Source code for

Utilities for comparing image results.

import atexit
import functools
import hashlib
import logging
import os
from pathlib import Path
import re
import shutil
import subprocess
import sys
from tempfile import TemporaryDirectory, TemporaryFile
import weakref

import numpy as np
from PIL import Image

import matplotlib as mpl
from matplotlib import _api, cbook
from matplotlib.testing.exceptions import ImageComparisonFailure

_log = logging.getLogger(__name__)

__all__ = ['calculate_rms', 'comparable_formats', 'compare_images']

def make_test_filename(fname, purpose):
    Make a new filename by inserting *purpose* before the file's extension.
    base, ext = os.path.splitext(fname)
    return '%s-%s%s' % (base, purpose, ext)

def _get_cache_path():
    cache_dir = Path(mpl.get_cachedir(), 'test_cache')
    cache_dir.mkdir(parents=True, exist_ok=True)
    return cache_dir

def get_cache_dir():
    return str(_get_cache_path())

def get_file_hash(path, block_size=2 ** 20):
    md5 = hashlib.md5()
    with open(path, 'rb') as fd:
        while True:
            data =
            if not data:

    if Path(path).suffix == '.pdf':
    elif Path(path).suffix == '.svg':

    return md5.hexdigest()

def make_external_conversion_command(cmd):
    def convert(old, new):
        cmdline = cmd(old, new)
        pipe = subprocess.Popen(cmdline, universal_newlines=True,
                                stdout=subprocess.PIPE, stderr=subprocess.PIPE)
        stdout, stderr = pipe.communicate()
        errcode = pipe.wait()
        if not os.path.exists(new) or errcode:
            msg = "Conversion command failed:\n%s\n" % ' '.join(cmdline)
            if stdout:
                msg += "Standard output:\n%s\n" % stdout
            if stderr:
                msg += "Standard error:\n%s\n" % stderr
            raise IOError(msg)

    return convert

# Modified from
_find_unsafe_bytes = re.compile(br'[^[email protected]%+=:,./-]').search

def _shlex_quote_bytes(b):
    return (b if _find_unsafe_bytes(b) is None
            else b"'" + b.replace(b"'", b"'\"'\"'") + b"'")

class _ConverterError(Exception):

class _Converter:
    def __init__(self):
        self._proc = None
        # Explicitly register deletion from an atexit handler because if we
        # wait until the object is GC'd (which occurs later), then some module
        # globals (e.g. signal.SIGKILL) has already been set to None, and
        # kill() doesn't work anymore...

    def __del__(self):
        if self._proc:
            for stream in filter(None, [self._proc.stdin,
            self._proc = None

    def _read_until(self, terminator):
        """Read until the prompt is reached."""
        buf = bytearray()
        while True:
            c =
            if not c:
                raise _ConverterError
            if buf.endswith(terminator):
                return bytes(buf[:-len(terminator)])

class _GSConverter(_Converter):
    def __call__(self, orig, dest):
        if not self._proc:
            self._proc = subprocess.Popen(
                 "-dNOSAFER", "-dNOPAUSE", "-sDEVICE=png16m"],
                # As far as I can see, ghostscript never outputs to stderr.
                stdin=subprocess.PIPE, stdout=subprocess.PIPE)
            except _ConverterError as err:
                raise OSError("Failed to start Ghostscript") from err

        def encode_and_escape(name):
            return (os.fsencode(name)
                    .replace(b"\\", b"\\\\")
                    .replace(b"(", br"\(")
                    .replace(b")", br"\)"))

            b"<< /OutputFile ("
            + encode_and_escape(dest)
            + b") >> setpagedevice ("
            + encode_and_escape(orig)
            + b") run flush\n")
        # GS> if nothing left on the stack; GS<n> if n items left on the stack.
        err = self._read_until(b"GS")
        stack = self._read_until(b">")
        if stack or not os.path.exists(dest):
            stack_size = int(stack[1:]) if stack else 0
            self._proc.stdin.write(b"pop\n" * stack_size)
            # Using the systemencoding should at least get the filenames right.
            raise ImageComparisonFailure(
                (err + b"GS" + stack + b">")
                .decode(sys.getfilesystemencoding(), "replace"))

