TensorFlow:执行此损失计算

Question

我的问题和问题在两个代码块下面说明.

---

损失函数

def loss(labels, logits, sequence_lengths, label_lengths, logit_lengths):    
    scores = []
    for i in xrange(runner.batch_size):
        sequence_length = sequence_lengths[i]
        for j in xrange(length):
            label_length = label_lengths[i, j]
            logit_length = logit_lengths[i, j]

             # get top k indices <==> argmax_k(labels[i, j, 0, :], label_length)
            top_labels = np.argpartition(labels[i, j, 0, :], -label_length)[-label_length:]
            top_logits = np.argpartition(logits[i, j, 0, :], -logit_length)[-logit_length:]

            scores.append(edit_distance(top_labels, top_logits))

    return np.mean(scores)

# Levenshtein distance
def edit_distance(s, t):
    n = s.size
    m = t.size
    d = np.zeros((n+1, m+1))
    d[:, 0] = np.arrange(n+1)
    d[0, :] = np.arrange(n+1)

    for j in xrange(1, m+1):
        for i in xrange(1, n+1):
            if s[i] == t[j]:
                d[i, j] = d[i-1, j-1]
            else:
                d[i, j] = min(d[i-1, j] + 1,
                              d[i, j-1] + 1,
                              d[i-1, j-1] + 1)

    return d[m, n]

---

被用于

我试图压扁我的代码,以便一切都在一个地方发生.如果有混淆/错误点,请告诉我.

sequence_lengths_placeholder = tf.placeholder(tf.int64, shape=(batch_size))
labels_placeholder = tf.placeholder(tf.float32, shape=(batch_size, max_feature_length, label_size))
label_lengths_placeholder = tf.placeholder(tf.int64, shape=(batch_size, max_feature_length))
loss_placeholder = tf.placeholder(tf.float32, shape=(1))

logit_W = tf.Variable(tf.zeros([lstm_units, label_size]))
logit_b = tf.Variable(tf.zeros([label_size]))

length_W = tf.Variable(tf.zeros([lstm_units, max_length]))
length_b = tf.Variable(tf.zeros([max_length]))

lstm = rnn_cell.BasicLSTMCell(lstm_units)
stacked_lstm = rnn_cell.MultiRNNCell([lstm] * layer_count)

rnn_out, state = rnn.rnn(stacked_lstm, features, dtype=tf.float32, sequence_length=sequence_lengths_placeholder)

logits = tf.concat(1, [tf.reshape(tf.matmul(t, logit_W) + logit_b, [batch_size, 1, 2, label_size]) for t in rnn_out])

logit_lengths = tf.concat(1, [tf.reshape(tf.matmul(t, length_W) + length_b, [batch_size, 1, max_length]) for t in rnn_out])

optimizer = tf.train.AdamOptimizer(learning_rate)
global_step = tf.Variable(0, name='global_step', trainable=False)
train_op = optimizer.minimize(loss_placeholder, global_step=global_step)

...
...
# Inside training loop

np_labels, np_logits, sequence_lengths, label_lengths, logit_lengths = sess.run([labels_placeholder, logits, sequence_lengths_placeholder, label_lengths_placeholder, logit_lengths], feed_dict=feed_dict)
loss = loss(np_labels, np_logits, sequence_lengths, label_lengths, logit_lengths)
_ = sess.run([train_op], feed_dict={loss_placeholder: loss})

---

我的问题

问题是这是返回错误:

  File "runner.py", line 63, in <module>
    train_op = optimizer.minimize(loss_placeholder, global_step=global_step)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 188, in minimize
    name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 277, in apply_gradients
    (grads_and_vars,))

  ValueError: No gradients provided for any variable: <all my variables>

所以我假设这是TensorFlow抱怨它无法计算我的损失的梯度,因为损失是由超出TF范围的numpy执行的.

所以很自然地解决这个问题,我会尝试在TensorFlow中实现它.问题是,我logit_lengths和label_lengths都是张量,所以当我尝试访问单个元素时,我返回了一个Tensor of shape [].这是当我试图用一个问题tf.nn.top_k()这需要一个Int为它k的参数.

另一个问题是我label_lengths是一个占位符,因为我loss需要在optimizer.minimize(loss)调用之前定义我的值,我还会收到一个错误,指出需要为占位符传递一个值.

我只是想知道如何尝试实现这种损失功能.或者,如果我遗漏了一些明显的东西.

---

编辑:在进一步阅读之后,我看到通常像我所描述的那样的损失用于验证和训练中使用与使用真实损失相同的最小化的替代损失.有谁知道代理损失用于像我这样的基于编辑距离的场景？

Answer 1

我要做的第一件事是使用张量流而不是 numpy 来计算损失。这将允许张量流为您计算梯度，因此您将能够反向传播，这意味着您可以最大限度地减少损失。

有 tf.edit_distance( https://www.tensorflow.org/api_docs/python/tf/edit_distance核心库中

因此，为了解决这个问题，我自然会尝试在 TensorFlow 中实现它。问题是，我的 logit_lengths 和 label_lengths 都是张量，所以当我尝试访问单个元素时，我返回一个形状为 [] 的张量。当我尝试使用 tf.nn.top_k() 时，这是一个问题，它采用 Int 作为其 k 参数。

您能否提供更多细节说明为什么这是一个问题？