#
# Copyright (c) 2021, NVIDIA CORPORATION.
#
# 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.
#
# Adapted from source code: https://github.com/karlhigley/ranking-metrics-torch
from abc import abstractmethod
import torch
import torchmetrics as tm
from torchmetrics.utilities.data import dim_zero_cat
from merlin_standard_lib import Registry
from .utils import torch_utils
ranking_metrics_registry = Registry.class_registry("torch.ranking_metrics")
[docs]class RankingMetric(tm.Metric):
"""
Metric wrapper for computing ranking metrics@K for session-based task.
Parameters
----------
top_ks : list, default [2, 5])
list of cutoffs
labels_onehot : bool
Enable transform the labels to one-hot representation
"""
def __init__(self, top_ks=None, labels_onehot=False):
super(RankingMetric, self).__init__()
self.top_ks = top_ks or [2, 5]
self.labels_onehot = labels_onehot
# Store the mean of the batch metrics (for each cut-off at topk)
self.add_state("metric_mean", default=[], dist_reduce_fx="cat")
[docs] def update(self, preds: torch.Tensor, target: torch.Tensor, **kwargs): # type: ignore
# Computing the metrics at different cut-offs
if self.labels_onehot:
target = torch_utils.tranform_label_to_onehot(target, preds.size(-1))
metric = self._metric(torch.LongTensor(self.top_ks), preds.view(-1, preds.size(-1)), target)
self.metric_mean.append(metric) # type: ignore
[docs] def compute(self):
# Computing the mean of the batch metrics (for each cut-off at topk)
return dim_zero_cat(self.metric_mean).mean(0)
@abstractmethod
def _metric(self, ks: torch.Tensor, preds: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
"""
Compute a ranking metric over a predictions and one-hot targets.
This method should be overridden by subclasses.
"""
[docs]@ranking_metrics_registry.register_with_multiple_names("precision_at", "precision")
class PrecisionAt(RankingMetric):
def __init__(self, top_ks=None, labels_onehot=False):
super(PrecisionAt, self).__init__(top_ks=top_ks, labels_onehot=labels_onehot)
def _metric(self, ks: torch.Tensor, scores: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
"""Compute precision@K for each of the provided cutoffs
Parameters
----------
ks : torch.Tensor or list
list of cutoffs
scores : torch.Tensor
predicted item scores
labels : torch.Tensor
true item labels
Returns
-------
torch.Tensor:
list of precisions at cutoffs
"""
ks, scores, labels = torch_utils.check_inputs(ks, scores, labels)
_, _, topk_labels = torch_utils.extract_topk(ks, scores, labels)
precisions = torch_utils.create_output_placeholder(scores, ks)
for index, k in enumerate(ks):
precisions[:, index] = torch.sum(topk_labels[:, : int(k)], dim=1) / float(k)
return precisions
[docs]@ranking_metrics_registry.register_with_multiple_names("recall_at", "recall")
class RecallAt(RankingMetric):
def __init__(self, top_ks=None, labels_onehot=False):
super(RecallAt, self).__init__(top_ks=top_ks, labels_onehot=labels_onehot)
def _metric(self, ks: torch.Tensor, scores: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
"""Compute recall@K for each of the provided cutoffs
Parameters
----------
ks : torch.Tensor or list
list of cutoffs
scores : torch.Tensor
predicted item scores
labels : torch.Tensor
true item labels
Returns
-------
torch.Tensor: list of recalls at cutoffs
"""
ks, scores, labels = torch_utils.check_inputs(ks, scores, labels)
_, _, topk_labels = torch_utils.extract_topk(ks, scores, labels)
recalls = torch_utils.create_output_placeholder(scores, ks)
# Compute recalls at K
num_relevant = torch.sum(labels, dim=-1)
rel_indices = (num_relevant != 0).nonzero()
rel_count = num_relevant[rel_indices].squeeze()
if rel_indices.shape[0] > 0:
for index, k in enumerate(ks):
rel_labels = topk_labels[rel_indices, : int(k)].squeeze()
recalls[rel_indices, index] = (
torch.div(torch.sum(rel_labels, dim=-1), rel_count)
.reshape(len(rel_indices), 1)
.to(dtype=torch.float32)
) # Ensuring type is double, because it can be float if --fp16
return recalls
[docs]@ranking_metrics_registry.register_with_multiple_names("avg_precision_at", "avg_precision", "map")
class AvgPrecisionAt(RankingMetric):
def __init__(self, top_ks=None, labels_onehot=False):
super(AvgPrecisionAt, self).__init__(top_ks=top_ks, labels_onehot=labels_onehot)
self.precision_at = PrecisionAt(top_ks)._metric
def _metric(self, ks: torch.Tensor, scores: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
"""Compute average precision at K for provided cutoffs
Parameters
----------
ks : torch.Tensor or list
list of cutoffs
scores : torch.Tensor
2-dim tensor of predicted item scores
labels : torch.Tensor
2-dim tensor of true item labels
Returns
