Creates a pie chart from a GNU Parallel joblog after running OCR-D

evaluate-speed-pie-chart.py

# -*- coding:utf-8 -*-

from itertools import dropwhile
import json
import re
import sys

import matplotlib.pyplot as plt


JOBLOG_FILENAME = sys.argv[1]
OUTPUT_FILENAME = sys.argv[2]

COMMAND_REGEX = re.compile(r'.* -O ([^" ]*).*|.* (ocrd-import) .*')
LABEL_MAP = {
  "ocrd-import": "ocrd-import",
  "OCR-D-BIN": "olena-binarize",
  "OCR-D-CROP": "anybaseocr-crop",
  "OCR-D-BIN2": "olena-binarize",
  "OCR-D-BIN-DENOISE": "cis-ocropy-denoise",
  "OCR-D-BIN-DENOISE-DESKEW": "cis-ocropy-deskew",
  "OCR-D-SEG-REG": "tesserocr-segment-region",
  "OCR-D-SEG-REPAIR": "segment-repair",
  "OCR-D-SEG-REG-DESKEW": "cis-ocropy-deskew",
  "OCR-D-SEG-REG-DESKEW-CLIP": "cis-ocropy-clip",
  "OCR-D-SEG-LINE": "cis-ocropy-segment",
  "OCR-D-SEG-REPAIR-LINE": "segment-repair",
  "OCR-D-SEG-LINE-RESEG-DEWARP": "cis-ocropy-dewarp",
  "OCR-D-OCR": "calamari-recognize",
  "OCR-D-TEXT": "fileformat-transform",
  "OCR-D-HOCR": "fileformat-transform",
}


def read_joblog(filename):
    successfully_read = 0
    with open(filename, 'rt') as f:
        lines = iter(f)
        next(lines)
        for line in lines:
            line = line.rstrip('\r\n').split('\t')
            runtime = float(line[3])
            exit_code = int(line[6])
            command = line[-1]
            successfully_read += 1
            yield (command, runtime, exit_code)


def evaluate():
    runtimes = {}
    joblog = read_joblog(JOBLOG_FILENAME)
    for command, runtime, exit_code in joblog:
        if exit_code != 0:
            continue
        match = re.fullmatch(COMMAND_REGEX, command)
        command = list(dropwhile(lambda x: x is None, match.groups()))[0]
        label = LABEL_MAP[command]
        if label not in runtimes:
            runtimes[label] = []
        runtimes[label].append(runtime)
    return runtimes


def pie_chart(runtimes):
    labels, sizes = zip(*runtimes.items())
    sizes = list(map(sum, sizes))
    colors = plt.get_cmap('tab20').colors
    patches, texts = plt.pie(sizes, colors=colors)
    plt.legend(patches, labels, loc='best')
    plt.axis('equal')
    plt.tight_layout()
    plt.savefig(OUTPUT_FILENAME)


def main():
    runtimes = evaluate()
    pie_chart(runtimes)


if __name__ == '__main__':
    main()

A GNU Parallel joblog may look like this:

$ cat joblog
Seq	Host	Starttime	JobRuntime	Send	Receive	Exitval	Signal	Command
39	epimetheus3	1601552167.824	   173.330	0	148	0	0	docker run --rm -u `id -u` -v /var/tmp/ocrd-workspace/1139/:/data -w /data -v /var/tmp/tesseract/calamari_models:/models -- ocrd/all:maximum nice -n 10 ocrd-import -P -i -r 300 /data
40	epimetheus3	1601552167.939	   184.570	0	148	0	0	docker run --rm -u `id -u` -v /var/tmp/ocrd-workspace/1138/:/data -w /data -v /var/tmp/tesseract/calamari_models:/models -- ocrd/all:maximum nice -n 10 ocrd-import -P -i -r 300 /data
56	epimetheus4	1601552169.878	   204.542	0	6570	0	0	docker run --rm -u `id -u` -v /var/tmp/ocrd-workspace/1141/:/data -w /data -v /var/tmp/tesseract/calamari_models:/models -- ocrd/all:maximum nice -n 10 ocrd-import -P -i -r 300 /data
...
3092	epimetheus1	1602657979.219	  9116.075	0	464	0	0	docker run --rm -u `id -u` -v /var/tmp/ocrd-workspace/837/:/data -w /data -v /var/tmp/tesseract/calamari_models:/models -- ocrd/all:maximum nice -n 10 ocrd process "fileformat-transform -I OCR-D-OCR -O OCR-D-HOCR -P from-to page\ hocr"
3064	epimetheus3	1602655541.439	 11996.573	0	464	0	0	docker run --rm -u `id -u` -v /var/tmp/ocrd-workspace/119/:/data -w /data -v /var/tmp/tesseract/calamari_models:/models -- ocrd/all:maximum nice -n 10 ocrd process "fileformat-transform -I OCR-D-OCR -O OCR-D-HOCR -P from-to page\ hocr"
3107	epimetheus3	1602659504.950	  8612.075	0	464	0	0	docker run --rm -u `id -u` -v /var/tmp/ocrd-workspace/358/:/data -w /data -v /var/tmp/tesseract/calamari_models:/models -- ocrd/all:maximum nice -n 10 ocrd process "fileformat-transform -I OCR-D-OCR -O OCR-D-HOCR -P from-to page\ hocr"

You would run the gist as follows:

$ python3 evaluate-speed-pie-chart.py joblog pie-chart.pdf

After running the gist, you would end up with this beautiful pie chart:

COMMAND_REGEX = re.compile(r'.* -O ([^" ]*).*|.* (ocrd-import) .*')
...
match = re.fullmatch(COMMAND_REGEX, command)
command = list(dropwhile(lambda x: x is None, match.groups()))[0]
label = LABEL_MAP[command]

Just wondering why you need the mapping from output group to command in LABEL_MAP?
The command you are searching for is also in the text and retrievable via regex (r'ocrd process "([^" ]+)|(ocrd-import)').

Just wondering why you need the mapping from output group to command in LABEL_MAP?
The command you are searching for is also in the text and retrievable via regex (r'ocrd process "([^" ]+)|(ocrd-import)').

You are right, we could simplify things and get rid of LABEL_MAP. Then again, the solution with COMMAND_REGEX and LABEL_MAP is more general, will work for many sorts of joblogs beside OCR-D (in my code base, both LABEL_MAP and COMMAND_REGEX are command-line arguments), and allows you to name the labels independently on the command names. For example, you could merge cis-ocropy-clip, cis-ocropy-segment, and cis-ocropy-dewarp into a single slice named cis-ocropy:

LABEL_MAP = {
  # [...]
  "OCR-D-BIN-DENOISE": "cis-ocropy",
  "OCR-D-BIN-DENOISE-DESKEW": "cis-ocropy",
  "OCR-D-SEG-REG-DESKEW": "cis-ocropy",
  "OCR-D-SEG-REG-DESKEW-CLIP": "cis-ocropy",
  "OCR-D-SEG-LINE": "cis-ocropy",
  "OCR-D-SEG-LINE-RESEG-DEWARP": "cis-ocropy",
}