Compare reference data

Installation

• Edit .git/configure to set up upstream remote:

[remote "upstream"]
    url = <URL of GitHub repository or Azure mirror>
    fetch = +refs/heads/*:refs/remotes/upstream/*
    fetch = +refs/pull/*/head:refs/remotes/upstream/pr/*

• Initialize LFS hooks and fetch references from upstream:

  $ git lfs install
  $ git fetch upstream

• This notebook uses the bokeh plotting backend by default. Install it alongside your tardis environment by doing:

  $ conda activate tardis
  $ conda install bokeh -c conda-forge --no-update-deps

Usage

• Run this notebook inside the tardis-refdata/notebooks directory.

• ReferenceComparer object loads two versions of the reference data by passing at least one Git label (e.g. hash, tag, branch name). If either is set to None it will just use the current data in the directory. For example:

  comparer = ReferenceComparer(ref1_hash=None, ref2_hash='upstream/master')
  

  compares unit_test_data.h5 from your local repository against the HEAD of the upstream remote.

  Please set the labels you want to compare now:

REF1_HASH_DEFAULT = None
REF2_HASH_DEFAULT = 'upstream/master'

• This feature is specially useful for CI pipelines: ref1_hash and ref2_hash parameters can be passed as environment variables before running the notebook, overwriting the defaults defined in the above cell.

  export REF2_HASH='upstream/master'

• If you want to switch to the matplotlib backend pass the mpl_backend=True option to the compare_output_nu and compare_spectrum function.

• Use .teardown() method to delete temporary files.

---

import os
import shutil
import tempfile
import subprocess
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from bokeh.plotting import figure, show, ColumnDataSource
from bokeh.layouts import gridplot
from bokeh.models.tools import HoverTool
from bokeh.io import output_notebook

output_notebook()

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
<ipython-input-1-9f363699fb7c> in <module>
      6 import pandas as pd
      7 import matplotlib.pyplot as plt
----> 8 from bokeh.plotting import figure, show, ColumnDataSource
      9 from bokeh.layouts import gridplot
     10 from bokeh.models.tools import HoverTool

ModuleNotFoundError: No module named 'bokeh'

Make sure you are in the root of tardis-refdata after running the following cell.

cd ..

/home/vsts/work/1/s/tardis-refdata

Define classes and functions

def highlight_missing(val):
    if val == True:
        return 'background-color: #BCF5A9'
    else:
        return 'background-color: #F5A9A9'
    
def highlight_relative_difference(val):
    ret = 'background-color: #BCF5A9'
    if val is None:
        ret = 'background-color: #BCF5A9'
    elif val > 1e-2:
        ret = 'background-color: #F2F5A9'
    elif val > 1e-1:
        ret = 'background-color: #F5D0A9'
    elif val > 1:
        ret = 'background-color: #F5A9A9'
    return ret

class ReferenceComparer(object):

    def __init__(self, ref1_hash=None, ref2_hash=None, compare_path='unit_test_data.h5'):
        assert not ((ref1_hash is None) and (ref2_hash is None)), "One hash can not be None"
        self.ref1_hash = ref1_hash
        self.ref2_hash = ref2_hash
        self.compare_path = compare_path
        self.tmp_dir = None
        self.setup()
    
    def setup(self):
        self.tmp_dir = tempfile.mkdtemp()
        print('Created temporary directory at {0}. Delete after use with .teardown'.format(self.tmp_dir))
        for ref_id, ref_hash in enumerate([self.ref1_hash, self.ref2_hash]):
            ref_id += 1
            if ref_hash is not None:
                self._copy_data_from_hash(ref_hash, 'ref{0}_'.format(ref_id))
            else:
                subprocess.Popen('cp {0} {1}'.format(self.compare_path, 
                                                     os.path.join(self.tmp_dir, 
                                                                  'ref{0}_{1}'.format(ref_id, self.compare_path))), 
                                                     shell=True)
            setattr(self, 'ref{0}_fname'.format(ref_id), 
                    os.path.join(self.tmp_dir, 'ref{0}_{1}'.format(ref_id, self.compare_path)))

    def teardown(self):
        shutil.rmtree(self.tmp_dir)

    def _copy_data_from_hash(self, ref_hash, prefix):
        git_cmd = ['git']
        git_cmd.append('--work-tree={0}'.format(self.tmp_dir))
        git_cmd += ['checkout', ref_hash, self.compare_path]
        p = subprocess.Popen(git_cmd)
        p.wait()
        shutil.move(os.path.join(self.tmp_dir, self.compare_path), 
                    os.path.join(self.tmp_dir, prefix + self.compare_path))

    def generate_test_table(self):
        rd1_hdfs = pd.HDFStore(self.ref1_fname, mode='r')
        rd2_hdfs = pd.HDFStore(self.ref2_fname, mode='r')
        rd1_keys = rd1_hdfs.keys()
        rd2_keys = rd2_hdfs.keys()
        rd1_hdfs.close()
        rd2_hdfs.close()
        rd1_df = pd.DataFrame(index=rd1_keys, columns=['exists'])
        rd2_df = pd.DataFrame(index=rd2_keys, columns=['exists'])
        rd1_df['exists'] = True
        rd2_df['exists'] = True
        joined_df = rd1_df.join(rd2_df, how='outer', lsuffix='_1', rsuffix='_2')
        joined_df = joined_df.fillna(False)
        return joined_df
    
    def compare_refdata(self, test_table):
        test_table['match'] = None
        test_table['abs_diff_mean'] = None
        test_table['abs_diff_max'] = None
        test_table['rel_diff_mean'] = None
        test_table['rel_diff_max'] = None
        for row_id, row in test_table.iterrows():
            
            if row[['exists_1', 'exists_2']].all():
                ref1_df = pd.read_hdf(self.ref1_fname, row_id)
                ref2_df = pd.read_hdf(self.ref2_fname, row_id)
                
                if isinstance(ref1_df, pd.Series):
                    try:
                        pd.testing.assert_series_equal(ref1_df, ref2_df)
                    except AssertionError:
                        test_table.loc[row_id, 'match'] = False
                        abs_diff = np.fabs(ref1_df - ref2_df)
                        rel_diff = (abs_diff / np.fabs(ref1_df))[ref1_df != 0]
                        test_table.loc[row_id, 'abs_diff_mean'] = abs_diff.mean()
                        test_table.loc[row_id, 'abs_diff_max'] = abs_diff.max()
                        test_table.loc[row_id, 'rel_diff_mean'] = rel_diff.mean()
                        test_table.loc[row_id, 'rel_diff_max'] = rel_diff.max()
                    else:
                        test_table.loc[row_id, 'match'] = True

                elif isinstance(ref1_df, pd.DataFrame):
                    try:
                        pd.testing.assert_frame_equal(ref1_df, ref2_df)
                    except AssertionError:
                        test_table.loc[row_id, 'match'] = False
                        abs_diff = np.fabs(ref1_df - ref2_df)
                        rel_diff = (abs_diff / np.fabs(ref1_df))[ref1_df != 0]
                        test_table.loc[row_id, 'abs_diff_mean'] = abs_diff.mean(skipna=True).mean()
                        test_table.loc[row_id, 'abs_diff_max'] = abs_diff.max(skipna=True).max()
                        test_table.loc[row_id, 'rel_diff_mean'] = rel_diff.mean(skipna=True).mean()
                        test_table.loc[row_id, 'rel_diff_max'] = rel_diff.max(skipna=True).max()
                    else:
                        test_table.loc[row_id, 'match'] = True

                else:
                    raise ValueError('Needs to be a Series or DataFrame but is' + str(type(ref1_df)))
        return test_table
                