-------
torch.Tensor:
list of average precisions at cutoffs
"""
ks, scores, labels = torch_utils.check_inputs(ks, scores, labels)
topk_scores, _, topk_labels = torch_utils.extract_topk(ks, scores, labels)
avg_precisions = torch_utils.create_output_placeholder(scores, ks)
# Compute average precisions at K
num_relevant = torch.sum(labels, dim=1)
max_k = ks.max().item()
precisions = self.precision_at(1 + torch.arange(max_k), topk_scores, topk_labels)
rel_precisions = precisions * topk_labels
for index, k in enumerate(ks):
total_prec = rel_precisions[:, : int(k)].sum(dim=1)
avg_precisions[:, index] = total_prec / num_relevant.clamp(min=1, max=k).to(
dtype=torch.float32, device=scores.device
) # Ensuring type is double, because it can be float if --fp16
return avg_precisions
[docs]@ranking_metrics_registry.register_with_multiple_names("dcg_at", "dcg")
class DCGAt(RankingMetric):
def __init__(self, top_ks=None, labels_onehot=False):
super(DCGAt, self).__init__(top_ks=top_ks, labels_onehot=labels_onehot)
def _metric(
self, ks: torch.Tensor, scores: torch.Tensor, labels: torch.Tensor, log_base: int = 2
) -> torch.Tensor:
"""Compute discounted cumulative gain at K for provided cutoffs (ignoring ties)
Parameters
----------
ks : torch.Tensor or list
list of cutoffs
scores : torch.Tensor
predicted item scores
labels : torch.Tensor
true item labels
Returns
-------
torch.Tensor :
list of discounted cumulative gains at cutoffs
"""
ks, scores, labels = torch_utils.check_inputs(ks, scores, labels)
topk_scores, topk_indices, topk_labels = torch_utils.extract_topk(ks, scores, labels)
dcgs = torch_utils.create_output_placeholder(scores, ks)
# Compute discounts
discount_positions = torch.arange(ks.max().item()).to(
device=scores.device, dtype=torch.float32
)
discount_log_base = torch.log(
torch.Tensor([log_base]).to(device=scores.device, dtype=torch.float32)
).item()
discounts = 1 / (torch.log(discount_positions + 2) / discount_log_base)
# Compute DCGs at K
for index, k in enumerate(ks):
dcgs[:, index] = torch.sum(
(topk_labels[:, :k] * discounts[:k].repeat(topk_labels.shape[0], 1)), dim=1
).to(
dtype=torch.float32, device=scores.device
) # Ensuring type is double, because it can be float if --fp16
return dcgs
[docs]@ranking_metrics_registry.register_with_multiple_names("ndcg_at", "ndcg")
class NDCGAt(RankingMetric):
def __init__(self, top_ks=None, labels_onehot=False):
super(NDCGAt, self).__init__(top_ks=top_ks, labels_onehot=labels_onehot)
self.dcg_at = DCGAt(top_ks)._metric
def _metric(
self, ks: torch.Tensor, scores: torch.Tensor, labels: torch.Tensor, log_base: int = 2
) -> torch.Tensor:
"""Compute normalized discounted cumulative gain at K for provided cutoffs (ignoring ties)
Parameters
----------
ks : torch.Tensor or list
list of cutoffs
scores : torch.Tensor
predicted item scores
labels : torch.Tensor
true item labels
Returns
-------
torch.Tensor :
list of discounted cumulative gains at cutoffs
"""
ks, scores, labels = torch_utils.check_inputs(ks, scores, labels)
topk_scores, topk_indices, topk_labels = torch_utils.extract_topk(ks, scores, labels)
# ndcgs = _create_output_placeholder(scores, ks) #TODO track if this line is needed
# Compute discounted cumulative gains
gains = self.dcg_at(ks, topk_scores, topk_labels)
normalizing_gains = self.dcg_at(ks, topk_labels, topk_labels)
# Prevent divisions by zero
relevant_pos = (normalizing_gains != 0).nonzero(as_tuple=True)
irrelevant_pos = (normalizing_gains == 0).nonzero(as_tuple=True)
gains[irrelevant_pos] = 0
gains[relevant_pos] /= normalizing_gains[relevant_pos]
return gains
[docs]@ranking_metrics_registry.register_with_multiple_names("mrr_at", "mrr")
class MeanRecipricolRankAt(RankingMetric):
def __init__(self, top_ks=None, labels_onehot=False):
super(MeanRecipricolRankAt, self).__init__(top_ks=top_ks, labels_onehot=labels_onehot)
def _metric(
self, ks: torch.Tensor, scores: torch.Tensor, labels: torch.Tensor, log_base: int = 2
) -> torch.Tensor:
"""Compute mean recipricol rank at K for provided cutoffs (ignoring ties)
Parameters
----------
ks : torch.Tensor or list
list of cutoffs
scores : torch.Tensor
predicted item scores
labels : torch.Tensor
true item labels
Returns
-------
torch.Tensor :
list of mean recipricol rank at cutoffs
"""
ks, scores, labels = torch_utils.check_inputs(ks, scores, labels)
topk_scores, topk_indices, topk_labels = torch_utils.extract_topk(ks, scores, labels)
results = torch.zeros(scores.shape[0], len(ks)).to(
device=scores.device, dtype=torch.float32
)
for index, k in enumerate(ks):
values, _ = (topk_labels[:, :k] / (torch.arange(k) + 1)).max(dim=1)
results[:, index] = values
return results