A deep learning framework, which was developed by Neil Jiang. It supports define-by-run computational graph and automatically calculate the gradients of functions. It supports GPU acceleration. It is powerful, flexible and easy to understand.
pip install pinenut
from pinenut import Tensor
import numpy as np
# define a computational graph
x = Tensor(np.array([1, 2, 3]))
y = Tensor(np.array([4, 5, 6]))
z = x + y
z.backward() # calculate the gradients of z with respect to x and y
print(x.grad) # [1, 1, 1]
print(y.grad) # [1, 1, 1]from pinenut import Tensor, Cuda, matmul, as_array
x = Tensor([1, 2, 3])
y = Tensor([4, 5, 6])
cuda_is_available = Cuda.available()
if cuda_is_available:
x.to_gpu()
y.to_gpu()
z = matmul(x, y.T)
assert z.data == as_array(32)
print(type(z.data))
z.backward()
assert (x.grad.data == [4, 5, 6]).all()
assert (y.grad.data == [1, 2, 3]).all()- [mnist]
import numpy as np
import pinenut.core.datasets as dss
from pinenut import MLP, SGD, relu, softmax
from pinenut import Cuda
def data_transform(x):
x = x.flatten()
x = x.astype(np.float32)
return x / 255.0
train = dss.MNIST(train=True, data_transform=data_transform)
test = dss.MNIST(train=False, data_transform=data_transform)
epochs = 5
batch_size = 100
lr = 0.1 # learning rate
model = MLP([784, 100, 10], hidden_activation=relu, output_activation=softmax)
optimizer = SGD(model, lr)
cuda_is_available = Cuda.available()
model.train(train, epochs, batch_size, optimizer, test, enable_cuda=cuda_is_available)
model.save_weights('mnist_weights.npz')- [1] Define-by-run computational graph
- [2] GPU acceleration
- [3] Automatic gradient calculation
- [4] Support for various activation, loss and optimizer functions
- [5] Pure Python implementation
- Python 3.6 or later
- NumPy
- CUDA (optional), CuPy (optional)
- Matplotlib (optional)