class _SVGConverter(_Converter):
    def __call__(self, orig, dest):
        old_inkscape = mpl._get_executable_info("inkscape").version < "1"
        terminator = b"\n>" if old_inkscape else b"> "
        if not hasattr(self, "_tmpdir"):
            self._tmpdir = TemporaryDirectory()
            # On Windows, we must make sure that self._proc has terminated
            # (which __del__ does) before clearing _tmpdir.
            weakref.finalize(self._tmpdir, self.__del__)
        if (not self._proc  # First run.
                or self._proc.poll() is not None):  # Inkscape terminated.
            env = {
                # If one passes e.g. a png file to Inkscape, it will try to
                # query the user for conversion options via a GUI (even with
                # `--without-gui`).  Unsetting `DISPLAY` prevents this (and
                # causes GTK to crash and Inkscape to terminate, but that'll
                # just be reported as a regular exception below).
                "DISPLAY": "",
                # Do not load any user options.
                "INKSCAPE_PROFILE_DIR": os.devnull,
            # Old versions of Inkscape (e.g. seem to sometimes
            # deadlock when stderr is redirected to a pipe, so we redirect it
            # to a temporary file instead.  This is not necessary anymore as of
            # Inkscape 0.92.1.
            stderr = TemporaryFile()
            self._proc = subprocess.Popen(
                ["inkscape", "--without-gui", "--shell"] if old_inkscape else
                ["inkscape", "--shell"],
                stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=stderr,
            # Slight abuse, but makes shutdown handling easier.
            self._proc.stderr = stderr
            except _ConverterError as err:
                raise OSError("Failed to start Inkscape in interactive "
                              "mode") from err

        # Inkscape's shell mode does not support escaping metacharacters in the
        # filename ("\n", and ":;" for inkscape>=1).  Avoid any problems by
        # running from a temporary directory and using fixed filenames.
        inkscape_orig = Path(, os.fsdecode(b"f.svg"))
        inkscape_dest = Path(, os.fsdecode(b"f.png"))
        except OSError:
            shutil.copyfile(orig, inkscape_orig)
            b"f.svg --export-png=f.png\n" if old_inkscape else
        except _ConverterError as err:
            # Inkscape's output is not localized but gtk's is, so the output
            # stream probably has a mixed encoding.  Using the filesystem
            # encoding should at least get the filenames right...
            raise ImageComparisonFailure(
                    sys.getfilesystemencoding(), "replace")) from err
        shutil.move(inkscape_dest, dest)

    def __del__(self):
        if hasattr(self, "_tmpdir"):

def _update_converter():
    except mpl.ExecutableNotFoundError:
        converter['pdf'] = converter['eps'] = _GSConverter()
    except mpl.ExecutableNotFoundError:
        converter['svg'] = _SVGConverter()

#: A dictionary that maps filename extensions to functions which
#: themselves map arguments `old` and `new` (filenames) to a list of strings.
#: The list can then be passed to Popen to convert files with that
#: extension to png format.
converter = {}

