Deeplearning4j - 使用nd4j导入tensorflow模型

Author: sjsdfg

Deeplearning4j - 使用nd4j导入tensorflow模型.pdf

@@ -0,0 +1,152 @@
+# Deeplearning4j - 使用nd4j导入tensorflow模型
+
+在dl4j-example里面新增了模型导入的例子，这里简单的说一下。
+
+---
+
+在dl4j新版本的特性介绍：https://github.com/deeplearning4j/deeplearning4j-docs/blob/releasenotes_100a/releasenotes.md 中，对于nd4j的模型导入进行了特别强调。
+```
+ND4J: New Features
+
+Technology preview of tensorflow import added (supports 1.4.0 and up)
+```
+其中一项就是对于tf模型的导入提供了功能预览版本，所支持的tf版本为1.4版本及其以上。
+
+并于最近增加了导入tensorflow模型的示例代码，导入模型为 MINST 手写数字分类模型。从代码注释上面来看，因为是预览版本，目前只支持cpu运行，还不支持gpu的加速。
+
+<center>![资源文件列表.png-23.3kB][1]</center>
+
+并且提供了如上的文件，用于本次示例的测试。
+
+ 1. freeze_model_after.py 和 generate_model.py 是生成模型的 python 文件。
+ 2. frozen_model.pb 为tensorflow的模型文件
+ 3. input_*.csv 和 input_*.shape为配套的特征数据文件。csv文件存放的是特征数据，一个特征一行；shape文件保存的是输入模型时的形状。
+ 4. prediction 文件同理，为预测的标签数据文件。
+
+**注：** 使用该示例的时候，最好 IDE 已经安装了的相对的 lombok 插件。
+
+# 资源文件夹
+```
+ //Python code for this can be found in resources/import/tensorflow under generate_model.py and freeze_model_after.py
+//Input node/Placeholder in this graph is names "input"
+//Output node/op in this graph is names "output"
+public final static String BASE_DIR = "import/tensorflow";
+```
+首先定义了一个根目录用于寻找对应的文件。
+
+接下来在主函数中获取模型文件的绝对路径
+```
+final String FROZEN_MLP = new ClassPathResource(BASE_DIR + "/frozen_model.pb").getFile().getPath();
+
+```
+
+# 读取tf中的占位符输入
+```
+//Load placeholder inputs and corresponding predictions generated from tensorflow
+Map<String, INDArray> inputsPredictions = readPlaceholdersAndPredictions();
+```
+这里面所用`readPlaceholdersAndPredictions`方法的全部代码如下：
+```
+//A simple helper function to load the inputs and corresponding outputs generated from tensorflow
+//Two cases: {input_a,prediction_a} and {input_b,prediction_b}
+protected static Map<String, INDArray> readPlaceholdersAndPredictions() throws IOException {
+    String[] toReadList = {"input_a", "input_b", "prediction_a", "prediction_b"};
+    Map<String, INDArray> arraysFromPython = new HashMap<>();
+    for (int i = 0; i < toReadList.length; i++) {
+        String varShapePath = new ClassPathResource(BASE_DIR + "/" + toReadList[i] + ".shape").getFile().getPath();
+        String varValuePath = new ClassPathResource(BASE_DIR + "/" + toReadList[i] + ".csv").getFile().getPath();
+        int[] varShape = Nd4j.readNumpy(varShapePath, ",").data().asInt();
+        float[] varContents = Nd4j.readNumpy(varValuePath).data().asFloat();
+        arraysFromPython.put(toReadList[i], Nd4j.create(varContents).reshape(varShape));
+    }
+    return arraysFromPython;
+}
+```
+这段代码不难理解，就是把前缀为`toReadList`数组内容中的数据成对读取出来，并且转换成为INDArray对象，并且返回回去。
+
+# 模型读取
+```
+//Load the graph into samediff
+val graph = TFGraphMapper.getInstance().importGraph(new File(FROZEN_MLP));
+```
+这里面的 val 并非java 10提供的变量自动推断，而是 lombok 所提供的功能。
+
+# 数据关联
+```
+//libnd4j executor
+//running with input_a array expecting to get prediction_a
+graph.associateArrayWithVariable(inputsPredictions.get("input_a"), graph.variableMap().get("input"));
+```
+这段代码是将从文件中读取出来的 `input_a` INDArray关联模型的数据。
+
+# 模型预测
