File size: 5,304 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 |
# 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
import torch
from torch import Tensor
def _check_shape_and_type_consistency(preds: Tensor, target: Tensor) -> None:
"""Check shape and type consistency of input vectors.
Args:
preds:
Logits or a unnormalized score assigned to each token in a sequence with shape [batch_size, seq_len,
vocab_size]. Scores will be normalized internally using softmax.
target:
Ground truth values with a shape [batch_size, seq_len].
Raises:
ValueError:
If ``preds`` tensor has no 3 dimensions.
ValueError:
If ``target`` tensor has no 2 dimensions.
ValueError:
If the first two dimensions of ``preds`` and ``target`` do not equal.
TypeError:
If ``preds`` dtype is not one of ``(torch.float16, torch.float32, torch.float64)``
TypeError:
If ``target`` is not of a type LongTensor (torch.int64)
"""
if len(preds.shape) != 3:
raise ValueError(
"Input tensor `preds` is expected to have 3 dimensions, [batch_size, seq_len, vocab_size],"
f" but got {len(preds.shape)}."
)
if len(target.shape) != 2:
raise ValueError(
"Input tensor `target` is expected to have 2 dimensions, [batch_size, seq_len],"
f" but got {len(target.shape)}."
)
if preds.shape[:2] != target.shape:
raise ValueError(
"Input tensors `preds` and `target` are expected to have equaling first two dimensions,"
f" [batch_size, seq_len], but got {preds.shape[:2]} and {target.shape}."
)
if not preds.is_floating_point():
raise TypeError(f"Input tensor `preds` is expected to be of floating point type but got {preds.dtype}.")
if target.dtype != torch.int64:
raise TypeError(f"Input tensor `target` is expected to be of a type {torch.int64} but got {target.dtype}.")
def _perplexity_update(preds: Tensor, target: Tensor, ignore_index: Optional[int] = None) -> tuple[Tensor, Tensor]:
"""Compute intermediate statistics for Perplexity.
Args:
preds:
Logits or a unnormalized score assigned to each token in a sequence with shape [batch_size, seq_len,
vocab_size]. Scores will be normalized internally using softmax.
target:
Ground truth values with a shape [batch_size, seq_len].
ignore_index:
Integer specifying a target class to ignore. If given, this class index does not contribute
to the returned score.
Returns:
Log probabilities, summed over all samples
Number of samples
"""
_check_shape_and_type_consistency(preds, target)
probs = torch.nn.functional.softmax(preds.reshape(-1, preds.shape[-1]), dim=1)
target = target.reshape(-1)
if ignore_index is not None:
mask = target.ne(ignore_index)
target = target.where(target != ignore_index, torch.tensor(0, device=target.device))
else:
mask = torch.ones_like(target, dtype=torch.bool)
probs = probs[torch.arange(target.numel()), target][mask]
total_log_probs = -probs.log().sum()
count = mask.sum()
return total_log_probs, count
def _perplexity_compute(total: Tensor, count: Tensor) -> Tensor:
"""Compute the Perplexity.
Args:
total: Log probabilities, summed over all samples
count: Number of samples
Returns:
Perplexity
"""
return torch.exp(total / count)
def perplexity(preds: Tensor, target: Tensor, ignore_index: Optional[int] = None) -> Tensor:
"""Perplexity measures how well a language model predicts a text sample.
This metric is calculated as the average number of bits per word a model needs to represent the sample.
Args:
preds:
Logits or a unnormalized score assigned to each token in a sequence with shape [batch_size, seq_len,
vocab_size], which is the output of a language model. Scores will be normalized internally using softmax.
target:
Ground truth values with a shape [batch_size, seq_len].
ignore_index:
Integer specifying a target class to ignore. If given, this class index does not contribute
to the returned score.
Returns:
Perplexity value
Examples:
>>> from torch import rand, randint
>>> preds = rand(2, 8, 5)
>>> target = randint(5, (2, 8))
>>> target[0, 6:] = -100
>>> perplexity(preds, target, ignore_index=-100)
tensor(5.8540)
"""
total, count = _perplexity_update(preds, target, ignore_index)
return _perplexity_compute(total, count)
|