Last active
February 24, 2021 10:23
-
-
Save Roffild/982ba8ffb1f7ee8a5f4f0183cbbf1cc0 to your computer and use it in GitHub Desktop.
Metrics from XGBoost for PyTorch
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
# 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. | |
# | |
# https://github.com/Roffild/RoffildLibrary | |
# ============================================================================== | |
import unittest | |
import torch | |
import math | |
# from roffild.autopytorch import Metrics | |
class Metrics: | |
@staticmethod | |
def rmse(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
result = (output - input).pow_(2) | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum).sqrt_() | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]).sqrt_() | |
return ("rmse", result) | |
@staticmethod | |
def rmsle(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
result = (output.log1p() - input.log1p()).pow_(2) | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum).sqrt_() | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]).sqrt_() | |
return ("rmsle", result) | |
@staticmethod | |
def mae(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
result = (output - input).abs_() | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum) | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]) | |
return ("mae", result) | |
@staticmethod | |
def mape(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
result = (output - input).div_(output).abs_() | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum) | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]) | |
return ("mape", result) | |
@staticmethod | |
def mphe(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
result = (output - input).pow_(2).add_(1.0).sqrt_().sub_(1.0) | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum) | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]) | |
return ("mphe", result) | |
@staticmethod | |
def error(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
result = torch.where(input > 0.5, 1.0 - output, output) | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum) | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]) | |
return ("error", result) | |
@staticmethod | |
def logloss(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
result = input.clone() | |
ilog = input.log() | |
i1m = 1.0 - input | |
i1mlog = i1m.log() | |
oneg = output.neg() | |
o1m = 1.0 - output | |
eps = torch.tensor(1e-16, dtype=torch.float32, device=result.device) | |
epslog = eps.log() | |
eps1mlog = (1.0 - eps).log_() ### BUG == nan == 0 !!! | |
con1 = input < eps | |
con2 = i1m < eps | |
con2 = con2.bitwise_xor(con1).bitwise_and_(con2) | |
con3 = con2.bitwise_or(con1).bitwise_not_() | |
if bool(con1.any()): | |
result[con1] = oneg[con1] * epslog - o1m[con1] * eps1mlog | |
if bool(con2.any()): | |
result[con2] = oneg[con2] * eps1mlog - o1m[con2] * epslog | |
if bool(con3.any()): | |
result[con3] = oneg[con3] * ilog[con3] - o1m[con3] * i1mlog[con3] | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum) | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]) | |
return ("logloss", result) | |
@staticmethod | |
def poisson_nloglik(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
eps = torch.tensor(1e-16, dtype=torch.float32, device=input.device) | |
result = torch.where(input < eps, eps, input) | |
result = (output + 1.0).lgamma_().add_(result).sub_(result.log_().mul_(output)) | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum) | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]) | |
return ("poisson-nloglik", result) | |
@staticmethod | |
def gamma_deviance(input: torch.Tensor, output: torch.Tensor, weights: torch.Tensor = None) -> (str, torch.Tensor): | |
"""Multiclass classification error.""" | |
with torch.no_grad(): | |
eps = torch.tensor(1e-16, dtype=torch.float32, device=input.device) | |
result = input.div(output.add(eps)).log_().add_(output.div(input.add(eps))).sub_(1.0).mul_(2.0) | |
if weights is not None: | |
result.mul_(weights) | |
weights_sum = weights.sum(dim=1, keepdim=True) | |
weights_sum[weights_sum == 0.0] = 1.0 | |
result = result.sum(dim=1, keepdim=True).div_(weights_sum) | |
else: | |
result = result.sum(dim=1, keepdim=True).div_(output.shape[1]) | |
return ("gamma-deviance", result) | |
class AutoPytorchMetricsTest(unittest.TestCase): | |
def test_rmse(self): | |
metric = Metrics.rmse | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[ 0.1, 0.9, 0.1, 0.9 ], | |
[ 0, 0, 1, 1 ]), | |
0.6403, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ -1, 1, 9, -9]), | |
