Question

我正在尝试使用 tf.TextLineReader（）读取CSV文件（以高效的方式，如Google推荐）。

正如您所看到的，我可以创建批次，我将在稍后用它来训练我的基本回归模型。问题是 tf.train.shuffle_batch 会返回一个张量。这意味着我不能在我的模型中使用占位符X Y_pred = X * W + b。在训练阶段，我现在使用 batch_variable 作为占位符，但我需要一个占位符才能使用不同的数据进行测试。

我做错了什么？

编辑：我修改了尼古拉斯建议的代码（非常感谢！）但现在我的准确度达到了0.0 ......这有点尴尬。我正在使用鲍鱼数据集（8个特征和3个类），其中我用一热矢量（1,0,0）改变了（M）ale（F）emale（I）nfant分类结果，（0， 1,0）......

 def getPartitionedDatasets(filenames):

    filename_queue = tf.train.string_input_producer(filenames)
    reader = tf.TextLineReader()
    _, value = reader.read(filename_queue) # return a key and value (key is for debugging

    record_defaults = [[1.0] for _ in range(N_FEATURES+1)]
    cont = tf.decode_csv(value, record_defaults=record_defaults)

    features = tf.stack([cont[1],cont[2],cont[3],cont[4],cont[5],cont[6],cont[7],cont[8]])


    label = tf.to_int32(cont[0])

    min_after_dequeue = 10 * BATCH_SIZE
    capacity = 20 * BATCH_SIZE

    data_batch, label_batch_raw = tf.train.shuffle_batch([features, label], batch_size=BATCH_SIZE,
                                        capacity=capacity, min_after_dequeue=min_after_dequeue)
    label_batch_hot = tf.one_hot(label_batch_raw,on_value=1,off_value=0, depth=3)

    return data_batch, label_batch_hot

def get_model_params():
    """Build the model parameters."""
    W = tf.Variable(tf.zeros([N_FEATURES,CLASSES]), name='weights')
    b = tf.Variable(tf.zeros([CLASSES]), name = "bias")
    return W, b

def build_model(data_batch, label_batch_hot, w, b):
    """Use `W` and `b` to build a model based on some data and labels."""
    Y_predicted = tf.matmul(data_batch , w) + b
    cross_entropy = tf.reduce_mean(
        tf.nn.softmax_cross_entropy_with_logits(labels=label_batch_hot,
                                                logits=Y_predicted))
    return Y_predicted, cross_entropy

def get_optimizer_op(cross_entropy_op):
    """Return an optimiser associated to a cross entropy op."""
    return tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy_op)



def run():
    train_data, train_label = getPartitionedDatasets(["ABA_Train.csv"])
    test_data, test_label   = getPartitionedDatasets(["ABA_Test.csv"])
    W,b = get_model_params();
    train_predicted, train_cross_entropy = build_model(train_data, train_label, W, b)
    optimizer = get_optimizer_op(train_cross_entropy)

    test_predicted, test_cross_entropy = build_model(test_data, test_label, W, b)

    correct_prediction = tf.equal(tf.argmax(test_predicted, 1), tf.argmax(test_label, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))


    with tf.Session() as sess:
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(coord=coord)
        sess.run(tf.global_variables_initializer())

        for i in range(300):  # generate 10 batches
            sess.run([optimizer])
        print(sess.run(accuracy))


def main():
    run();
if __name__ == '__main__':
    main()

Answer 1

你可以做的是实现各种方法

定义您的模型参数，
根据这些参数和某些数据构建模型，
根据您的模型构建优化器。

例如，你可以有类似的东西（假设您的火车和测试数据在两组不同的文件中）：

fun main(args: Array<String>) {
//  doNoSubjectExample()
  doSubjectExample()
}

private fun doSubjectExample() {
  val disposable = TwitterSubject().observe().subscribe(
      { status -> println("Status: {$status)}") },
      { error -> println("Error callback: $error") })

  TimeUnit.SECONDS.sleep(10)
  disposable.dispose()
}

private class TwitterSubject {
  val subject = PublishSubject.create<Status>()

  init {
    val twitterStream = TwitterStreamFactory().instance
    // See: https://stackoverflow.com/questions/37672023/how-to-create-an-instance-of-anonymous-interface-in-kotlin/37672334
    val listner = object : StatusListener {
      override fun onStatus(status: Status?) {
        subject.onNext(status)
      }

      override fun onException(ex: Exception?) {
        subject.onError(ex)
      }

      override fun onTrackLimitationNotice(numberOfLimitedStatuses: Int) {
        // Not implemented.
      }

      override fun onStallWarning(warning: StallWarning?) {
        // Not implemented.
      }

      override fun onDeletionNotice(statusDeletionNotice: StatusDeletionNotice?) {
        // Not implemented.
      }

      override fun onScrubGeo(userId: Long, upToStatusId: Long) {
        // Not implemented.
      }
    }

    Twitter4JHelper.addStatusListner(twitterStream, listner)

    twitterStream.sample()
  }

  fun observe(): Observable<Status> = subject
}

TensorFlow CSV输入

1 个答案: