160 lines
5.8 KiB
Python
160 lines
5.8 KiB
Python
|
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
|
||
|
#
|
||
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||
|
# you may not use this file except in compliance with the License.
|
||
|
# You may obtain a copy of the License at
|
||
|
#
|
||
|
# http://www.apache.org/licenses/LICENSE-2.0
|
||
|
#
|
||
|
# Unless required by applicable law or agreed to in writing, software
|
||
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
||
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||
|
# See the License for the specific language governing permissions and
|
||
|
# limitations under the License.
|
||
|
# ==============================================================================
|
||
|
"""Package for histogram compression."""
|
||
|
|
||
|
|
||
|
import dataclasses
|
||
|
import numpy as np
|
||
|
|
||
|
from typing import Tuple
|
||
|
|
||
|
# Normal CDF for std_devs: (-Inf, -1.5, -1, -0.5, 0, 0.5, 1, 1.5, Inf)
|
||
|
# naturally gives bands around median of width 1 std dev, 2 std dev, 3 std dev,
|
||
|
# and then the long tail.
|
||
|
NORMAL_HISTOGRAM_BPS = (0, 668, 1587, 3085, 5000, 6915, 8413, 9332, 10000)
|
||
|
|
||
|
|
||
|
@dataclasses.dataclass(frozen=True)
|
||
|
class CompressedHistogramValue:
|
||
|
"""Represents a value in a compressed histogram.
|
||
|
|
||
|
Attributes:
|
||
|
basis_point: Compression point represented in basis point, 1/100th of a
|
||
|
percent.
|
||
|
value: Cumulative weight at the basis point.
|
||
|
"""
|
||
|
|
||
|
basis_point: float
|
||
|
value: float
|
||
|
|
||
|
def as_tuple(self) -> Tuple[float, float]:
|
||
|
"""Returns the basis point and the value as a tuple."""
|
||
|
return (self.basis_point, self.value)
|
||
|
|
||
|
|
||
|
# TODO(@jart): Unfork these methods.
|
||
|
def compress_histogram_proto(histo, bps=NORMAL_HISTOGRAM_BPS):
|
||
|
"""Creates fixed size histogram by adding compression to accumulated state.
|
||
|
|
||
|
This routine transforms a histogram at a particular step by interpolating its
|
||
|
variable number of buckets to represent their cumulative weight at a constant
|
||
|
number of compression points. This significantly reduces the size of the
|
||
|
histogram and makes it suitable for a two-dimensional area plot where the
|
||
|
output of this routine constitutes the ranges for a single x coordinate.
|
||
|
|
||
|
Args:
|
||
|
histo: A HistogramProto object.
|
||
|
bps: Compression points represented in basis points, 1/100ths of a percent.
|
||
|
Defaults to normal distribution.
|
||
|
|
||
|
Returns:
|
||
|
List of values for each basis point.
|
||
|
"""
|
||
|
# See also: Histogram::Percentile() in core/lib/histogram/histogram.cc
|
||
|
if not histo.num:
|
||
|
return [CompressedHistogramValue(b, 0.0).as_tuple() for b in bps]
|
||
|
bucket = np.array(histo.bucket)
|
||
|
bucket_limit = list(histo.bucket_limit)
|
||
|
weights = (bucket * bps[-1] / (bucket.sum() or 1.0)).cumsum()
|
||
|
values = []
|
||
|
j = 0
|
||
|
while j < len(bps):
|
||
|
i = np.searchsorted(weights, bps[j], side="right")
|
||
|
while i < len(weights):
|
||
|
cumsum = weights[i]
|
||
|
cumsum_prev = weights[i - 1] if i > 0 else 0.0
|
||
|
if cumsum == cumsum_prev: # prevent lerp divide by zero
|
||
|
i += 1
|
||
|
continue
|
||
|
if not i or not cumsum_prev:
|
||
|
lhs = histo.min
|
||
|
else:
|
||
|
lhs = max(bucket_limit[i - 1], histo.min)
|
||
|
rhs = min(bucket_limit[i], histo.max)
|
||
|
weight = _lerp(bps[j], cumsum_prev, cumsum, lhs, rhs)
|
||
|
values.append(CompressedHistogramValue(bps[j], weight).as_tuple())
|
||
|
j += 1
|
||
|
break
|
||
|
else:
|
||
|
break
|
||
|
while j < len(bps):
|
||
|
values.append(CompressedHistogramValue(bps[j], histo.max).as_tuple())
|
||
|
j += 1
|
||
|
return values
|
||
|
|
||
|
|
||
|
def compress_histogram(buckets, bps=NORMAL_HISTOGRAM_BPS):
|
||
|
"""Creates fixed size histogram by adding compression to accumulated state.
|
||
|
|
||
|
This routine transforms a histogram at a particular step by linearly
|
||
|
interpolating its variable number of buckets to represent their cumulative
|
||
|
weight at a constant number of compression points. This significantly reduces
|
||
|
the size of the histogram and makes it suitable for a two-dimensional area
|
||
|
plot where the output of this routine constitutes the ranges for a single x
|
||
|
coordinate.
|
||
|
|
||
|
Args:
|
||
|
buckets: A list of buckets, each of which is a 3-tuple of the form
|
||
|
`(min, max, count)`.
|
||
|
bps: Compression points represented in basis points, 1/100ths of a percent.
|
||
|
Defaults to normal distribution.
|
||
|
|
||
|
Returns:
|
||
|
List of values for each basis point.
|
||
|
"""
|
||
|
# See also: Histogram::Percentile() in core/lib/histogram/histogram.cc
|
||
|
buckets = np.array(buckets)
|
||
|
if not buckets.size:
|
||
|
return [CompressedHistogramValue(b, 0.0).as_tuple() for b in bps]
|
||
|
(minmin, maxmax) = (buckets[0][0], buckets[-1][1])
|
||
|
counts = buckets[:, 2]
|
||
|
right_edges = list(buckets[:, 1])
|
||
|
weights = (counts * bps[-1] / (counts.sum() or 1.0)).cumsum()
|
||
|
|
||
|
result = []
|
||
|
bp_index = 0
|
||
|
while bp_index < len(bps):
|
||
|
i = np.searchsorted(weights, bps[bp_index], side="right")
|
||
|
while i < len(weights):
|
||
|
cumsum = weights[i]
|
||
|
cumsum_prev = weights[i - 1] if i > 0 else 0.0
|
||
|
if cumsum == cumsum_prev: # prevent division-by-zero in `_lerp`
|
||
|
i += 1
|
||
|
continue
|
||
|
if not i or not cumsum_prev:
|
||
|
lhs = minmin
|
||
|
else:
|
||
|
lhs = max(right_edges[i - 1], minmin)
|
||
|
rhs = min(right_edges[i], maxmax)
|
||
|
weight = _lerp(bps[bp_index], cumsum_prev, cumsum, lhs, rhs)
|
||
|
result.append(
|
||
|
CompressedHistogramValue(bps[bp_index], weight).as_tuple()
|
||
|
)
|
||
|
bp_index += 1
|
||
|
break
|
||
|
else:
|
||
|
break
|
||
|
while bp_index < len(bps):
|
||
|
result.append(
|
||
|
CompressedHistogramValue(bps[bp_index], maxmax).as_tuple()
|
||
|
)
|
||
|
bp_index += 1
|
||
|
return result
|
||
|
|
||
|
|
||
|
def _lerp(x, x0, x1, y0, y1):
|
||
|
"""Affinely map from [x0, x1] onto [y0, y1]."""
|
||
|
return y0 + (x - x0) * float(y1 - y0) / (x1 - x0)
|