File size: 8,238 Bytes
9c6594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 |
# Copyright The Lightning team.
#
# 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.
from typing import Optional, Union
import torch
from torch import Tensor
from typing_extensions import Literal
from torchmetrics.utilities import rank_zero_warn
def _safe_matmul(x: Tensor, y: Tensor) -> Tensor:
"""Safe calculation of matrix multiplication.
If input is float16, will cast to float32 for computation and back again.
"""
if x.dtype == torch.float16 or y.dtype == torch.float16:
return (x.float() @ y.T.float()).half()
return x @ y.T
def _safe_xlogy(x: Tensor, y: Tensor) -> Tensor:
"""Compute x * log(y). Returns 0 if x=0.
Example:
>>> import torch
>>> x = torch.zeros(1)
>>> _safe_xlogy(x, 1/x)
tensor([0.])
"""
res = x * torch.log(y)
res[x == 0] = 0.0
return res
def _safe_divide(
num: Tensor,
denom: Tensor,
zero_division: Union[float, Literal["warn", "nan"]] = 0.0,
) -> Tensor:
"""Safe division, by preventing division by zero.
Function will cast to float if input is not already to secure backwards compatibility.
Args:
num: numerator tensor
denom: denominator tensor, which may contain zeros
zero_division: value to replace elements divided by zero
Example:
>>> import torch
>>> num = torch.tensor([1.0, 2.0, 3.0])
>>> denom = torch.tensor([0.0, 1.0, 2.0])
>>> _safe_divide(num, denom)
tensor([0.0000, 2.0000, 1.5000])
"""
num = num if num.is_floating_point() else num.float()
denom = denom if denom.is_floating_point() else denom.float()
if isinstance(zero_division, (float, int)) or zero_division == "warn":
if zero_division == "warn" and torch.any(denom == 0):
rank_zero_warn("Detected zero division in _safe_divide. Setting 0/0 to 0.0")
zero_division = 0.0 if zero_division == "warn" else zero_division
zero_division_tensor = torch.tensor(zero_division, dtype=num.dtype).to(num.device, non_blocking=True)
return torch.where(denom != 0, num / denom, zero_division_tensor)
return torch.true_divide(num, denom)
def _adjust_weights_safe_divide(
score: Tensor, average: Optional[str], multilabel: bool, tp: Tensor, fp: Tensor, fn: Tensor, top_k: int = 1
) -> Tensor:
if average is None or average == "none":
return score
if average == "weighted":
weights = tp + fn
else:
weights = torch.ones_like(score)
if not multilabel:
weights[tp + fp + fn == 0 if top_k == 1 else tp + fn == 0] = 0.0
return _safe_divide(weights * score, weights.sum(-1, keepdim=True)).sum(-1)
def _auc_format_inputs(x: Tensor, y: Tensor) -> tuple[Tensor, Tensor]:
"""Check that auc input is correct."""
x = x.squeeze() if x.ndim > 1 else x
y = y.squeeze() if y.ndim > 1 else y
if x.ndim > 1 or y.ndim > 1:
raise ValueError(
f"Expected both `x` and `y` tensor to be 1d, but got tensors with dimension {x.ndim} and {y.ndim}"
)
if x.numel() != y.numel():
raise ValueError(
f"Expected the same number of elements in `x` and `y` tensor but received {x.numel()} and {y.numel()}"
)
return x, y
def _auc_compute_without_check(x: Tensor, y: Tensor, direction: float, axis: int = -1) -> Tensor:
"""Compute area under the curve using the trapezoidal rule.
Assumes increasing or decreasing order of `x`.
"""
with torch.no_grad():
auc_score: Tensor = torch.trapz(y, x, dim=axis) * direction
return auc_score
def _auc_compute(x: Tensor, y: Tensor, reorder: bool = False) -> Tensor:
"""Compute area under the curve using the trapezoidal rule.
