https://github.com/facebookresearch/pythia
Tip revision: 1fbf93c64f067f4d6e5519b066b7b7ec671c8dcf authored by Vedanuj Goswami on 26 June 2021, 20:26:56 UTC
[chores] PT upgrade
[chores] PT upgrade
Tip revision: 1fbf93c
metrics.py
# Copyright (c) Facebook, Inc. and its affiliates.
"""
The metrics module contains implementations of various metrics used commonly to
understand how well our models are performing. For e.g. accuracy, vqa_accuracy,
r@1 etc.
For implementing your own metric, you need to follow these steps:
1. Create your own metric class and inherit ``BaseMetric`` class.
2. In the ``__init__`` function of your class, make sure to call
``super().__init__('name')`` where 'name' is the name of your metric. If
you require any parameters in your ``__init__`` function, you can use
keyword arguments to represent them and metric constructor will take care of
providing them to your class from config.
3. Implement a ``calculate`` function which takes in ``SampleList`` and
`model_output` as input and return back a float tensor/number.
4. Register your metric with a key 'name' by using decorator,
``@registry.register_metric('name')``.
Example::
import torch
from mmf.common.registry import registry
from mmf.modules.metrics import BaseMetric
@registry.register_metric("some")
class SomeMetric(BaseMetric):
def __init__(self, some_param=None):
super().__init__("some")
....
def calculate(self, sample_list, model_output):
metric = torch.tensor(2, dtype=torch.float)
return metric
Example config for above metric::
model_config:
pythia:
metrics:
- type: some
params:
some_param: a
"""
import collections
import warnings
from typing import Dict
import torch
from mmf.common.registry import registry
from mmf.datasets.processors.processors import EvalAIAnswerProcessor
from mmf.utils.logger import log_class_usage
from sklearn.metrics import (
average_precision_score,
f1_score,
precision_recall_curve,
roc_auc_score,
)
from torch import Tensor
def _convert_to_one_hot(expected, output):
# This won't get called in case of multilabel, only multiclass or binary
# as multilabel will anyways be multi hot vector
if output.squeeze().dim() != expected.squeeze().dim() and expected.dim() == 1:
expected = torch.nn.functional.one_hot(
expected.long(), num_classes=output.size(-1)
).float()
return expected
class Metrics:
"""Internally used by MMF, Metrics acts as wrapper for handling
calculation of metrics over various metrics specified by the model in
the config. It initializes all of the metrics and when called it runs
calculate on each of them one by one and returns back a dict with proper
naming back. For e.g. an example dict returned by Metrics class:
``{'val/vqa_accuracy': 0.3, 'val/r@1': 0.8}``
Args:
metric_list (ListConfig): List of DictConfigs where each DictConfig
specifies name and parameters of the
metrics used.
"""
def __init__(self, metric_list):
if not isinstance(metric_list, collections.abc.Sequence):
metric_list = [metric_list]
self.metrics = self._init_metrics(metric_list)
def _init_metrics(self, metric_list):
metrics = {}
self.required_params = {"dataset_name", "dataset_type"}
for metric in metric_list:
params = {}
dataset_names = []
if isinstance(metric, collections.abc.Mapping):
if "type" not in metric:
raise ValueError(
f"Metric {metric} needs to have 'type' attribute "
+ "or should be a string"
)
metric_type = key = metric.type
params = metric.get("params", {})
# Support cases where uses need to give custom metric name
if "key" in metric:
key = metric.key
# One key should only be used once
if key in metrics:
raise RuntimeError(
f"Metric with type/key '{metric_type}' has been defined more "
+ "than once in metric list."
)
# a custom list of dataset where this metric will be applied
if "datasets" in metric:
dataset_names = metric.datasets
else:
if not isinstance(metric, str):
raise TypeError(
"Metric {} has inappropriate type"
"'dict' or 'str' allowed".format(metric)
)
metric_type = key = metric
metric_cls = registry.get_metric_class(metric_type)
if metric_cls is None:
raise ValueError(
f"No metric named {metric_type} registered to registry"
)
metric_instance = metric_cls(**params)
metric_instance.name = key
metric_instance.set_applicable_datasets(dataset_names)
metrics[key] = metric_instance
self.required_params.update(metrics[key].required_params)
return metrics
def __call__(self, sample_list, model_output, *args, **kwargs):
values = {}
dataset_type = sample_list.dataset_type
dataset_name = sample_list.dataset_name
with torch.no_grad():
for metric_name, metric_object in self.metrics.items():
if not metric_object.is_dataset_applicable(dataset_name):
continue
key = f"{dataset_type}/{dataset_name}/{metric_name}"
values[key] = metric_object._calculate_with_checks(
sample_list, model_output, *args, **kwargs
)
if not isinstance(values[key], torch.Tensor):
values[key] = torch.tensor(values[key], dtype=torch.float)
else:
values[key] = values[key].float()
if values[key].dim() == 0:
values[key] = values[key].view(1)
registry.register(
"{}.{}.{}".format("metrics", sample_list.dataset_name, dataset_type), values
)
return values
class BaseMetric:
"""Base class to be inherited by all metrics registered to MMF. See
the description on top of the file for more information. Child class must
implement ``calculate`` function.
Args:
name (str): Name of the metric.
"""
def __init__(self, name, *args, **kwargs):
self.name = name
self.required_params = ["scores", "targets"]
# the set of datasets where this metric will be applied
# an empty set means it will be applied on *all* datasets
self._dataset_names = set()
log_class_usage("Metric", self.__class__)
@property
def name(self):
return self._name
@name.setter
def name(self, name):
self._name = name
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Abstract method to be implemented by the child class. Takes
in a ``SampleList`` and a dict returned by model as output and
returns back a float tensor/number indicating value for this metric.
Args:
sample_list (SampleList): SampleList provided by the dataloader for the
current iteration.
model_output (Dict): Output dict from the model for the current
SampleList
Returns:
torch.Tensor|float: Value of the metric.
"""
# Override in your child class
raise NotImplementedError("'calculate' must be implemented in the child class")
def __call__(self, *args, **kwargs):
return self.calculate(*args, **kwargs)
def _calculate_with_checks(self, *args, **kwargs):
value = self.calculate(*args, **kwargs)
return value
def set_applicable_datasets(self, dataset_names):
self._dataset_names = set(dataset_names)
def is_dataset_applicable(self, dataset_name):
return len(self._dataset_names) == 0 or dataset_name in self._dataset_names
@registry.register_metric("accuracy")
class Accuracy(BaseMetric):
"""Metric for calculating accuracy.
**Key:** ``accuracy``
"""
def __init__(self, score_key="scores", target_key="targets", topk=1):
super().__init__("accuracy")
self.score_key = score_key
self.target_key = target_key
self.topk = topk
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate accuracy and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: accuracy.
"""
output = model_output[self.score_key]
batch_size = output.shape[0]
expected = sample_list[self.target_key]
assert (
output.dim() <= 2
), "Output from model shouldn't have more than dim 2 for accuracy"
assert (
expected.dim() <= 2
), "Expected target shouldn't have more than dim 2 for accuracy"
if output.dim() == 2:
output = output.topk(self.topk, 1, True, True)[1].t().squeeze()
# If more than 1
# If last dim is 1, we directly have class indices
if expected.dim() == 2 and expected.size(-1) != 1:
expected = expected.topk(self.topk, 1, True, True)[1].t().squeeze()
correct = (expected == output.squeeze()).sum().float()
return correct / batch_size
@registry.register_metric("topk_accuracy")
class TopKAccuracy(Accuracy):
def __init__(self, score_key: str, k: int):
super().__init__(score_key=score_key, topk=k)
@registry.register_metric("caption_bleu4")
class CaptionBleu4Metric(BaseMetric):
"""Metric for calculating caption accuracy using BLEU4 Score.