Load data

Check if REF1_HASH and REF2_HASH are environment variables. If not defined (or empty), use the defaults defined above.

try:
    REF1_HASH = os.environ['REF1_HASH']

    if not REF1_HASH:
        raise ValueError

except (KeyError, ValueError):
    REF1_HASH = REF1_HASH_DEFAULT

try:
    REF2_HASH = os.environ['REF2_HASH']

    if not REF2_HASH:
        raise ValueError

except (KeyError, ValueError):
    REF2_HASH = REF2_HASH_DEFAULT

REF1_HASH, REF2_HASH

('b6e5238', 'ae743ac')

comparer = ReferenceComparer(ref1_hash=REF1_HASH, ref2_hash=REF2_HASH)

Created temporary directory at /tmp/tmp27baxs0n. Delete after use with .teardown

---------------------------------------------------------------------------
FileNotFoundError                         Traceback (most recent call last)
/usr/share/miniconda/envs/tardis/lib/python3.8/shutil.py in move(src, dst, copy_function)
    790     try:
--> 791         os.rename(src, real_dst)
    792     except OSError:

FileNotFoundError: [Errno 2] No such file or directory: '/tmp/tmp27baxs0n/unit_test_data.h5' -> '/tmp/tmp27baxs0n/ref1_unit_test_data.h5'

During handling of the above exception, another exception occurred:

FileNotFoundError                         Traceback (most recent call last)
<ipython-input-1-7430fed7be7e> in <module>
----> 1 comparer = ReferenceComparer(ref1_hash=REF1_HASH, ref2_hash=REF2_HASH)

<ipython-input-1-aa05e4877f43> in __init__(self, ref1_hash, ref2_hash, compare_path)
      7         self.compare_path = compare_path
      8         self.tmp_dir = None
----> 9         self.setup()
     10 
     11     def setup(self):

<ipython-input-1-aa05e4877f43> in setup(self)
     15             ref_id += 1
     16             if ref_hash is not None:
---> 17                 self._copy_data_from_hash(ref_hash, 'ref{0}_'.format(ref_id))
     18             else:
     19                 subprocess.Popen('cp {0} {1}'.format(self.compare_path, 

<ipython-input-1-aa05e4877f43> in _copy_data_from_hash(self, ref_hash, prefix)
     33         p = subprocess.Popen(git_cmd)
     34         p.wait()
---> 35         shutil.move(os.path.join(self.tmp_dir, self.compare_path), 
     36                     os.path.join(self.tmp_dir, prefix + self.compare_path))
     37 

/usr/share/miniconda/envs/tardis/lib/python3.8/shutil.py in move(src, dst, copy_function)
    803             rmtree(src)
    804         else:
--> 805             copy_function(src, real_dst)
    806             os.unlink(src)
    807     return real_dst

/usr/share/miniconda/envs/tardis/lib/python3.8/shutil.py in copy2(src, dst, follow_symlinks)
    433     if os.path.isdir(dst):
    434         dst = os.path.join(dst, os.path.basename(src))
--> 435     copyfile(src, dst, follow_symlinks=follow_symlinks)
    436     copystat(src, dst, follow_symlinks=follow_symlinks)
    437     return dst

/usr/share/miniconda/envs/tardis/lib/python3.8/shutil.py in copyfile(src, dst, follow_symlinks)
    262         os.symlink(os.readlink(src), dst)
    263     else:
--> 264         with open(src, 'rb') as fsrc, open(dst, 'wb') as fdst:
    265             # macOS
    266             if _HAS_FCOPYFILE:

FileNotFoundError: [Errno 2] No such file or directory: '/tmp/tmp27baxs0n/unit_test_data.h5'

Test table

tt = comparer.generate_test_table()

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-8117ee489402> in <module>
----> 1 tt = comparer.generate_test_table()

NameError: name 'comparer' is not defined

tt = comparer.compare_refdata(tt)