Example:
>>> import torch
>>> x = torch.tensor([1, 2, 3, 4])
>>> y = torch.tensor([1, 2, 3, 4])
>>> _auc_compute(x, y)
tensor(7.5000)
"""
with torch.no_grad():
if reorder:
x, x_idx = torch.sort(x, stable=True)
y = y[x_idx]
dx = x[1:] - x[:-1]
if (dx < 0).any():
if (dx <= 0).all():
direction = -1.0
else:
raise ValueError(
"The `x` tensor is neither increasing or decreasing. Try setting the reorder argument to `True`."
)
else:
direction = 1.0
return _auc_compute_without_check(x, y, direction)
def auc(x: Tensor, y: Tensor, reorder: bool = False) -> Tensor:
"""Compute Area Under the Curve (AUC) using the trapezoidal rule.
Args:
x: x-coordinates, must be either increasing or decreasing
y: y-coordinates
reorder: if True, will reorder the arrays to make it either increasing or decreasing
Return:
Tensor containing AUC score
"""
x, y = _auc_format_inputs(x, y)
return _auc_compute(x, y, reorder=reorder)
def interp(x: Tensor, xp: Tensor, fp: Tensor) -> Tensor:
"""One-dimensional linear interpolation for monotonically increasing sample points.
Returns the one-dimensional piecewise linear interpolation to a function with
given discrete data points :math:`(xp, fp)`, evaluated at :math:`x`.
Adjusted version of this https://github.com/pytorch/pytorch/issues/50334#issuecomment-1000917964
Args:
x: the :math:`x`-coordinates at which to evaluate the interpolated values.
xp: the :math:`x`-coordinates of the data points, must be increasing.
fp: the :math:`y`-coordinates of the data points, same length as `xp`.
Returns:
the interpolated values, same size as `x`.
Example:
>>> x = torch.tensor([0.5, 1.5, 2.5])
>>> xp = torch.tensor([1, 2, 3])
>>> fp = torch.tensor([1, 2, 3])
>>> interp(x, xp, fp)
tensor([0.5000, 1.5000, 2.5000])
"""
m = _safe_divide(fp[1:] - fp[:-1], xp[1:] - xp[:-1])
b = fp[:-1] - (m * xp[:-1])
indices = torch.sum(torch.ge(x[:, None], xp[None, :]), 1) - 1
indices = torch.clamp(indices, 0, len(m) - 1)
return m[indices] * x + b[indices]
def normalize_logits_if_needed(tensor: Tensor, normalization: Literal["sigmoid", "softmax"]) -> Tensor:
"""Normalize logits if needed.
If input tensor is outside the [0,1] we assume that logits are provided and apply the normalization.
Use torch.where to prevent device-host sync.
Args:
tensor: input tensor that may be logits or probabilities
normalization: normalization method, either 'sigmoid' or 'softmax'
Returns:
normalized tensor if needed
Example:
>>> import torch
>>> tensor = torch.tensor([-1.0, 0.0, 1.0])
>>> normalize_logits_if_needed(tensor, normalization="sigmoid")
tensor([0.2689, 0.5000, 0.7311])
>>> tensor = torch.tensor([[-1.0, 0.0, 1.0], [1.0, 0.0, -1.0]])
>>> normalize_logits_if_needed(tensor, normalization="softmax")
tensor([[0.0900, 0.2447, 0.6652],
[0.6652, 0.2447, 0.0900]])
>>> tensor = torch.tensor([0.0, 0.5, 1.0])
>>> normalize_logits_if_needed(tensor, normalization="sigmoid")
tensor([0.0000, 0.5000, 1.0000])
"""
# decrease sigmoid on cpu .
if tensor.device == torch.device("cpu"):
if not torch.all((tensor >= 0) * (tensor <= 1)):
tensor = tensor.sigmoid() if normalization == "sigmoid" else torch.softmax(tensor, dim=1)
return tensor
# decrease device-host sync on device .
condition = ((tensor < 0) | (tensor > 1)).any()
return torch.where(
condition,
torch.sigmoid(tensor) if normalization == "sigmoid" else torch.softmax(tensor, dim=1),
tensor,
)
|