PDLL's mission is to build a backend-agnostic python deep learning framework for education, pluging and playing for different numpy-like computing engines. It includes autograd, nn, optimizer, io, .etc.
In [1]: import pdll as L
In [2]: a = L.rand(2, 2, 3, requires_grad=True)
In [3]: b = L.rand(3, 3)
In [4]: c = a @ b + L.rand(2, 3)
In [5]: d = (c ** 2) * 2 - L.ones_like(c)
In [6]: d.mean().backward()
In [7]: a.grad
Out[7]:
Tensor([[[0.71458999 0.87984239 0.73015823]
[0.76491385 1.04176047 0.89780678]]
[[1.07044392 1.33654949 1.12387667]
[0.74022419 1.01408228 0.86196845]]])| Name | Performance | Commits |
|---|---|---|
| mnist | acc=0.99 | an exmaple modified from pytorch mnist, but with new network achitecture. |
PDLL is a python deep learning library. To see details in achitecture and todo-list.
| Module | Description |
|---|---|
| pdll.backend | a numpy-like library, types ans operations |
| pdll.autograd | an automatic differentiation library, that can record operations on Tensor type |
| pdll.nn | a neural network library based on autograd |
| pdll.optim | an optimizer algorithm library for deep learning |
| pdll.io | dataset, dataloader and serialization |
To get complete source code of PDLL, please contact me.
To learn more about contributing to PDLL, please contact me.
- pdll
- caffe
- pytorch