pybp1_running.py

model = MoonsModel(n_features=2, n_neurons=50)

cost_func = nn.BCELoss()
optimizer = tr.optim.Adam(params=model.parameters(), lr=0.01)

num_epochs = 20

losses = []
accs = []
for e in range(num_epochs):

    for ix, (_x, _y) in enumerate(ds):

        #=========make inpur differentiable=======================

        _x = Variable(_x).float()
        _y = Variable(_y).float()

        #========forward pass=====================================
        yhat = model(_x).float()
        loss = cost_func(yhat, _y)
        acc = tr.eq(yhat.round(), _y).float().mean() # accuracy

        #=======backward pass=====================================
        optimizer.zero_grad() # zero the gradients on each pass before the update
        loss.backward() # backpropagate the loss through the model
        optimizer.step() # update the gradients w.r.t the loss

        losses.append(loss.data[0])
        accs.append(acc.data[0])

    if e % 1 == 0:
        print("[{}/{}], loss: {} acc: {}".format(e,
        num_epochs, np.round(loss.data[0], 3), np.round(acc.data[0], 3)))