def train(self, data, numepochs):
from NN.loss import MSELoss
mse = MSELoss()
for i in range(numepochs):
x, y = data
yhat = self.forward(x)
dldy = mse.loss_gradient(y, yhat)
self.backward(dldy)
print(f"Epoch {i}, loss: {mse.loss(y, yhat)}")
def train(network, data, numepochs):
from NN.loss import MSELoss
mse = MSELoss()
for i in range(numepochs):
x, y = data
yhat, cachelist = network.forward(x)
dldy = mse.loss_gradient(y, yhat)
network.backward(dldy, cachelist)
print(f"Epoch {i}, loss: {mse.loss(y, yhat)}")
def train(network, data, numepochs, loss=None, optim=None):
if not loss:
from NN.loss import MSELoss
loss = MSELoss()
if not optim:
optim = sgd_optimiser(0.01)
for i in range(numepochs):
x, y = data
yhat, cachelist = network.forward(x)
dldy = loss.loss_gradient(y, yhat)
network.backward(dldy, cachelist, optim)
print(f"Epoch {i}, loss: {loss.loss(y, yhat)}")
def test_network():
from NN.loss import MSELoss
x = np.random.randn(100, 10)
y = np.random.randn(100, 3)
net = layer.Network(layer.FullyConnected(10, 20), layer.Tanh(),
layer.FullyConnected(20, 3), layer.Tanh())
mse = MSELoss()
yhat = net.forward(x, {})
initloss = mse.loss(y, yhat)
layer.train(net, (x, y), 10)
yhat = net.forward(x, {})
finloss = mse.loss(yhat, y)
assert initloss > finloss