def test():
""" test """
indata = [0.1, 0.2, 0.3, 0.4, 0.4, 0.5, 0.1, 0.4]
stats = RollingStats()
for i in indata:
stats.append(i)
print stats.get()
print calculate_stats(indata)
def get_chi2(tasks, bad_users):
"""Perform chi^2 test on the stats from each worker"""
results = {'chi2': 0.0, 'dof': 0, 'p': 0.0, 'residual': {}}
# Aggregate results by worker
users = {}
for task in tasks:
task['worker_key'] = get_worker_key(task)
if 'worker_info' not in task:
continue
key = get_worker_key(task)
if key in bad_users:
continue
stats = task.get('stats', {})
wld = [float(stats.get('wins', 0)), float(stats.get('losses', 0)), float(stats.get('draws', 0))]
if wld == [0.0, 0.0, 0.0]:
continue
if key in users:
for idx in range(len(wld)):
users[key][idx] += wld[idx]
else:
users[key] = wld
if len(users) == 0:
return results
observed = numpy.array(users.values())
rows,columns = observed.shape
df = (rows - 1) * (columns - 1)
column_sums = numpy.sum(observed, axis=0)
row_sums = numpy.sum(observed, axis=1)
grand_total = numpy.sum(column_sums)
if grand_total == 0:
return results
expected = numpy.outer(row_sums, column_sums) / grand_total
diff = observed - expected
adj = numpy.outer((1 - row_sums / grand_total), (1 - column_sums / grand_total))
residual = diff / numpy.sqrt(expected * adj)
for idx in range(len(users)):
users[users.keys()[idx]] = numpy.max(numpy.abs(residual[idx]))
chi2 = numpy.sum(diff * diff / expected)
return {
'chi2': chi2,
'dof': df,
'p': 1 - scipy.stats.chi2.cdf(chi2, df),
'residual': users,
}
def calculate_residuals(run):
bad_users = set()
chi2 = get_chi2(run['tasks'], bad_users)
residuals = chi2['residual']
# Limit bad users to 1 for now
for _ in range(1):
worst_user = {}
for task in run['tasks']:
if task['worker_key'] in bad_users:
continue
task['residual'] = residuals.get(task['worker_key'], 0.0)
# Special case crashes or time losses
stats = task.get('stats', {})
crashes = stats.get('crashes', 0)
time_losses = stats.get('time_losses', 0)
if crashes + time_losses > 3:
task['residual'] = 8.0
if abs(task['residual']) < 2.0:
task['residual_color'] = '#44EB44'
elif abs(task['residual']) < 2.7:
task['residual_color'] = 'yellow'
else:
task['residual_color'] = '#FF6A6A'
if chi2['p'] < 0.01 or task['residual'] > 7.0:
if len(worst_user) == 0 or task['residual'] > worst_user['residual']:
worst_user['worker_key'] = task['worker_key']
worst_user['residual'] = task['residual']
if len(worst_user) == 0:
break
bad_users.add(worst_user['worker_key'])
residuals = get_chi2(run['tasks'], bad_users)['residual']
chi2['bad_users'] = bad_users
return chi2
def get_chi2(tasks, bad_users):
""" Perform chi^2 test on the stats from each worker """
results = {"chi2": 0.0, "dof": 0, "p": 0.0, "residual": {}}
# Aggregate results by worker
users = {}
for task in tasks:
task["worker_key"] = get_worker_key(task)
if "worker_info" not in task:
continue
key = get_worker_key(task)
if key in bad_users:
continue
stats = task.get("stats", {})
wld = [
float(stats.get("wins", 0)),
float(stats.get("losses", 0)),
float(stats.get("draws", 0)),
]
if wld == [0.0, 0.0, 0.0]:
continue
if key in users:
for idx in range(len(wld)):
users[key][idx] += wld[idx]
else:
users[key] = wld
if len(users) == 0:
return results
observed = numpy.array(list(users.values()))
rows, columns = observed.shape
# Results only from one worker: skip the test for workers homogeneity
if rows == 1:
return {"chi2": float("nan"), "dof": 0, "p": float("nan"), "residual": {}}
column_sums = numpy.sum(observed, axis=0)
columns_not_zero = sum(i > 0 for i in column_sums)
df = (rows - 1) * (columns - 1)
if columns_not_zero == 0:
return results
# Results only of one type: workers are identical wrt the test
elif columns_not_zero == 1:
results = {"chi2": 0.0, "dof": df, "p": 1.0, "residual": {}}
return results
# Results only of two types: workers are identical wrt the missing result type
# Change the data shape to avoid divide by zero
elif columns_not_zero == 2:
idx = numpy.argwhere(numpy.all(observed[..., :] == 0, axis=0))
observed = numpy.delete(observed, idx, axis=1)
column_sums = numpy.sum(observed, axis=0)
row_sums = numpy.sum(observed, axis=1)
grand_total = numpy.sum(column_sums)
expected = numpy.outer(row_sums, column_sums) / grand_total
raw_residual = observed - expected
std_error = numpy.sqrt(
expected
* numpy.outer((1 - row_sums / grand_total), (1 - column_sums / grand_total))
)
adj_residual = raw_residual / std_error
for idx in range(len(users)):
users[list(users.keys())[idx]] = numpy.max(numpy.abs(adj_residual[idx]))
chi2 = numpy.sum(raw_residual * raw_residual / expected)
return {
"chi2": chi2,
"dof": df,
"p": 1 - scipy.stats.chi2.cdf(chi2, df),
"residual": users,
}
def get_chi2(tasks, bad_users):
"""Perform chi^2 test on the stats from each worker"""
results = {"chi2": 0.0, "dof": 0, "p": 0.0, "residual": {}}
# Aggregate results by worker
users = {}
has_pentanomial = None
for task in tasks:
task["worker_key"] = get_worker_key(task)
if "worker_info" not in task:
continue
key = get_worker_key(task)
if key in bad_users:
continue
stats = task.get("stats", {})
if has_pentanomial is None:
has_pentanomial = "pentanomial" in stats
if not has_pentanomial:
wld = [
float(stats.get("wins", 0)),
float(stats.get("losses", 0)),
float(stats.get("draws", 0)),
]
else:
p = stats["pentanomial"]
# The ww and ll frequencies will typically be too small for
# the full pentanomial chi2 test to be valid. See e.g. the last page of
# https://www.open.ac.uk/socialsciences/spsstutorial/files/tutorials/chi-square.pdf.
# So we combine the ww and ll frequencies with the wd and ld frequencies.
wld = [float(p[4] + p[3]), float(p[0] + p[1]), float(p[2])]
if wld == [0.0, 0.0, 0.0]:
continue
if key in users:
for idx in range(len(wld)):
users[key][idx] += wld[idx]
else:
users[key] = wld
if len(users) == 0:
return results
observed = numpy.array(list(users.values()))
rows, columns = observed.shape
# Results only from one worker: skip the test for workers homogeneity
if rows == 1:
return {
"chi2": float("nan"),
"dof": 0,
"p": float("nan"),
"residual": {}
}
column_sums = numpy.sum(observed, axis=0)
columns_not_zero = numpy.count_nonzero(column_sums)
df = (rows - 1) * (columns - 1)
if columns_not_zero == 0:
return results
# Results only of one type: workers are homogeneous
elif columns_not_zero == 1:
return {"chi2": 0.0, "dof": df, "p": 1.0, "residual": {}}
# Results only of two types: drop the column of zeros to avoid divide by zero
elif columns_not_zero == 2:
observed = observed[:, ~numpy.all(observed == 0, axis=0)]
column_sums = numpy.sum(observed, axis=0)
row_sums = numpy.sum(observed, axis=1)
grand_total = numpy.sum(column_sums)
expected = numpy.outer(row_sums, column_sums) / grand_total
raw_residual = observed - expected
std_error = numpy.sqrt(expected * numpy.outer(
(1 - row_sums / grand_total), (1 - column_sums / grand_total)))
adj_residual = raw_residual / std_error
for idx in range(len(users)):
users[list(users.keys())[idx]] = numpy.max(numpy.abs(
adj_residual[idx]))
chi2 = numpy.sum(raw_residual * raw_residual / expected)
return {
"chi2": chi2,
"dof": df,
"p": 1 - scipy.stats.chi2.cdf(chi2, df),
"residual": users,
}