**Key:** ``caption_bleu4``
"""
def __init__(self):
import nltk.translate.bleu_score as bleu_score
self._bleu_score = bleu_score
super().__init__("caption_bleu4")
self.caption_processor = registry.get("coco_caption_processor")
self.required_params = ["scores", "answers", "captions"]
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate accuracy and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: bleu4 score.
"""
# Create reference and hypotheses captions.
references = []
hypotheses = []
# References
targets = sample_list.answers
for j, _ in enumerate(targets):
img_captions = [
self.caption_processor(c)["tokens"] for c in targets[j].tolist()
]
references.append(img_captions)
# Hypotheses
if "captions" in model_output:
scores = model_output["captions"]
else:
scores = torch.max(model_output["scores"], dim=-1)[1]
scores = scores.tolist()
predictions = []
for j, _ in enumerate(scores):
caption = self.caption_processor(scores[j])["tokens"]
predictions.append(caption)
hypotheses.extend(predictions)
assert len(references) == len(hypotheses)
bleu4 = self._bleu_score.corpus_bleu(references, hypotheses)
return targets.new_tensor(bleu4, dtype=torch.float)
@registry.register_metric("vqa_accuracy")
class VQAAccuracy(BaseMetric):
"""
Calculate VQAAccuracy. Find more information here_
**Key**: ``vqa_accuracy``.
.. _here: https://visualqa.org/evaluation.html
"""
def __init__(self):
super().__init__("vqa_accuracy")
def _masked_unk_softmax(self, x, dim, mask_idx):
x1 = torch.nn.functional.softmax(x, dim=dim)
x1[:, mask_idx] = 0
x1_sum = torch.sum(x1, dim=1, keepdim=True)
y = x1 / x1_sum
return y
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate vqa accuracy and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: VQA Accuracy
"""
output = model_output["scores"]
# for three branch movie+mcan model
if output.dim() == 3:
output = output[:, 0]
expected = sample_list["targets"]
output = self._masked_unk_softmax(output, 1, 0)
output = output.argmax(dim=1) # argmax
one_hots = expected.new_zeros(*expected.size())
one_hots.scatter_(1, output.view(-1, 1), 1)
scores = one_hots * expected
accuracy = torch.sum(scores) / expected.size(0)
return accuracy
@registry.register_metric("vqa_evalai_accuracy")
class VQAEvalAIAccuracy(BaseMetric):
"""
Calculate Eval AI VQAAccuracy. Find more information here_
This is more accurate and similar comparision to Eval AI
but is slower compared to vqa_accuracy.
**Key**: ``vqa_evalai_accuracy``.
.. _here: https://visualqa.org/evaluation.html
"""
def __init__(self):
super().__init__("vqa_evalai_accuracy")
self.evalai_answer_processor = EvalAIAnswerProcessor()
self.required_params = ["scores", "answers", "context_tokens"]
def _masked_unk_softmax(self, x, dim, mask_idx):
x1 = torch.nn.functional.softmax(x, dim=dim)
x1[:, mask_idx] = 0
x1_sum = torch.sum(x1, dim=1, keepdim=True)
y = x1 / x1_sum
return y
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate vqa accuracy and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: VQA Accuracy
"""
output = model_output["scores"]
expected = sample_list["answers"]
answer_processor = registry.get(sample_list.dataset_name + "_answer_processor")
answer_space_size = answer_processor.get_true_vocab_size()
output = self._masked_unk_softmax(output, 1, 0)
output = output.argmax(dim=1).clone().tolist()
accuracy = []
for idx, answer_id in enumerate(output):
if answer_id >= answer_space_size:
answer_id -= answer_space_size
answer = sample_list["context_tokens"][idx][answer_id]
else:
answer = answer_processor.idx2word(answer_id)
answer = self.evalai_answer_processor(answer)
