Java中的模型优化器tensorflow

Question

我正在尝试用Java语言创建一个不使用Python语言的TensorFlow模型。我设法用Java编写了很多Python代码，但是我遗漏了一些要完成的元素。我阻塞了优化器。Python中的原始代码是一个非常简单的模型。

import tensorflow as tf

# Batch of input and target output (1x1 matrices)
x = tf.placeholder(tf.float32, shape=[None, 1, 1], name='input')
y = tf.placeholder(tf.float32, shape=[None, 1, 1], name='target')

# Trivial linear model
y_ = tf.identity(tf.layers.dense(x, 1), name='output')

# Optimize loss
loss = tf.reduce_mean(tf.square(y_ - y), name='loss')
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)
train_op = optimizer.minimize(loss, name='train')

init = tf.global_variables_initializer()

我开始了向Java的转换，虽然离最后还不远，但我被优化器卡住了。

try (Graph g = new Graph()) {
   //# Batch of input and target output (1x1 matrices)
   //x = tf.placeholder(tf.float32, shape=[None, 1, 1], name='input')
   Output<OperationBuilder> x = g.opBuilder("Placeholder", "input")
                .setAttr("dtype", DataType.FLOAT)
                .build().output(0);
   //y = tf.placeholder(tf.float32, shape=[None, 1, 1], name=target')
   Output<OperationBuilder> y = g.opBuilder("Placeholder", "target")
                .setAttr("dtype", DataType.FLOAT)
                .build().output(0);
   //# Trivial linear model
   //y_ = tf.identity(tf.layers.dense(x, 1), name='output')
   Tensor t = Tensor.create(new int[] {0});
   Output reductionIndices = g.opBuilder("Const", "layer")
                .setAttr("dtype", t.dataType()).setAttr("value", t)
                .build().output(0);
   Output dense = g.opBuilder("layersdense", "dense")
                .setAttr("T", DataType.FLOAT)
                .setAttr("Tidx", DataType.INT32)
                .addInput(input).addInput(reductionIndices)
                .build().output(0);

   Tensor<?> t2 = Tensor.create(dense);
   Output<OperationBuilder> y_ = g.opBuilder("Identity", "output")
                .setAttr("value", t2)
                .build().output(0);
   //# Optimize loss
   //loss = tf.reduce_mean(tf.square(y_ - y), name='loss')
   Output<OperationBuilder> sub=g.opBuilder("Sub","sub")
        .addInput(y_).addInput(y)
        .build().output(0);
   Output<OperationBuilder> sq = g.opBuilder("Square", "Square")
        .addInput(sub)
        .build().output(0);
   //optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)
   Code java ???
   //train_op = optimizer.minimize(loss, name='train')
   Code java ???

}

Answer 1

在对推断的输出进行一些后处理时，我也遇到了很多麻烦。我可以建议的一个解决方案是导入整个图形.pb文件(包括优化器和损失函数的必要节点)，该文件是使用python定义的，并在2“s.runner”循环中运行该循环，该循环提供批处理并执行时期。请注意，我还没有尝试过这种方法，但根据我的经验，避免使用特定于java的图形生成器会对您有很大帮助。祝你好运