根据论文中的公式实现了下mini-batch aware regularization。因为对tensorflow的api掌握的不深，而且这个实现也只是为了加深理解，所以代码会比较丑陋。如果有更好的实现方式，请指正。
`
alpha = 0.01
embedding_size = 16
cate_size = 10
batch_size = 32
cate2_batch = [np.random.randint(0, cate_size) for _ in range(batch_size)]
graph = tf.Graph()
with graph.as_default():
cate_emb_w = tf.get_variable("cate_emb_w", [cate_size, embedding_size])
cate2_emb = tf.nn.embedding_lookup(cate_emb_w, cate2_batch)
...
cur_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=y)
opt = tf.train.GradientDescentOptimizer(0.1)
grads_vals = opt.compute_gradients(cur_loss, [cate_emb_w])
for i, (grad, val) in enumerate(grads_vals):
grad_buffer = [[] for _ in range(cate_size)]
cnt_buffer = [[] for _ in range(cate_size)]
# 记录梯度
for value in cate2_batch:
g_1 = tf.nn.embedding_lookup(grad.values, value)
cnt_buffer[value].append(1)
grad_buffer[value].append(g_1)
# 对梯度求平均
for ind in range(len(grad_buffer)):
if len(grad_buffer[ind]) > 0:
grad_buffer[ind] = tf.reduce_mean(grad_buffer[ind], axis=0)
else:
grad_buffer[ind] = tf.gather(grad.values, ind)
# 根据公式更新梯度
new_grad = [0] * cate_size
for ind in range(cate_size):
new_grad[ind] = grad_buffer[ind]
cnt_buffer[ind] = [sum(cnt_buffer[ind])] * embedding_size
new_grad += tf.divide(val, cnt_buffer) * alpha
grads_vals[i] = (new_grad, val)
train_op = opt.apply_gradients(grads_vals)
with tf.Session(graph=graph) as sess:
sess.run(tf.global_variables_initializer())
print(sess.run(train_op))`
根据论文中的公式实现了下mini-batch aware regularization。因为对tensorflow的api掌握的不深，而且这个实现也只是为了加深理解，所以代码会比较丑陋。如果有更好的实现方式，请指正。
`