Apache Spark:如何从DataFrame创建矩阵？

Question

我在Apache Spark中有一个带有整数数组的DataFrame,源是一组图像.我最终想在它上面做PCA,但是我在从数组中创建一个矩阵时遇到了麻烦.如何从RDD创建矩阵？

> imagerdd = traindf.map(lambda row: map(float, row.image))
> mat = DenseMatrix(numRows=206456, numCols=10, values=imagerdd)
Traceback (most recent call last):

  File "<ipython-input-21-6fdaa8cde069>", line 2, in <module>
mat = DenseMatrix(numRows=206456, numCols=10, values=imagerdd)

  File "/usr/local/spark/current/python/lib/pyspark.zip/pyspark/mllib/linalg.py", line 815, in __init__
values = self._convert_to_array(values, np.float64)

  File     "/usr/local/spark/current/python/lib/pyspark.zip/pyspark/mllib/linalg.py", line 806, in _convert_to_array
    return np.asarray(array_like, dtype=dtype)

  File "/usr/local/python/conda/lib/python2.7/site-        packages/numpy/core/numeric.py", line 462, in asarray
    return array(a, dtype, copy=False, order=order)

TypeError: float() argument must be a string or a number

我从我能想到的每一种可能的安排中得到了同样的错误:

imagerdd = traindf.map(lambda row: Vectors.dense(row.image))
imagerdd = traindf.map(lambda row: row.image)
imagerdd = traindf.map(lambda row: np.array(row.image))

如果我试试

> imagedf = traindf.select("image")
> mat = DenseMatrix(numRows=206456, numCols=10, values=imagedf)

Traceback(最近一次调用最后一次):

  File "<ipython-input-26-a8cbdad10291>", line 2, in <module>
mat = DenseMatrix(numRows=206456, numCols=10, values=imagedf)

  File "/usr/local/spark/current/python/lib/pyspark.zip/pyspark/mllib/linalg.py", line 815, in __init__
    values = self._convert_to_array(values, np.float64)

  File "/usr/local/spark/current/python/lib/pyspark.zip/pyspark/mllib/linalg.py", line 806, in _convert_to_array
    return np.asarray(array_like, dtype=dtype)

  File "/usr/local/python/conda/lib/python2.7/site-packages/numpy/core/numeric.py", line 462, in asarray
    return array(a, dtype, copy=False, order=order)

ValueError: setting an array element with a sequence.

Answer 1

由于您没有提供示例输入,因此我假设它看起来或多或少类似于id行号并image包含值.

traindf = sqlContext.createDataFrame([
    (1, [1, 2, 3]),
    (2, [4, 5, 6]),
    (3, (7, 8, 9))
], ("id", "image"))

首先要了解的是,这DenseMatrix是一个本地数据结构.确切地说,它是一个包装numpy.ndarray.至于现在(Spark 1.4.1),PySpark MLlib中没有分布式的等价物.

密集矩阵采取三个强制性参数numRows,numCols,values这里values是一个本地数据结构.在你的情况下,你必须先收集:

values = (traindf.
    rdd.
    map(lambda r: (r.id, r.image)). # Extract row id and data
    sortByKey(). # Sort by row id
    flatMap(lambda (id, image): image).
    collect())


ncol = len(traindf.rdd.map(lambda r: r.image).first())
nrow = traindf.count()

dm = DenseMatrix(nrow, ncol, values)

最后:

> print dm.toArray()
[[ 1.  4.  7.]
 [ 2.  5.  8.]
 [ 3.  6.  9.]]

编辑:

在Spark 1.5+中,您可以使用mllib.linalg.distributed如下:

from pyspark.mllib.linalg.distributed import IndexedRow, IndexedRowMatrix

mat = IndexedRowMatrix(traindf.map(lambda row: IndexedRow(*row)))
mat.numRows()
## 4
mat.numCols()
## 3

虽然目前API仍然限于在实践中有用.