The cTensor (crafted tensor) is a super light-weight deep learning library (perhaps we cannot even call it a libray). It's based on numpy and furthermore its only one dependency is actually numpy. Features include dynamic graph, autograd and user-defined operations in numpy. The line number of core code is within 300 400, making it friendly for study and teaching purpose. It mimics PyTorch framework defacto.
Stars are welcomed! : )
The file contains the definition of Tensor and its many basic operations. The definition of Operator enables user-defined operators on Tensor and therefore forward and backward in numpy.ndarray or Tensor (see example View class).
The file includes some pre-defined Operators on Tensor, e.g., ReLU and Conv2d. Note that misc things about Conv2d are exposed here as I am lazy to separate details and abstracts about it.
An Operator subclass could be applied to Tensor only when an instance of the class are initiated. The file consists of functions that make an instance of the Operator and pass through it.
Optimizers here.
High level abstracts of Operators and their parameters. They are totally in PyTorch fashion.
Some testcases. Maybe too hard to read although I've written up some comments.
Readme in more detail.
Add nn.Conv2d.
Speed 100x up conv2d by using np.tensordot, which operates automatically in multicores.
Now support max_pooling operation with any stride, kernel size and padding.
Fix buges related to conv2d backpropagation.
Now support conv2d with any stride and padding.
We have supported batch conv2d operation in pytorch fashion (limited in stride 1, no padding) !!
Of course, cTensor has not yet support convolutional operations.