[docs]def comparable_formats(): """ Return the list of file formats that `.compare_images` can compare on this system. Returns ------- list of str E.g. ``['png', 'pdf', 'svg', 'eps']``. """ return ['png', *converter]
def convert(filename, cache): """ Convert the named file to png; return the name of the created file. If *cache* is True, the result of the conversion is cached in `matplotlib.get_cachedir() + '/test_cache/'`. The caching is based on a hash of the exact contents of the input file. Old cache entries are automatically deleted as needed to keep the size of the cache capped to twice the size of all baseline images. """ path = Path(filename) if not path.exists(): raise IOError(f"{path} does not exist") if path.suffix[1:] not in converter: import pytest pytest.skip(f"Don't know how to convert {path.suffix} files to png") newpath = path.parent / f"{path.stem}_{path.suffix[1:]}.png" # Only convert the file if the destination doesn't already exist or # is out of date. if not newpath.exists() or newpath.stat().st_mtime < path.stat().st_mtime: cache_dir = _get_cache_path() if cache else None if cache_dir is not None: _register_conversion_cache_cleaner_once() hash_value = get_file_hash(path) cached_path = cache_dir / (hash_value + newpath.suffix) if cached_path.exists(): _log.debug("For %s: reusing cached conversion.", filename) shutil.copyfile(cached_path, newpath) return str(newpath) _log.debug("For %s: converting to png.", filename) converter[path.suffix[1:]](path, newpath) if cache_dir is not None: _log.debug("For %s: caching conversion result.", filename) shutil.copyfile(newpath, cached_path) return str(newpath) def _clean_conversion_cache(): # This will actually ignore mpl_toolkits baseline images, but they're # relatively small. baseline_images_size = sum( path.stat().st_size for path in Path(mpl.__file__).parent.glob("**/baseline_images/**/*")) # 2x: one full copy of baselines, and one full copy of test results # (actually an overestimate: we don't convert png baselines and results). max_cache_size = 2 * baseline_images_size # Reduce cache until it fits. with cbook._lock_path(_get_cache_path()): cache_stat = { path: path.stat() for path in _get_cache_path().glob("*")} cache_size = sum(stat.st_size for stat in cache_stat.values()) paths_by_atime = sorted( # Oldest at the end. cache_stat, key=lambda path: cache_stat[path].st_atime, reverse=True) while cache_size > max_cache_size: path = paths_by_atime.pop() cache_size -= cache_stat[path].st_size path.unlink() @functools.lru_cache() # Ensure this is only registered once. def _register_conversion_cache_cleaner_once(): atexit.register(_clean_conversion_cache) def crop_to_same(actual_path, actual_image, expected_path, expected_image): # clip the images to the same size -- this is useful only when # comparing eps to pdf if actual_path[-7:-4] == 'eps' and expected_path[-7:-4] == 'pdf': aw, ah, ad = actual_image.shape ew, eh, ed = expected_image.shape actual_image = actual_image[int(aw / 2 - ew / 2):int( aw / 2 + ew / 2), int(ah / 2 - eh / 2):int(ah / 2 + eh / 2)] return actual_image, expected_image
[docs]def calculate_rms(expected_image, actual_image): """ Calculate the per-pixel errors, then compute the root mean square error. """ if expected_image.shape != actual_image.shape: raise ImageComparisonFailure( "Image sizes do not match expected size: {} " "actual size {}".format(expected_image.shape, actual_image.shape)) # Convert to float to avoid overflowing finite integer types. return np.sqrt(((expected_image - actual_image).astype(float) ** 2).mean())
# NOTE: compare_image and save_diff_image assume that the image does not have # 16-bit depth, as Pillow converts these to RGB incorrectly.
[docs]def compare_images(expected, actual, tol, in_decorator=False): """ Compare two "image" files checking differences within a tolerance. The two given filenames may point to files which are convertible to PNG via the `.converter` dictionary. The underlying RMS is calculated with the `.calculate_rms` function. Parameters ---------- expected : str The filename of the expected image. actual : str The filename of the actual image. tol : float The tolerance (a color value difference, where 255 is the maximal difference). The test fails if the average pixel difference is greater than this value. in_decorator : bool Determines the output format. If called from image_comparison decorator, this should be True. (default=False) Returns ------- None or dict or str Return *None* if the images are equal within the given tolerance. If the images differ, the return value depends on *in_decorator*. If *in_decorator* is true, a dict with the following entries is returned: - *rms*: The RMS of the image difference. - *expected*: The filename of the expected image. - *actual*: The filename of the actual image. - *diff_image*: The filename of the difference image. - *tol*: The comparison tolerance. Otherwise, a human-readable multi-line string representation of this information is returned. Examples -------- :: img1 = "./baseline/plot.png" img2 = "./output/plot.png" compare_images(img1, img2, 0.001) """ actual = os.fspath(actual) if not os.path.exists(actual): raise Exception("Output image %s does not exist." % actual) if os.stat(actual).st_size == 0: raise Exception("Output image file %s is empty." % actual) # Convert the image to png expected = os.fspath(expected) if not os.path.exists(expected): raise IOError('Baseline image %r does not exist.' % expected) extension = expected.split('.')[-1] if extension != 'png': actual = convert(actual, cache=True) expected = convert(expected, cache=True) # open the image files and remove the alpha channel (if it exists) expected_image = np.asarray("RGB")) actual_image = np.asarray("RGB")) actual_image, expected_image = crop_to_same( actual, actual_image, expected, expected_image) diff_image = make_test_filename(actual, 'failed-diff') if tol <= 0: if np.array_equal(expected_image, actual_image): return None # convert to signed integers, so that the images can be subtracted without # overflow expected_image = expected_image.astype(np.int16) actual_image = actual_image.astype(np.int16) rms = calculate_rms(expected_image, actual_image) if rms <= tol: return None save_diff_image(expected, actual, diff_image) results = dict(rms=rms, expected=str(expected), actual=str(actual), diff=str(diff_image), tol=tol) if not in_decorator: # Then the results should be a string suitable for stdout. template = ['Error: Image files did not match.', 'RMS Value: {rms}', 'Expected: \n {expected}', 'Actual: \n {actual}', 'Difference:\n {diff}', 'Tolerance: \n {tol}', ] results = '\n '.join([line.format(**results) for line in template]) return results
def save_diff_image(expected, actual, output): """ Parameters ---------- expected : str File path of expected image. actual : str File path of actual image. output : str File path to save difference image to. """ # Drop alpha channels, similarly to compare_images. expected_image = np.asarray("RGB")) actual_image = np.asarray("RGB")) actual_image, expected_image = crop_to_same( actual, actual_image, expected, expected_image) expected_image = np.array(expected_image).astype(float) actual_image = np.array(actual_image).astype(float) if expected_image.shape != actual_image.shape: raise ImageComparisonFailure( "Image sizes do not match expected size: {} " "actual size {}".format(expected_image.shape, actual_image.shape)) abs_diff_image = np.abs(expected_image - actual_image) # expand differences in luminance domain abs_diff_image *= 255 * 10 save_image_np = np.clip(abs_diff_image, 0, 255).astype(np.uint8) height, width, depth = save_image_np.shape # The PDF renderer doesn't produce an alpha channel, but the # matplotlib PNG writer requires one, so expand the array if depth == 3: with_alpha = np.empty((height, width, 4), dtype=np.uint8) with_alpha[:, :, 0:3] = save_image_np save_image_np = with_alpha # Hard-code the alpha channel to fully solid save_image_np[:, :, 3] = 255 Image.fromarray(save_image_np).save(output, format="png")