM?
In pydpm/model/hybrid_pm/whai.py it seems that the implementation of the encoder does not include the stochastic downward part from the WHAI paper. Is this intentional or something you will change later?

I hope that you can clarify this for me. Thank you for your time!

With best regards,

Martine Hjelkrem Tan

PhD student, University of Oslo

Digital Signal Processing and Image Analysis Group
''

Hi, I find some new model in this new version library, but there is no corresponding document to explanation them. Cound you update them?
Best wish

How to make the model continuously train on two datasets. Specifically, how to continue training on the second data without initialization after training on the first one.

Thanks

I’m glad that this repository has collected many generative models, like the diffusion model, variational autoencoder, Generative Adversarial Networks and so on. But I didn’t find the flow-like model, would you like to add a model and demo of Normalizing Flow?

Thanks for your effort in building this repository of the generative model.

Thanks for your contributions to the cuda based sampler module. I want to build some new distributions' PRNG by this sampler, and find out there are different files in sampler/_compact. wanna to know where and how to edit them for new PRNG

您好，我发现这个文档https://dustone-mu.github.io/Getting%20Started/ 和exampleshttps://github.com/BoChenGroup/PyDPM/tree/master/pydpm/example 里面提供的代码结构和内容有一些不一致的地方，请问后续能统一一下吗，非常感谢！

Thanks for your efforts in summarizing Bayesian models.

I am wondering if you would like to include the widely used Gaussian Process in your library. Because some recent studies of my colleagues have difficulties in speeding up the sampling efficiency, where the previous implementation of Gaussian Process on CPU is too slow for us.

Thanks

I couldn't find any details about whai in documentreadme. could you provide more information?

DPGDS error
I meet this warning when running dpgds demo:
Training Stage: epoch 0 takes 4.90 seconds. Likelihood: -0.425
Training Stage: epoch 1 takes 5.11 seconds. Likelihood: -0.367
Training Stage: epoch 2 takes 4.83 seconds. Likelihood: -0.373
Training Stage: epoch 3 takes 4.65 seconds. Likelihood: -0.376
Training Stage: epoch 4 takes 4.54 seconds. Likelihood: -0.378
Training Stage: epoch 5 takes 4.49 seconds. Likelihood: -0.377
Training Stage: epoch 6 takes 4.44 seconds. Likelihood: -0.377
Training Stage: epoch 7 takes 4.41 seconds. Likelihood: -0.377
Training Stage: epoch 8 takes 4.36 seconds. Likelihood: -0.374
Training Stage: epoch 9 takes 4.37 seconds. Likelihood: -0.369
Training Stage: epoch 10 takes 4.36 seconds. Likelihood: -0.367
Training Stage: epoch 11 takes 4.35 seconds. Likelihood: -0.363
Training Stage: epoch 12 takes 4.28 seconds. Likelihood: -0.360
Training Stage: epoch 13 takes 4.26 seconds. Likelihood: -0.357
Training Stage: epoch 14 takes 4.25 seconds. Likelihood: -0.349
Training Stage: epoch 15 takes 4.25 seconds. Likelihood: -0.343
Training Stage: epoch 16 takes 4.24 seconds. Likelihood: -0.334
Training Stage: epoch 17 takes 4.28 seconds. Likelihood: -0.326
Training Stage: epoch 18 takes 4.25 seconds. Likelihood: -0.315
Training Stage: epoch 19 takes 4.23 seconds. Likelihood: -0.306
Training Stage: epoch 20 takes 4.20 seconds. Likelihood: -0.294
Warning: shape -36.502342 <= 0 in threads idx: 8258 [thread:(2125698480, 0), block:(66, 0)]
Warning: shape -0.142884 <= 0 in threads idx: 8260 [thread:(2125698480, 0), block:(68, 0)]
Warning: shape -2.206542 <= 0 in threads idx: 8262 [thread:(2125698480, 0), block:(70, 0)]
Warning: shape -0.727849 <= 0 in threads idx: 8264 [thread:(2125698480, 0), block:(72, 0)]
Warning: shape -0.010437 <= 0 in threads idx: 8265 [thread:(2125698480, 0), block:(73, 0)]
Warning: shape -0.446480 <= 0 in threads idx: 8266 [thread:(2125698480, 0), block:(74, 0)]
Warning: shape -0.286718 <= 0 in threads idx: 8267 [thread:(2125698480, 0), block:(75, 0)]

Why is the gpu version of PGDS slower than the cpu version? The time those two cost per epoch is about 1.256 to 1.032 seconds. That's very different from other models like PGBN.

I found that the sampler part in this project is similar to cupy(cuda version of numpy). So what the difference between them? Does this pydpm sampler faster? btw, appreciate for your contribution in DTM

How to fix the error below?
ptxas fatal : Unresolved extern function '_Z3powfi' Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "D:\Anaconda3\envs\mmdet\lib\site-packages\pydpm\_model\_pgbn.py", line 52, in __init__
self._sampler = Basic_Sampler(self._model_setting.device) File "D:\Anaconda3\envs\mmdet\lib\site-packages\pydpm\_sampler\_basic_sampler.py", line 37, in __init__ self._gpu_sampler_initial()
File "D:\Anaconda3\envs\mmdet\lib\site-packages\pydpm\_sampler\_basic_sampler.py", line 63, in _gpu_sampler_initial
sampler = distribution_sampler_gpu(self.system_type)
File "D:\Anaconda3\envs\mmdet\lib\site-packages\pydpm\_sampler\_distribution_sampler_gpu.py", line 56, in __init__
dll = ctypes.cdll.LoadLibrary(compact_path)
File "D:\Anaconda3\envs\mmdet\lib\ctypes\__init__.py", line 451, in LoadLibrary
return self._dlltype(name) File "D:\Anaconda3\envs\mmdet\lib\ctypes\__init__.py", line 373, in __init__
self._handle = _dlopen(self._name, mode)
FileNotFoundError: Could not find module 'D:\Anaconda3\envs\mmdet\lib\site-packages\pydpm\_sampler\_compact\sampler_kernel.dll' (or one of its dependencies). Try using the full path with constructor syntax.

when running
from pydpm._model import PGBN
model = PGBN([128,64,32], device='gpu')
Thanks.

thanks for the pre q&a about recompilelation, and I still wonder how to compile sampler library files manually on my server. want to depoly this project on AWS

I find some intereseting work published in recent top conference. Can your guys include these projects into this library for convience?

The work list:
Alleviating “Posterior Collapse” in Deep Topic Models via Policy Gradient", NeurIPS 2022
HyperMiner: Topic Taxonomy Mining with Hyperbolic Embedding , NeurIPS 2022
Knowledge-Aware Bayesian Deep Topic Model, NeurIPS2022