kihyuk-nam
commented
2 years ago Two months ago I found it and tested a simple example in the SPFlow (the 'PyTorch test' in https://drive.google.com/file/d/1shAeYqXv7EwI6c1cEcALhIWAxcBEpP_P/view?usp=sharing), but I couldn't understand the reason why it's more efficient other than the use of tensors instead of numpy arrays. Maybe I should read the 'Einsum Networks' paper.
The size of codes that we need to modify seems relatively small but the logic to understand seems not easy :-)
Sum-Product Networks (SPN) for PyTorch
Brief Summary
According to the Github of the SPFow, I think that SPFlow supports not only the Tensorflow framework, but also the PyTorch framework (torchspn). The implementation of the torchspn is already included in the current version of the SPFlow. In particular, the performance of the torchspn outperforms over native SPFlow and SPFlow-TF(tensorflow) in terms of the processing time.
So, we'd better review the source code for the torchspn for implementing unsupervised learning in approximate query processing.
Related Results
The following remarks referred to GitHub in SPFlow.
https://github.com/SPFlow/SPFlow/tree/master/src/spn/algorithms/layerwise
The example architecture above has been used to benchmark the runtime with varying number of input features (batch size = 1024) and varying batch size (number of input features = 1024).
The comparison is against a node-wise implementation of SPNs in SPFlow on the CPU and a node-wise implementation of SPNs in SPFlow on the GPU using Tensorflow.