gt_answers = [self.evalai_answer_processor(x) for x in expected[idx]]
gt_answers = list(enumerate(gt_answers))
gt_acc = []
for gt_answer in gt_answers:
other_answers = [item for item in gt_answers if item != gt_answer]
matching_answers = [item for item in other_answers if item[1] == answer]
acc = min(1, float(len(matching_answers)) / 3)
gt_acc.append(acc)
avgGTAcc = float(sum(gt_acc)) / len(gt_acc)
accuracy.append(avgGTAcc)
accuracy = float(sum(accuracy)) / len(accuracy)
return model_output["scores"].new_tensor(accuracy, dtype=torch.float)
class RecallAtK(BaseMetric):
def __init__(self, name="recall@k"):
super().__init__(name)
def score_to_ranks(self, scores):
# sort in descending order - largest score gets highest rank
sorted_ranks, ranked_idx = scores.sort(1, descending=True)
# convert from ranked_idx to ranks
ranks = ranked_idx.clone().fill_(0)
for i in range(ranked_idx.size(0)):
for j in range(100):
ranks[i][ranked_idx[i][j]] = j
ranks += 1
return ranks
def get_gt_ranks(self, ranks, ans_ind):
_, ans_ind = ans_ind.max(dim=1)
ans_ind = ans_ind.view(-1)
gt_ranks = torch.LongTensor(ans_ind.size(0))
for i in range(ans_ind.size(0)):
gt_ranks[i] = int(ranks[i, ans_ind[i].long()])
return gt_ranks
def process_ranks(self, ranks):
num_opts = 100
# none of the values should be 0, there is gt in options
if torch.sum(ranks.le(0)) > 0:
num_zero = torch.sum(ranks.le(0))
warnings.warn(f"Some of ranks are zero: {num_zero}")
ranks = ranks[ranks.gt(0)]
# rank should not exceed the number of options
if torch.sum(ranks.ge(num_opts + 1)) > 0:
num_ge = torch.sum(ranks.ge(num_opts + 1))
warnings.warn(f"Some of ranks > 100: {num_ge}")
ranks = ranks[ranks.le(num_opts + 1)]
return ranks
def get_ranks(self, sample_list, model_output, *args, **kwargs):
output = model_output["scores"]
expected = sample_list["targets"]
ranks = self.score_to_ranks(output)
gt_ranks = self.get_gt_ranks(ranks, expected)
ranks = self.process_ranks(gt_ranks)
return ranks.float()
def calculate(self, sample_list, model_output, k, *args, **kwargs):
ranks = self.get_ranks(sample_list, model_output)
recall = float(torch.sum(torch.le(ranks, k))) / ranks.size(0)
return recall
@registry.register_metric("r@1")
class RecallAt1(RecallAtK):
"""
Calculate Recall@1 which specifies how many time the chosen candidate
was rank 1.
**Key**: ``r@1``.
"""
def __init__(self):
super().__init__("r@1")
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate Recall@1 and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: Recall@1
"""
return super().calculate(sample_list, model_output, k=1)
@registry.register_metric("r@5")
class RecallAt5(RecallAtK):
"""
Calculate Recall@5 which specifies how many time the chosen candidate
was among first 5 rank.
**Key**: ``r@5``.
"""
def __init__(self):
super().__init__("r@5")
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate Recall@5 and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: Recall@5
"""
return super().calculate(sample_list, model_output, k=5)
@registry.register_metric("r@10")
class RecallAt10(RecallAtK):
"""
Calculate Recall@10 which specifies how many time the chosen candidate
was among first 10 ranks.
**Key**: ``r@10``.
"""
def __init__(self):
super().__init__("r@10")
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate Recall@10 and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: Recall@10
"""
return super().calculate(sample_list, model_output, k=10)
@registry.register_metric("mean_r")
class MeanRank(RecallAtK):
"""
Calculate MeanRank which specifies what was the average rank of the chosen
candidate.
