PyToyTorch

PyToyTorch, or ptt, a toy neural network framework which depends merely on numpy.

If you have pytorch installed, you can export ptt model to pytorch model (but no one would do that apparently).

This project is only an exercise. Do not rely this on your project.

Installation

NOTE: to run test(s) in test/, you should have pytorch, tqdm, and sklearn installed.

git clone https://github.com/mnicnc404/PyToyTorch.git
cd PyToyTorch
python setup.py install

or,

pip install git+https://github.com/mnicnc404/PyToyTorch.git

Example

Here's a minimal example; or you can checkout example/sklearn_dataset_breast_cancer_example.py.

>>> import PyToyTorch as ptt
>>> import numpy as np
>>> model = ptt.nn.Sequential(
...     ptt.nn.FC(4, 5),
...     ptt.nn.Mish(),
...     ptt.nn.FC(5, 1),
...     ptt.nn.Sigmoid(),
... )
>>> x = np.random.rand(2, 4).astype(np.float32)
>>> out = model(x)
>>> out
array([[0.7493436],
       [0.8790313]], dtype=float32)


# train with ptt
>>> y = np.array([[1], [0]], dtype=np.float32)
>>> crit = ptt.loss.BCE()
>>> optim = ptt.optim.SGD(model, 0.1)
>>> out = model(x)
>>> loss = crit(out, y)
>>> loss
1.2003906
>>> grad = crit.backward(out, y)
>>> model.backward(grad)
# optim.update() also cleans grad, so no need to clean grad manually
>>> optim.update()  
# see tests/test_with_pytorch.py for training phase example


# if you have pytorch installed:
>>> pytorch_model = model.export()
>>> pytorch_model
Sequential(
  (0): Linear(in_features=4, out_features=5, bias=True)
  (1): _TorchMish()
  (2): Linear(in_features=5, out_features=1, bias=True)
  (3): Sigmoid()
)
>>> tout = pytorch_model(tx)
>>> tout
tensor([[0.7493],
        [0.8790]], grad_fn=<SigmoidBackward>)
>>> out
array([[0.7493436],
       [0.8790313]], dtype=float32)
>>> (np.power(out - tout.detach().numpy(), 2)).mean()
0.0