Chapter 13: @merge multiple input into a single result¶

The previous chapter explained how Ruffus allows large jobs to be split into small pieces with @split and analysed in parallel using for example, our old friend @transform.

Having done this, our next task is to recombine the fragments into a seamless whole.

This is the role of the @merge decorator.

@transform tasks multiple inputs and produces a single output, Ruffus is again agnostic as to the sort of data contained within this single output. It can be a single (string) file name, an arbitrary complicated nested structure with numbers, objects etc. Or even a list.

The main thing is that downstream tasks will interpret this output as a single entity leading to a single job.

@split and @merge are, in other words, about network topology.

Because of this @merge is also very useful for summarising the progress in our pipeline. At key selected points, we can gather data from the multitude of data or disparate inputs and @merge them to a single set of summaries.

The previous chapter we had almost completed all the pieces of our flowchart:

What remains is to take the partial solutions from the different .sums files and turn these into the variance as follows:

variance = (sum_squared - sum * sum / N)/N

where N is the number of values

See the wikipedia entry for a discussion of why this is a very naive approach.

To do this, all we have to do is iterate through all the values in *.sums, add up the sums and sum_squared, and apply the above (naive) formula.

#
#   @merge files together
#
@merge(sum_of_squares, "variance.result")
def calculate_variance (input_file_names, output_file_name):
    """
    Calculate variance naively
    """
    #
    #   initialise variables
    #
    all_sum_squared = 0.0
    all_sum         = 0.0
    all_cnt_values  = 0.0
    #
    # added up all the sum_squared, and sum and cnt_values from all the chunks
    #
    for input_file_name in input_file_names:
        sum_squared, sum, cnt_values = map(float, open(input_file_name).readlines())
        all_sum_squared += sum_squared
        all_sum         += sum
        all_cnt_values  += cnt_values
    all_mean = all_sum / all_cnt_values
    variance = (all_sum_squared - all_sum * all_mean)/(all_cnt_values)
    #
    #   print output
    #
    open(output_file_name,  "w").write("%s\n" % variance)

This results in the following equivalent function call:

calculate_variance (["1.sums", "2.sums", "3.sums",
                     "4.sums", "5.sums", "6.sums",
                     "7.sums", "8.sums", "9.sums, "10.sums"], "variance.result")

and the following display:

>>> pipeline_run()
    Job  = [[1.sums, 10.sums, 2.sums, 3.sums, 4.sums, 5.sums, 6.sums, 7.sums, 8.sums, 9.sums] -> variance.result] completed
Completed Task = calculate_variance

The final result is in variance.result

Have a look at the complete example code for this chapter.