**Key**: ``mean_r``.
"""
def __init__(self):
super().__init__("mean_r")
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate Mean Rank and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: mean rank
"""
ranks = self.get_ranks(sample_list, model_output)
return torch.mean(ranks)
@registry.register_metric("mean_rr")
class MeanReciprocalRank(RecallAtK):
"""
Calculate reciprocal of mean rank..
**Key**: ``mean_rr``.
"""
def __init__(self):
super().__init__("mean_rr")
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate Mean Reciprocal Rank and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: Mean Reciprocal Rank
"""
ranks = self.get_ranks(sample_list, model_output)
return torch.mean(ranks.reciprocal())
@registry.register_metric("textvqa_accuracy")
class TextVQAAccuracy(BaseMetric):
def __init__(self):
super().__init__("textvqa_accuracy")
import mmf.utils.m4c_evaluators as evaluators
self.evaluator = evaluators.TextVQAAccuracyEvaluator()
self.required_params = ["scores", "answers", "context_tokens"]
self.gt_key = "answers"
def calculate(self, sample_list, model_output, *args, **kwargs):
answer_processor = registry.get(sample_list.dataset_name + "_answer_processor")
batch_size = sample_list.context_tokens.size(0)
pred_answers = model_output["scores"].argmax(dim=-1)
context_tokens = sample_list.context_tokens.cpu().numpy()
answers = sample_list.get(self.gt_key).cpu().numpy()
answer_space_size = answer_processor.get_true_vocab_size()
predictions = []
from mmf.utils.distributed import byte_tensor_to_object
from mmf.utils.text import word_tokenize
for idx in range(batch_size):
tokens = byte_tensor_to_object(context_tokens[idx])
answer_words = []
for answer_id in pred_answers[idx].tolist():
if answer_id >= answer_space_size:
answer_id -= answer_space_size
answer_words.append(word_tokenize(tokens[answer_id]))
else:
if answer_id == answer_processor.EOS_IDX:
break
answer_words.append(
answer_processor.answer_vocab.idx2word(answer_id)
)
pred_answer = " ".join(answer_words).replace(" 's", "'s")
gt_answers = byte_tensor_to_object(answers[idx])
predictions.append({"pred_answer": pred_answer, "gt_answers": gt_answers})
accuracy = self.evaluator.eval_pred_list(predictions)
accuracy = torch.tensor(accuracy).to(sample_list.context_tokens.device)
return accuracy
@registry.register_metric("stvqa_anls")
class STVQAANLS(TextVQAAccuracy):
def __init__(self):
super().__init__()
self.name = "stvqa_anls"
import mmf.utils.m4c_evaluators as evaluators
self.evaluator = evaluators.STVQAANLSEvaluator()
@registry.register_metric("stvqa_accuracy")
class STVQAAccuracy(TextVQAAccuracy):
def __init__(self):
super().__init__()
self.name = "stvqa_accuracy"
import mmf.utils.m4c_evaluators as evaluators
self.evaluator = evaluators.STVQAAccuracyEvaluator()
@registry.register_metric("ocrvqa_accuracy")
class OCRVQAAccuracy(STVQAAccuracy):
def __init__(self):
super().__init__()
# same as STVQAAccuracy except for the name
self.name = "ocrvqa_accuracy"
@registry.register_metric("textcaps_bleu4")
class TextCapsBleu4(TextVQAAccuracy):
def __init__(self):
super().__init__()
self.name = "textcaps_bleu4"
self.required_params = ["scores", "ref_strs", "context_tokens"]
self.gt_key = "ref_strs"
import mmf.utils.m4c_evaluators as evaluators
self.evaluator = evaluators.TextCapsBleu4Evaluator()
@registry.register_metric("f1")
class F1(BaseMetric):
"""Metric for calculating F1. Can be used with type and params
argument for customization. params will be directly passed to sklearn
f1 function.
**Key:** ``f1``
"""
def __init__(self, *args, **kwargs):
super().__init__("f1")
self._multilabel = kwargs.pop("multilabel", False)
self._sk_kwargs = kwargs
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate f1 and return it back.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration
model_output (Dict): Dict returned by model.
