high level
optimizer=tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
train=optimizer.minimize(cost, name='train')
low level
for one hide layer model
dc_dw_out, dc_db_out=tf.gradients(cost, [weights['out'], biases['out']])
dc_dw_1, dc_db_1=tf.gradients(cost, [weights['h1'], biases['b1']])
#update
upd_w_1=tf.assign(weights['h1'], weights['h1']-learning_rate*dc_dw_1)
upd_b_1=tf.assign(biases['b1'], biases['b1']-learning_rate*dc_db_1)
upd_w_out=tf.assign(weights['out'], weights['out']-learning_rate*dc_dw_out)
upd_b_out=tf.assign(biases['out'], biases['out']-learning_rate*dc_db_out)
train=tf.group(upd_w_1, upd_b_1, upd_w_out, upd_b_out, name='train')