+```
+val executioner = new NativeGraphExecutioner();
+val results = executioner.executeGraph(graph); //returns an array of the outputs
+INDArray libnd4jPred = ((INDArray[]) results)[0];
+System.out.println("LIBND4J exec prediction for input_a:\n" + libnd4jPred);
+```
+模型预测，并且获取模型的结果输出。并且将其打印到控制台上。
+
+# 结果判断
+```
+if (libnd4jPred.equals(inputsPredictions.get("prediction_a"))) {
+    //this is true and therefore predictions are equal
+    System.out.println("Predictions are equal to tensorflow");
+} else {
+    throw new RuntimeException("Predictions don't match!");
+}
+```
+用于判断结果预测，和所给的标签是否相同
+
+# 使用不同的API用于预测 input_b 的值
+```
+//Now to run with the samediff executor, with input_b array expecting to get prediction_b
+val graphSD = TFGraphMapper.getInstance().importGraph(new File(FROZEN_MLP)); //Reimport graph here, necessary for the 1.0 alpha release
+graphSD.associateArrayWithVariable(inputsPredictions.get("input_b"), graph.variableMap().get("input"));
+INDArray samediffPred = graphSD.execAndEndResult();
+System.out.println("SameDiff exec prediction for input_b:\n" + samediffPred);
+if (samediffPred.equals(inputsPredictions.get("prediction_b"))) {
+    //this is true and therefore predictions are equal
+    System.out.println("Predictions are equal to tensorflow");
+}
+```
+
+# 对模型进行新增op
+```
+//add to graph to demonstrate pytorch like capability
+System.out.println("Adding new op to graph..");
+SDVariable linspaceConstant = graphSD.var("linspace", Nd4j.linspace(1, 10, 10));
+SDVariable totalOutput = graphSD.getVariable("output").add(linspaceConstant);
+INDArray totalOutputArr = totalOutput.eval();
+System.out.println(totalOutputArr);
+```
+这个代码的意思就是对原有模型添加新的操作。
+
+ 1. 首先使用`graphSD.var("linspace", Nd4j.linspace(1, 10, 10))`获取[1,2,3 ... 10]，10个整数的向量
+ 2. `graphSD.getVariable("output").add(linspaceConstant);`将这个向量加入到模型的输出中。
+
+# 整体输出
+```
+LIBND4J exec prediction for input_a:
+[[         0,         0,         0,         0,         0,         0,         0,    1.0000,         0,         0]]
+Predictions are equal to tensorflow
+22:39:16,498 WARN  ~ No input found for Add and op name mmul
+22:39:16,498 WARN  ~ No input found for Add_1 and op name mmul
+SameDiff exec prediction for input_b:
+[[         0,         0,    1.0000,         0,         0,         0,         0,         0,         0,         0]]
+Predictions are equal to tensorflow
+Adding new op to graph..
+[[    1.0000,    2.0000,    4.0000,    4.0000,    5.0000,    6.0000,    7.0000,    8.0000,    9.0000,   10.0000]]
+```
+在最后我们可以看到，因为增加了新的op操作， 模型的原本输出`[[         0,         0,    1.0000,         0,         0,         0,         0,         0,         0,         0]]` 加上了`[1,2,3 ... 10]`就会变成对应的`[[    1.0000,    2.0000,    4.0000,    4.0000,    5.0000,    6.0000,    7.0000,    8.0000,    9.0000,   10.0000]]`。
+
+
+---
+更多文档可以查看 https://github.com/sjsdfg/deeplearning4j-issues。
+你的star是我持续分享的动力
+
+代码地址已经放在github上面，自行下载即可： https://github.com/sjsdfg/dl4j-tutorials
+ 
+ 
+  [1]: http://static.zybuluo.com/ZzzJoe/zrkt70us7dx0dmnpsqjud2w0/%E8%B5%84%E6%BA%90%E6%96%87%E4%BB%B6%E5%88%97%E8%A1%A8.png