Autograd But Smaller
Examples
from autograd.tensor import Tensor
X = Tensor.uniform(shape=(100, 10))
W = Tensor.uniform(shape=(10, 1))
b = Tensor.zeros(shape=(1, 10))
linear = (X @ W + b).relu()
linear.backward()
print(X.grad, W.grad, b.grad) # df/dX, df/dW, df/db
Regression
class RegNet(nn.Module):
def __init__(self, n_features):
super().__init__()
self.l1 = nn.Linear(n_features, 32)
self.l2 = nn.Linear(32, 1)
def forward(self, X):
out = self.l1(X).relu()
return self.l2(out)
X_, y_ = load_boston()["data"], load_boston()["target"]
X, y = Tensor(X_), Tensor(y_).reshape(-1, 1)
# net training
net = RegNet(X.shape[1])
optimizer = SGD(net.parameters(), lr=1e-6)
for _ in range(500):
net.zero_grad()
loss = MSELoss(net(X), y)
loss.backward()
optimizer.step()
Classification
class ClassifyNet(nn.Module):
def __init__(self, n_features, n_classes):
super().__init__()
self.l1 = nn.Linear(n_features, 32)
self.l2 = nn.Linear(32, n_classes)
def forward(self, X):
out = self.l1(X).relu()
return self.l2(out).softmax(axis=1)
X_, y_ = load_breast_cancer()["data"], load_breast_cancer()["target"]
X, y = Tensor(X_), validation.one_hot_encode(y_)
net = ClassifyNet(X.shape[1], n_classes=2)
optimizer = SGD(net.parameters(), lr=1e-5)
for _ in range(n_iter):
net.zero_grad()
loss = CrossEntropyLoss(net(X), y)
loss.backward()
optimizer.step()
References