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-197587a6f814> in <module>
----> 1 tt = comparer.compare_refdata(tt)

NameError: name 'comparer' is not defined

tt[["exists_1", "exists_2", 'rel_diff_mean', 'rel_diff_max', 'match']].style.applymap(
    highlight_missing, subset=['exists_1', 'exists_2', 'match']).applymap(
    highlight_relative_difference, subset=['rel_diff_mean', 'rel_diff_max'])

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-49edc6aa32a1> in <module>
----> 1 tt[["exists_1", "exists_2", 'rel_diff_mean', 'rel_diff_max', 'match']].style.applymap(
      2     highlight_missing, subset=['exists_1', 'exists_2', 'match']).applymap(
      3     highlight_relative_difference, subset=['rel_diff_mean', 'rel_diff_max'])

NameError: name 'tt' is not defined

Detailed inspection of the reference data

If parts of the reference data show differences between revisions, you should invest some time examining these differences in detail. Often, visualizing the relevant data blocks already helps.

You can use the following plotting routines as a blueprint and adjust and extend them to your needs.

def compare_output_nu(df1, df2, mpl_backend=False):
    nu_min = np.min([df1.min(), df2.min()])
    nu_max = np.max([df1.max(), df2.max()])
    
    if mpl_backend:
        plt.figure(figsize=(14, 6))
        plt.subplot(121)
        plt.plot(df1, df2, ',')
        plt.xlabel("output_nu, ref 1")
        plt.ylabel("output_nu, ref 2")
        plt.subplot(122)
        plt.hist(df1, bins=np.linspace(nu_min, nu_max, 100), histtype="step", label="ref 1")
        plt.hist(df2, bins=np.linspace(nu_min, nu_max, 100), histtype="step", label="ref 2")
        plt.xlabel("output_nu")
        plt.legend(frameon=False)
        
        return

    
    TOOLTIPS = [("(x,y)", "(@x, @y)")]
    hover = HoverTool(tooltips=TOOLTIPS)
    
    p = figure()
    output_nu = ColumnDataSource(pd.DataFrame.from_records({'x': df1.values, 
                                                            'y': df2.values}))
    p.circle('x', 'y', size=1, source=output_nu)
    p.xaxis.axis_label = "output_nu, ref 1"
    p.yaxis.axis_label = "output_nu, ref 2"
    p.xaxis.formatter.precision = 1
    p.yaxis.formatter.precision = 1
    p.add_tools(hover)

    # Step lines are hacky way to make histograms with Bokeh
    arr_hist1, edges1 = np.histogram(df1.values, 
                                     bins = 100, 
                                     range = [nu_min, nu_max])
    arr_hist2, edges2 = np.histogram(df1.values, 
                                     bins = 100, 
                                     range = [nu_min, nu_max])
    
    hist1 = ColumnDataSource(pd.DataFrame.from_records({'x': np.linspace(nu_min, nu_max, 100),
                                                        'y': arr_hist1}))
    hist2 = ColumnDataSource(pd.DataFrame.from_records({'x': np.linspace(nu_min, nu_max, 100),
                                                        'y': arr_hist2}))
    q = figure()
    q.step('x', 'y', source=hist1, legend_label='ref 1')
    q.step('x', 'y', source=hist2, legend_label='ref 2', color='#ff7f0e')
    q.xaxis.axis_label = "output_nu"
    q.xaxis.formatter.precision = 1
    q.legend.click_policy="hide"
    
