如何将形状参数传递给tf.nn.ctc_greedy_decoder？

Question

我试图对加载的图形执行推断：

ds_graph = load_graph(model)
graph_input = ds_graph.get_tensor_by_name('prefix/input_node:0')
graph_seqlength = ds_graph.get_tensor_by_name('prefix/input_lengths:0')
graph_output = ds_graph.get_tensor_by_name('prefix/output_node:0')

我正在迭代的变量是

inp[i]
sl[i]

在循环中

for i in range(num):
    with tf.Session(graph=ds_graph) as sess:
        logits = sess.run(graph_output,feed_dict={graph_input:inp[i],graph_seqlength:sl[i]})
        logits = tf.nn.softmax(logits, dim=-1, name=None)
        logits = sess.run(logits)
        output_length=np.array([logits.shape[0]])
        tf_greedy_path, _ = tf.nn.ctc_greedy_decoder(logits,output_length,merge_repeated=True)
        tf_greedy_path = tf.convert_to_tensor([tf.sparse_tensor_to_dense(sparse_tensor) for sparse_tensor in tf_greedy_path])
        greed_out = ndarray_to_text(sess.run(tf_greedy_path)[0][0])  

    return greed_out

我知道这个片段在每次迭代时都会将元素添加到图中。但我不知道如何具体解决这个问题。

我有限的理解告诉我要在循环之外创建图形元素：

logits = tf.nn.softmax(graph_output, dim=-1, name=None)
tf_greedy_path, _ = tf.nn.ctc_greedy_decoder(logits,output_length,merge_repeated=True)
tf_greedy_path = tf.convert_to_tensor([tf.sparse_tensor_to_dense(sparse_tensor) for sparse_tensor in tf_greedy_path])

for i in range(num):
    with tf.Session(graph=ds_graph) as sess:
        sess.run(graph_output,feed_dict={graph_input:inp[i],graph_seqlength:sl[i]})
        sess.run(logits)
        output_length=np.array([logits.shape[0]])
        greed_out = ndarray_to_text(sess.run(tf_greedy_path)[0][0]) 

但我仍然需要处理一个事实，即output_length是在执行过程中计算出来的。不幸的是，ctc_greedy_decoder没有将output_length作为张量。否则我会通过tf.shape(logits)的

Answer 1

没有完整的代码是很难回答的，但是是的，您是对的，您应该在进入循环之前将所有操作添加到图中。似乎没有什么能阻止您使用tensor shape of您的graph_output张量(顺便记住，不需要中间调用，只需计算您感兴趣的张量，任何中间张量都会自动计算)：

import tensorflow as tf

graph_output = tf.placeholder(tf.float32, shape=[None, 1, 2]) # graph_output has a dynamic shape
logits = tf.nn.softmax(graph_output, dim=-1, name=None)
tf_greedy_path, _ = tf.nn.ctc_greedy_decoder(logits,[graph_output.shape[0]],merge_repeated=True)
tf_greedy_path = tf.convert_to_tensor([tf.sparse_tensor_to_dense(sparse_tensor) for sparse_tensor in tf_greedy_path])

for i in range(10):
    with tf.Session() as sess:
        print(sess.run(tf_greedy_path, feed_dict={graph_output:[[[1., 2.]], [[3., 4.]]]})))