def testNonLinear(X, Y):
sigmoide = Non_linear.Sigmoide()
tanh = Non_linear.Tanh()
coche1 = Linear(2, 5)
coche2 = Linear(5, 1)
mse = MSE()
res_sigmoide = None
loss = []
maxIter = 100
for _ in range(maxIter):
#forward
res_lin1 = coche1.forward(X)
res_tanh = tanh.forward(res_lin1)
res_lin2 = coche2.forward(res_tanh)
res_sigmoide = sigmoide.forward(res_lin2)
#loss
res = np.array([[1 if res_sigmoide[i] > 0.5 else 0]
for i in range(len(res_sigmoide))])
# res = np.array([res_sigmoide[i] > 0 for i in range(len(res_sigmoide))])
# print("res",res.shape)
loss.append(sum(mse.forward(Y.reshape(-1, 1), res)))
#retro-propager
res_mse = mse.backward(Y.reshape(-1, 1), res)
# print("mse",res_mse.shape)
delta_sig = sigmoide.backward_delta(res_lin2, res_mse)
coche2.zero_grad()
coche2.backward_update_gradient(res_tanh, delta_sig)
coche2.update_parameters(0.05)
delta_lin2 = coche2.backward_delta(X, delta_sig)
delta_tanh = tanh.backward_delta(res_lin1, delta_lin2)
coche1.zero_grad()
coche1.backward_update_gradient(X, delta_tanh)
coche1.update_parameters(0.05)
return loss, "Nonlinear", maxIter
def testlinear(datax, datay):
## Lineaire et MSE
linear = Linear(10, 1)
linear.zero_grad()
mse = MSE()
# print(datax,datay)
l_mse = []
for _ in range(100):
res_lin = linear.forward(datax)
res_mse = mse.forward(datay, res_lin)
delta_mse = mse.backward(datay, res_lin)
l_mse.append(sum(res_mse))
linear.zero_grad()
linear.backward_update_gradient(datax, delta_mse)
linear.update_parameters()
return l_mse, "Linear", 100