Add pypy support (#80)

* Add pypy support

* Switch to builtin statistics library. Drop scipy requirement due to negligible difference in benchmark performance.

* Using custom ranking function

* Drop 3.7 support due to nearing EOL

* Skip doctest for *.rst files

* Adjust relative tolerance of doctest number comparison

* Change doctest glob to *.py

* Add changelog fragments

Author: vivekjoshy

.github/workflows/main.yml

@@ -8,7 +8,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [Ubuntu, MacOS, Windows]
-        python-version: ["3.7", "3.8", "3.9", "3.10", "3.11"]
+        python-version: ["3.8", "3.9", "3.10", "3.11", "pypy3.8", "pypy3.9"]
         experimental: [false]
         include:
           - os: Ubuntu
@@ -42,7 +42,7 @@ jobs:
               python -m build
               twine check dist/*
               black --check .
-              pytest . --cov --cov-report=term-missing -vv
+              pytest . --cov --cov-report=term-missing -vv --doctest-glob="*.py"
               coverage report
               coverage html
               coverage xml --ignore-errors

README.md

@@ -170,7 +170,7 @@ Sometimes you want to know what the likelihood is someone will place at a partic
 >>> team_1, team_2, team_3 = [a1, a2, a3], [b1, b2, b3], [c1, c2, c3]
 >>> draw_probability = predict_draw(teams=[team_1, team_2, team_3])
 >>> draw_probability
-0.3295385074666581
+0.329538507466658
 >>> rank_probability = predict_rank(teams=[team_1, team_2, team_3])
 >>> rank_probability
 [(1, 0.4450361350569973), (2, 0.19655022513040032), (3, 0.028875132345944337)]

benchmark/benchmark.py

@@ -5,7 +5,6 @@
 from prompt_toolkit.completion import WordCompleter
 from prompt_toolkit.validation import ValidationError, Validator
 
-from benchmark.processors import Competition
 from openskill.models import (
     BradleyTerryFull,
     BradleyTerryPart,

changes/80.breaking.rst

@@ -0,0 +1 @@
+Drop support for Python 3.7

changes/80.doc.rst

@@ -0,0 +1 @@
+Improve SEO of Documentation

changes/80.feature.rst

@@ -0,0 +1 @@
+Add support for PyPy 3.8 and 3.9

docs/manual.rst

@@ -137,10 +137,10 @@ Sometimes you want to know what the likelihood is someone will place at a partic
    >>> team_1, team_2, team_3 = [a1, a2, a3], [b1, b2, b3], [c1, c2, c3]
    >>> draw_probability = predict_draw(teams=[team_1, team_2, team_3])
    >>> draw_probability
-   0.3295385074666581
+   0.329538507466658
    >>> rank_probability = predict_rank(teams=[team_1, team_2, team_3])
    >>> rank_probability
-   [(1, 0.4450361350569973), (2, 0.19655022513040032), (3, 0.028875132345944337)]
+   [(1, 0.4450361350569973), (2, 0.19655022513040035), (3, 0.02887513234594437)]
    >>> sum([y for x, y in rank_probability]) + draw_probability
    1.0
 

openskill/rate.py

@@ -4,14 +4,12 @@
 from functools import reduce
 from typing import List, Optional, Tuple, Union
 
-from scipy.stats import rankdata
-
 from openskill.constants import Constants, beta
 from openskill.constants import mu as default_mu
 from openskill.constants import sigma as default_sigma
 from openskill.models.plackett_luce import PlackettLuce
 from openskill.statistics import phi_major, phi_major_inverse
-from openskill.util import rankings, unwind
+from openskill.util import rank_minimum, rankings, unwind
 
 
 class Rating:
@@ -461,7 +459,7 @@ def predict_rank(
     ]
 
     ranked_probability = [abs(_) for _ in win_probability]
-    ranks = list(rankdata(ranked_probability, method="min"))
+    ranks = list(rank_minimum(ranked_probability))
     max_ordinal = max(ranks)
     ranks = [abs(_ - max_ordinal) + 1 for _ in ranks]
     predictions = list(zip(ranks, ranked_probability))

openskill/statistics.py

@@ -1,16 +1,15 @@
 import sys
+from statistics import NormalDist
 
-import scipy.stats
-
-normal = scipy.stats.norm(0, 1)
+normal = NormalDist()
 
 
 def phi_major(x):
     return normal.cdf(x)
 
 
 def phi_major_inverse(x):
-    return normal.ppf(x)
+    return normal.inv_cdf(x)
 
 
 def phi_minor(x):

openskill/util.py

@@ -119,3 +119,25 @@ def sorter(teams):
         return [x for x, _ in sorted_list], [x for _, x in sorted_list]
 
     return sorter(teams) if isinstance(teams, list) else sorter
+
+
+def rank_simple(vector):
+    return sorted(range(len(vector)), key=vector.__getitem__)
+
+
+def rank_minimum(a):
+    n = len(a)
+    i_vec = rank_simple(a)
+    s_vec = [a[rank] for rank in i_vec]
+    sum_ranks = 0
+    duplicate_count = 0
+    new_array = [0] * n
+    for i in range(n):
+        sum_ranks += i
+        duplicate_count += 1
+        if i == n - 1 or s_vec[i] != s_vec[i + 1]:
+            for j in range(i - duplicate_count + 1, i + 1):
+                new_array[i_vec[j]] = i + 1 - duplicate_count + 1
+            sum_ranks = 0
+            duplicate_count = 0
+    return new_array