    # Currently HoverTool does not work for step line glyph. See: https://github.com/bokeh/bokeh/issues/7419
    q.add_tools(hover)
    
    plot = gridplot([p, q], ncols=2, plot_width=420, plot_height=360)
    
    show(plot)

def compare_spectrum(ref1_nu, ref1_L, ref2_nu, ref2_L, mpl_backend=False):
    
    if mpl_backend:
        plt.figure(figsize=(14, 6))
        plt.subplot(121)
        plt.plot(ref1_nu, ref1_L, label="ref 1")
        plt.plot(ref2_nu, ref2_L, label="ref 2")
        plt.xlabel("nu")
        plt.ylabel("L")
        plt.legend(frameon=False)
        plt.subplot(122)
        plt.plot(ref1_nu, ref1_L / ref2_L)
        plt.xlabel("nu")
        plt.ylabel("L ref 1 / L ref 2")
        
        return
    
    
    TOOLTIPS = [("(x,y)", "(@x, @y)")]
    hover = HoverTool(tooltips=TOOLTIPS)
    
    p = figure()
    spectrum1 = ColumnDataSource(pd.DataFrame.from_records({'x': ref1_nu.values, 
                                                            'y': ref1_L}))
    spectrum2 = ColumnDataSource(pd.DataFrame.from_records({'x': ref2_nu.values, 
                                                            'y': ref2_L}))
    p.line('x', 'y', source=spectrum1, legend_label='ref 1')
    p.line('x', 'y', source=spectrum2, legend_label='ref 2', color='#ff7f0e')
    p.xaxis.axis_label = "L"
    p.yaxis.axis_label = "nu"
    p.xaxis.formatter.precision = 1
    p.yaxis.formatter.precision = 1
    p.legend.click_policy="hide"
    p.add_tools(hover)
    
    q = figure()
    lum_ratio = ColumnDataSource(pd.DataFrame.from_records({'x': ref1_nu.values, 
                                                            'y': ref1_L.values/ref2_L.values}))
    q.circle('x', 'y', size=1, source=lum_ratio)
    q.xaxis.axis_label = "nu"
    q.yaxis.axis_label = "L ref 1 / L ref 2"
    q.xaxis.formatter.precision = 1
    q.yaxis.formatter.precision = 1
    q.add_tools(hover)
    
    
    plot = gridplot([p, q], ncols=2, plot_width=420, plot_height=360)
    
    show(plot)

Get the data and find all the entries for which differences exist:

tmp1 = pd.HDFStore(comparer.ref1_fname, "r")
tmp2 = pd.HDFStore(comparer.ref2_fname, "r")

diff_entries = tt.loc[(tt["match"] == False) & (tt["exists_1"] == True) & (tt["exists_2"] == True)].index

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-2b40df7d3038> in <module>
----> 1 tmp1 = pd.HDFStore(comparer.ref1_fname, "r")
      2 tmp2 = pd.HDFStore(comparer.ref2_fname, "r")
      3 
      4 diff_entries = tt.loc[(tt["match"] == False) & (tt["exists_1"] == True) & (tt["exists_2"] == True)].index

NameError: name 'comparer' is not defined

Results

compare_output_nu(tmp1['/test_simulation/output_nu'], tmp2['/test_simulation/output_nu'])

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-54d5fb0ce665> in <module>
----> 1 compare_output_nu(tmp1['/test_simulation/output_nu'], tmp2['/test_simulation/output_nu'])

NameError: name 'tmp1' is not defined

compare_spectrum(tmp1['/test_runner_simple/spectrum/_frequency'][:-1], 
                 tmp1['/test_runner_simple/spectrum/luminosity'],
                 tmp2['/test_runner_simple/spectrum/_frequency'][:-1], 
                 tmp2['/test_runner_simple/spectrum/luminosity'])

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-0257c581a562> in <module>
----> 1 compare_spectrum(tmp1['/test_runner_simple/spectrum/_frequency'][:-1], 
      2                  tmp1['/test_runner_simple/spectrum/luminosity'],
      3                  tmp2['/test_runner_simple/spectrum/_frequency'][:-1],
      4                  tmp2['/test_runner_simple/spectrum/luminosity'])

NameError: name 'tmp1' is not defined

comparer.teardown()

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-2244ed23302f> in <module>
----> 1 comparer.teardown()

NameError: name 'comparer' is not defined