Returns:
torch.FloatTensor: f1.
"""
scores = model_output["scores"]
expected = sample_list["targets"]
if self._multilabel:
output = torch.sigmoid(scores)
output = torch.round(output)
expected = _convert_to_one_hot(expected, output)
else:
# Multiclass, or binary case
output = scores.argmax(dim=-1)
if expected.dim() != 1:
# Probably one-hot, convert back to class indices array
expected = expected.argmax(dim=-1)
value = f1_score(expected.cpu(), output.cpu(), **self._sk_kwargs)
return expected.new_tensor(value, dtype=torch.float)
@registry.register_metric("macro_f1")
class MacroF1(F1):
"""Metric for calculating Macro F1.
**Key:** ``macro_f1``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="macro", **kwargs)
self.name = "macro_f1"
@registry.register_metric("micro_f1")
class MicroF1(F1):
"""Metric for calculating Micro F1.
**Key:** ``micro_f1``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="micro", **kwargs)
self.name = "micro_f1"
@registry.register_metric("binary_f1")
class BinaryF1(F1):
"""Metric for calculating Binary F1.
**Key:** ``binary_f1``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="micro", labels=[1], **kwargs)
self.name = "binary_f1"
@registry.register_metric("multilabel_f1")
class MultiLabelF1(F1):
"""Metric for calculating Multilabel F1.
**Key:** ``multilabel_f1``
"""
def __init__(self, *args, **kwargs):
super().__init__(multilabel=True, **kwargs)
self.name = "multilabel_f1"
@registry.register_metric("multilabel_micro_f1")
class MultiLabelMicroF1(MultiLabelF1):
"""Metric for calculating Multilabel Micro F1.
**Key:** ``multilabel_micro_f1``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="micro", **kwargs)
self.name = "multilabel_micro_f1"
@registry.register_metric("multilabel_macro_f1")
class MultiLabelMacroF1(MultiLabelF1):
"""Metric for calculating Multilabel Macro F1.
**Key:** ``multilabel_macro_f1``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="macro", **kwargs)
self.name = "multilabel_macro_f1"
@registry.register_metric("roc_auc")
class ROC_AUC(BaseMetric):
"""Metric for calculating ROC_AUC.
See more details at `sklearn.metrics.roc_auc_score <http://scikit-learn.org/stable/modules/generated/sklearn.metrics.roc_auc_score.html#sklearn.metrics.roc_auc_score>`_ # noqa
**Note**: ROC_AUC is not defined when expected tensor only contains one
label. Make sure you have both labels always or use it on full val only
**Key:** ``roc_auc``
"""
def __init__(self, *args, **kwargs):
super().__init__("roc_auc")
self._sk_kwargs = kwargs
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate ROC_AUC and returns it back. The function performs softmax
on the logits provided and then calculated the ROC_AUC.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration.
model_output (Dict): Dict returned by model. This should contain "scores"
field pointing to logits returned from the model.
Returns:
torch.FloatTensor: ROC_AUC.
"""
output = torch.nn.functional.softmax(model_output["scores"], dim=-1)
expected = sample_list["targets"]
expected = _convert_to_one_hot(expected, output)
value = roc_auc_score(expected.cpu(), output.cpu(), **self._sk_kwargs)
return expected.new_tensor(value, dtype=torch.float)
@registry.register_metric("micro_roc_auc")
class MicroROC_AUC(ROC_AUC):
"""Metric for calculating Micro ROC_AUC.
**Key:** ``micro_roc_auc``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="micro", **kwargs)
self.name = "micro_roc_auc"
@registry.register_metric("macro_roc_auc")
class MacroROC_AUC(ROC_AUC):
"""Metric for calculating Macro ROC_AUC.
