pythonのtensorflowで交差検定を行う際にエラーがでる

前提・実現したいこと

一般的に用いられる交差検定は、データをトレーニングとテストに分けてトレーニングデータで訓練し損失関数にトレーニングデータの実ラベルとの出力誤差を最小化し、テスト用データでそのモデルを評価する一連の流れだと思うが、今回実装したいコードはトレーニングデータで訓練し、損失関数にテストデータの実ラベルとの出力誤差の最小化を行いたい。

発生している問題・エラーメッセージ

エラーメッセージ
`No gradients provided for any variable, check your graph for ops that do not support gradients, between variables ["<tf.Variable 'Variable:0' shape=() dtype=float32_ref>", "<tf.Variable 'Variable_1:0' shape=(564, 512) dtype=float32_ref>", "<tf.Variable 'Variable_2:0' shape=(512,) dtype=float32_ref>", "<tf.Variable 'Variable_3:0' shape=(512, 128) dtype=float32_ref>", "<tf.Variable 'Variable_4:0' shape=(128,) dtype=float32_ref>", "<tf.Variable 'Variable_5:0' shape=(128, 2) dtype=float32_ref>", "<tf.Variable 'Variable_6:0' shape=(2,) dtype=float32_ref>"] and loss Tensor("Assign:0", shape=(), dtype=float32_ref).
`
###　該当のソースコード
 cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
 custom_loss = tf.assign(v, sess.run(cross_entropy, feed_dict={x: feature1, y_: box1}))
 train_step = tf.train.GradientDescentOptimizer(0.01).minimize(custom_loss)

### 　ソースコード
# number for CV
fold_num = 5
num_classes = 2
epochs = 500

# fix random seed for reproducibility
seed = 7
numpy.random.seed(seed)

# load features
feature = pd.read_csv('test.csv' ,usecols = range(1, 565), header=0)
x = df(feature)

#load label
label = pd.read_csv('test1.csv', usecols=[8], header=0)
y = df(label)

partner_id = pd.read_csv('test1.csv', usecols=[0], header=0)
z = df(partner_id)
z = z.values.tolist()

#Formatting data
x_train = x.astype(np.float)
y_train = y.astype(np.float)
X = x_train.values
Y = y_train.values



# define X-fold cross validation
kfold = StratifiedKFold(n_splits=fold_num, shuffle=True, random_state=seed)
cvscores = []


# input
x = tf.placeholder(tf.float32, [None, 564])
# label
y_ = tf.placeholder(tf.float32, [None, 2])

v = tf.Variable(1.0)

# FC1
W1 = tf.Variable(tf.random_normal([564, 512], mean=0.0, stddev=0.05))
b1 = tf.Variable(tf.zeros([512]))
# layer output
h1 = tf.nn.sigmoid(tf.matmul(x, W1) + b1)

# FC2
W2 = tf.Variable(tf.random_normal([512, 128], mean=0.0, stddev=0.05))
b2 = tf.Variable(tf.zeros([128]))
# layer output
h2 = tf.nn.relu(tf.matmul(h1, W2) + b2)

# FC3
W3 = tf.Variable(tf.random_normal([128, 2], mean=0.0, stddev=0.05))
b3 = tf.Variable(tf.zeros([2]))
# output
y = tf.nn.softmax(tf.matmul(h2, W3) + b3)

sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)

for train, test in kfold.split(X, Y):
    feature = df(X[train])
    feature1 = df(X[test])

    box = []
    for i in Y[train]:
        if i == 0:
            box.append([1, 0])
        elif i == 1:
            box.append([0, 1])

    box1 = []
    for i in Y[test]:
        if i == 0:
            box1.append([1, 0])
        elif i == 1:
            box1.append([0, 1])



    # training

    cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
    custom_loss = tf.assign(v, sess.run(cross_entropy, feed_dict={x: feature1, y_: box1}))
    train_step = tf.train.GradientDescentOptimizer(0.1).minimize(custom_loss)

    for step in range(500):
        sess.run(train_step, feed_dict={x: feature, y_: box})



    predict = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
    accuracy = tf.reduce_mean(tf.cast(predict, tf.float32))

    # テストデータを使って、正解率を求める
    acc = sess.run(accuracy, feed_dict={x: feature1, y_: box1})
    print("正解率=", acc)