2.8284, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ 1, 2, 9, 8]), | |
0.6708, 0.001) | |
def test_rmsle(self): | |
metric = Metrics.rmsle | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.2, 0.4, 0.8, 1.6], | |
[1.0, 1.0, 1.0, 1.0, 1.0]), | |
0.40632, 1e-4) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.2, 0.4, 0.8, 1.6], | |
[1.0, 1.0, 1.0, 1.0, 1.0], | |
[ 0, -1, 1, -9, 9]), | |
0.6212, 1e-4) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.2, 0.4, 0.8, 1.6], | |
[1.0, 1.0, 1.0, 1.0, 1.0], | |
[ 0, 1, 2, 9, 8]), | |
0.2415, 1e-4) | |
def test_mae(self): | |
metric = Metrics.mae | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1]), | |
0.5, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ -1, 1, 9, -9]), | |
8.0, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ 1, 2, 9, 8]), | |
0.54, 0.001) | |
def test_mape(self): | |
metric = Metrics.mape | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [150, 300], [100, 200]), 0.5, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[50, 400, 500, 4000], | |
[100, 200, 500, 1000]), | |
1.125, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[50, 400, 500, 4000], | |
[100, 200, 500, 1000], | |
[ -1, 1, 9, -9]), | |
-26.5, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[50, 400, 500, 4000], | |
[100, 200, 500, 1000], | |
[ 1, 2, 9, 8]), | |
1.3250, 0.001) | |
def test_mphe(self): | |
metric = Metrics.mphe | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1]), | |
0.17517, 1e-4) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ -1, 1, 9, -9]), | |
3.4037, 1e-4) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ 1, 2, 9, 8]), | |
0.1922, 1e-4) | |
def test_logloss(self): | |
metric = Metrics.logloss | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.5, 1e-17, 1.0+1e-17, 0.9], | |
[ 0, 0, 1, 1]), | |
0.1996, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1]), | |
1.2039, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ -1, 1, 9, -9]), | |
21.9722, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ 1, 2, 9, 8]), | |
1.3138, 0.001) | |
def test_error(self): | |
metric = Metrics.error | |
# For error@0.5 | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1]), | |
0.5, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ -1, 1, 9, -9]), | |
10.0, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ 1, 2, 9, 8]), | |
0.55, 0.001) | |
def test_poisson_nloglik(self): | |
metric = Metrics.poisson_nloglik | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0.5, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.5, 1e-17, 1.0+1e-17, 0.9], | |
[ 0, 0, 1, 1]), | |
0.6263, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1]), | |
1.1019, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ -1, 1, 9, -9]), | |
13.3750, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ 1, 2, 9, 8]), | |
1.5783, 0.001) | |
def test_gamma_deviance(self): | |
metric = Metrics.gamma_deviance | |
self.EXPECT_NEAR(self.GetMetricEval(metric, [0, 1], [0, 1]), 0.5, 1e-10) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1]), | |
1.1020, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ -1, 1, 9, -9]), | |
4.4079, 0.001) | |
self.EXPECT_NEAR(self.GetMetricEval(metric, | |
[0.1, 0.9, 0.1, 0.9], | |
[ 0, 0, 1, 1], | |
[ 1, 2, 9, 8]), | |
0.2204, 0.001) | |
def Stest_merror(self): | |
metric = Metrics.merror | |
def Stest_mlogloss(self): | |
metric = Metrics.mlogloss | |
def EXPECT_NEAR(self, val1, val2, abs_error): | |
if abs(val1 - val2) >= abs_error: print(f"{val1} != {val2} >= {abs_error} = ", abs(val1-val2)) | |
return 0 | |
self.assertTrue(abs(val1 - val2) <= abs_error, f"{val1} != {val2}") | |
def GetMetricEval(self, metric, preds, labels, weights=None): # , groups = None): | |
input = torch.tensor(list(preds) * 5, dtype=torch.float32).reshape((5, -1)) | |
output = torch.tensor(list(labels) * 5, dtype=torch.float32).reshape((5, -1)) | |
input_old = input.clone() | |
output_old = output.clone() | |
if weights is not None: | |
weights = torch.tensor(list(weights) * 5, dtype=torch.float32).reshape((5, -1)) | |
weights_old = weights.clone() | |
name, result = metric(input, output, weights) | |
self.assertEqual(name.replace("-", "_"), metric.__name__) | |
self.assertTrue(bool(input.eq(input_old).all())) | |
self.assertTrue(bool(output.eq(output_old).all())) | |
if weights is not None: | |
self.assertTrue(bool(weights.eq(weights_old).all())) | |
print(name, " = ", result.mean()) | |
return float(result.mean()) | |
if __name__ == "__main__": | |
unittest.main() |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment