[英]How can I feed a sparse placeholder in a TensorFlow model from Java
我正在嘗試使用 TensorFlow 中的 kNN 算法計算給定地址的最佳匹配,該算法效果很好,但是當我嘗試導出模型並在我們的 Java 環境中使用它時,我陷入了如何提供來自 Java 的稀疏占位符。
這是 python 部分的一個非常精簡的版本,它返回測試名稱和最佳參考名稱之間的最小距離。 到目前為止,這項工作符合預期。 當我導出模型並將其導入我的 Java 程序時,它總是返回相同的值(占位符默認的距離)。 我假設,python 函數sparse_from_word_vec(word_vec)
不在模型中,這對我來說完全有意義,但是我應該如何制作這個稀疏張量? 我的輸入是一個字符串,我需要創建一個合適的稀疏張量(值)來計算距離。 我也在Java端搜索了一種生成稀疏張量的方法,但沒有成功。
import tensorflow as tf
import pandas as pd
d = {'NAME': ['max mustermann',
'erika musterfrau',
'joseph haydn',
'johann sebastian bach',
'wolfgang amadeus mozart']}
df = pd.DataFrame(data=d)
input_name = tf.placeholder_with_default('max musterman',(), name='input_name')
output_dist = tf.placeholder(tf.float32, (), name='output_dist')
test_name = tf.sparse_placeholder(dtype=tf.string)
ref_names = tf.sparse_placeholder(dtype=tf.string)
output_dist = tf.edit_distance(test_name, ref_names, normalize=True)
def sparse_from_word_vec(word_vec):
num_words = len(word_vec)
indices = [[xi, 0, yi] for xi,x in enumerate(word_vec) for yi,y in enumerate(x)]
chars = list(''.join(word_vec))
return(tf.SparseTensorValue(indices, chars, [num_words,1,1]))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
t_data_names=tf.constant(df['NAME'])
reference_names = [el.decode('UTF-8') for el in (t_data_names.eval())]
sparse_ref_names = sparse_from_word_vec(reference_names)
sparse_test_name = sparse_from_word_vec([str(input_name.eval().decode('utf-8'))]*5)
feeddict={test_name: sparse_test_name,
ref_names: sparse_ref_names,
}
output_dist = sess.run(output_dist, feed_dict=feeddict)
output_dist = tf.reduce_min(output_dist, 0)
print(output_dist.eval())
tf.saved_model.simple_save(sess,
"model-simple",
inputs={"input_name": input_name},
outputs={"output_dist": output_dist})
這是我的 Java 方法:
public void run(ApplicationArguments args) throws Exception {
log.info("Loading model...");
SavedModelBundle savedModelBundle = SavedModelBundle.load("/model", "serve");
byte[] test_name = "Max Mustermann".toLowerCase().getBytes("UTF-8");
List<Tensor<?>> output = savedModelBundle.session().runner()
.feed("input_name", Tensor.<String>create(test_names))
.fetch("output_dist")
.run();
System.out.printl("Nearest distance: " + output.get(0).floatValue());
}
我能夠讓你的例子工作。 在深入研究之前,我對您的 Python 代碼有一些評論。
您在整個代碼中將變量output_dist
用於 3 種不同的值類型。 我不是 Python 專家,但我認為這是不好的做法。 您也從未實際使用input_name
占位符,除非將其導出為輸入。 最后一個是tf.saved_model.simple_save
已棄用,您應該改用tf.saved_model.Builder
。
現在為解決方案。
使用命令jar tvf libtensorflow-xxxjar
(感謝這篇文章)查看libtensorflow
jar 文件,您可以看到沒有用於創建稀疏張量的有用綁定(也許提出功能請求?)。 因此,我們必須將輸入更改為稠密張量,然后向圖中添加操作以將其轉換為稀疏張量。 在您的原始代碼中,稀疏轉換在 python 端,這意味着 java 中加載的圖形不會有任何操作。
這是新的python代碼:
import tensorflow as tf
import pandas as pd
def model():
#use dense tensors then convert to sparse for edit_distance
test_name = tf.placeholder(shape=(None, None), dtype=tf.string, name="test_name")
ref_names = tf.placeholder(shape=(None, None), dtype=tf.string, name="ref_names")
#Java Does not play well with the empty character so use "/" instead
test_name_sparse = tf.contrib.layers.dense_to_sparse(test_name, "/")
ref_names_sparse = tf.contrib.layers.dense_to_sparse(ref_names, "/")
output_dist = tf.edit_distance(test_name_sparse, ref_names_sparse, normalize=True)
#output the index to the closest ref name
min_idx = tf.argmin(output_dist)
return test_name, ref_names, min_idx
#Python code to be replicated in Java
def pad_string(s, max_len):
return s + ["/"] * (max_len - len(s))
d = {'NAME': ['joseph haydn',
'max mustermann',
'erika musterfrau',
'johann sebastian bach',
'wolfgang amadeus mozart']}
df = pd.DataFrame(data=d)
input_name = 'max musterman'
#pad dense tensor input
max_len = max([len(n) for n in df['NAME']])
test_input = [list(input_name)]*len(df['NAME'])
#no need to pad, all same length
ref_input = list(map(lambda x: pad_string(x, max_len), [list(n) for n in df['NAME']]))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
test_name, ref_names, min_idx = model()
#run a test to make sure the model works
feeddict = {test_name: test_input,
ref_names: ref_input,
}
out = sess.run(min_idx, feed_dict=feeddict)
print("test output:", out)
#save the model with the new Builder API
signature_def_map= {
"predict": tf.saved_model.signature_def_utils.predict_signature_def(
inputs= {"test_name": test_name, "ref_names": ref_names},
outputs= {"min_idx": min_idx})
}
builder = tf.saved_model.Builder("model")
builder.add_meta_graph_and_variables(sess, ["serve"], signature_def_map=signature_def_map)
builder.save()
這是加載和運行它的java。 這里可能有很大的改進空間(java 不是我的主要語言),但它給了你這個想法。
import org.tensorflow.Graph;
import org.tensorflow.Session;
import org.tensorflow.Tensor;
import org.tensorflow.TensorFlow;
import org.tensorflow.SavedModelBundle;
import java.util.ArrayList;
import java.util.List;
import java.util.Arrays;
public class Test {
public static byte[][] makeTensor(String s, int padding) throws Exception
{
int len = s.length();
int extra = padding - len;
byte[][] ret = new byte[len + extra][];
for (int i = 0; i < len; i++) {
String cur = "" + s.charAt(i);
byte[] cur_b = cur.getBytes("UTF-8");
ret[i] = cur_b;
}
for (int i = 0; i < extra; i++) {
byte[] cur = "/".getBytes("UTF-8");
ret[len + i] = cur;
}
return ret;
}
public static byte[][][] makeTensor(List<String> l, int padding) throws Exception
{
byte[][][] ret = new byte[l.size()][][];
for (int i = 0; i < l.size(); i++) {
ret[i] = makeTensor(l.get(i), padding);
}
return ret;
}
public static void main(String[] args) throws Exception {
System.out.println("Loading model...");
SavedModelBundle savedModelBundle = SavedModelBundle.load("model", "serve");
List<String> str_test_name = Arrays.asList("Max Mustermann",
"Max Mustermann",
"Max Mustermann",
"Max Mustermann",
"Max Mustermann");
List<String> names = Arrays.asList("joseph haydn",
"max mustermann",
"erika musterfrau",
"johann sebastian bach",
"wolfgang amadeus mozart");
//get the max length for each array
int pad1 = str_test_name.get(0).length();
int pad2 = 0;
for (String var : names) {
if(var.length() > pad2)
pad2 = var.length();
}
byte[][][] test_name = makeTensor(str_test_name, pad1);
byte[][][] ref_names = makeTensor(names, pad2);
//use a with block so the close method is called
try(Tensor t_test_name = Tensor.<String>create(test_name))
{
try (Tensor t_ref_names = Tensor.<String>create(ref_names))
{
List<Tensor<?>> output = savedModelBundle.session().runner()
.feed("test_name", t_test_name)
.feed("ref_names", t_ref_names)
.fetch("ArgMin")
.run();
System.out.println("Nearest distance: " + output.get(0).longValue());
}
}
}
}
聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.