for i in range(param_niter):
        # prolaz unaprijed
        s1 = np.dot(X, W1.transpose()) + b1   # N x h
        h1 = np.vectorize(relu)(s1) # N x h
        s2 = np.dot(h1, W2.transpose()) + b2 # N x C

        expscores = np.vectorize(math.exp)(s2) # N x C
        sumexp = np.sum(expscores, axis = 1)   # N x 1

        # logaritmirane vjerojatnosti razreda
        probs = expscores / sumexp[:, None]    # N x C
        logprobs = np.vectorize(math.log)(probs)  # N x C

        # gubitak
        loss  = np.sum([ -logprobs[i][Y_[i]] for i in range(N)])     # scalar

        # dijagnostički ispis
        if i % 10 == 0:
            print("iteration {}: loss {}".format(i, loss))

        # prolaz unazad
        # derivacije komponenata gubitka po rezultatu
        Ycode = [ [1 if Y_[i] == j else 0 for j in range(C)] for i in range(N) ]
        Gs2 = probs - Ycode     # N x C
        grad_W2 = 1./N * np.dot(Gs2.transpose(), h1) # C x h
        grad_b2 = 1./N * np.sum(Gs2.transpose(), axis = 1) # C x 1

        Gh1 = np.dot(Gs2, W2) # N x h
        Gh1[h1 == 0] = 0     # N x h
        Gs1 = Gh1

        grad_W1 = 1./N * np.dot(Gs1.transpose(), X) # h x D
        grad_b1 = 1./N * np.sum(Gs1.transpose(), axis = 1) # h x 1

        # poboljšani parametri
        W1 += -param_delta * grad_W1
        b1 += -param_delta * grad_b1
        W2 += -param_delta * grad_W2
        b2 += -param_delta * grad_b2