**Key:** ``macro_roc_auc``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="macro", **kwargs)
self.name = "macro_roc_auc"
@registry.register_metric("ap")
class AveragePrecision(BaseMetric):
"""Metric for calculating Average Precision.
See more details at `sklearn.metrics.average_precision_score <http://scikit-learn.org/stable/modules/generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score>`_ # noqa
If you are looking for binary case, please take a look at binary_ap
**Key:** ``ap``
"""
def __init__(self, *args, **kwargs):
super().__init__("ap")
self._sk_kwargs = kwargs
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate AP and returns it back. The function performs softmax
on the logits provided and then calculated the AP.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration.
model_output (Dict): Dict returned by model. This should contain "scores"
field pointing to logits returned from the model.
Returns:
torch.FloatTensor: AP.
"""
output = torch.nn.functional.softmax(model_output["scores"], dim=-1)
expected = sample_list["targets"]
expected = _convert_to_one_hot(expected, output)
value = average_precision_score(expected.cpu(), output.cpu(), **self._sk_kwargs)
return expected.new_tensor(value, dtype=torch.float)
@registry.register_metric("binary_ap")
class BinaryAP(AveragePrecision):
"""Metric for calculating Binary Average Precision.
See more details at `sklearn.metrics.average_precision_score <http://scikit-learn.org/stable/modules/generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score>`_ # noqa
**Key:** ``binary_ap``
"""
def __init__(self, *args, **kwargs):
super().__init__(**kwargs)
self.name = "binary_ap"
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate Binary AP and returns it back. The function performs softmax
on the logits provided and then calculated the binary AP.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration.
model_output (Dict): Dict returned by model. This should contain "scores"
field pointing to logits returned from the model.
Returns:
torch.FloatTensor: AP.
"""
output = torch.nn.functional.softmax(model_output["scores"], dim=-1)
# Take the score for positive (1) label
output = output[:, 1]
expected = sample_list["targets"]
# One hot format -> Labels
if expected.dim() == 2:
expected = expected.argmax(dim=1)
value = average_precision_score(expected.cpu(), output.cpu(), **self._sk_kwargs)
return expected.new_tensor(value, dtype=torch.float)
@registry.register_metric("micro_ap")
class MicroAP(AveragePrecision):
"""Metric for calculating Micro Average Precision.
**Key:** ``micro_ap``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="micro", **kwargs)
self.name = "micro_ap"
@registry.register_metric("macro_ap")
class MacroAP(AveragePrecision):
"""Metric for calculating Macro Average Precision.
**Key:** ``macro_ap``
"""
def __init__(self, *args, **kwargs):
super().__init__(average="macro", **kwargs)
self.name = "macro_ap"
@registry.register_metric("r@pk")
class RecallAtPrecisionK(BaseMetric):
"""Metric for calculating recall when precision is above a
particular threshold. Use `p_threshold` param to specify the
precision threshold i.e. k. Accepts precision in both 0-1
and 1-100 format.
**Key:** ``r@pk``
"""
def __init__(self, p_threshold, *args, **kwargs):
"""Initialization function recall @ precision k
Args:
p_threshold (float): Precision threshold
"""
super().__init__(name="r@pk")
self.name = "r@pk"
self.p_threshold = p_threshold if p_threshold < 1 else p_threshold / 100
def calculate(self, sample_list, model_output, *args, **kwargs):
"""Calculate Recall at precision k and returns it back. The function
performs softmax on the logits provided and then calculated the metric.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration.
model_output (Dict): Dict returned by model. This should contain "scores"
field pointing to logits returned from the model.
Returns:
torch.FloatTensor: Recall @ precision k.
"""
output = torch.nn.functional.softmax(model_output["scores"], dim=-1)[:, 1]
expected = sample_list["targets"]
# One hot format -> Labels
if expected.dim() == 2:
expected = expected.argmax(dim=1)
precision, recall, thresh = precision_recall_curve(expected.cpu(), output.cpu())
try:
value, _ = max(
(r, p) for p, r in zip(precision, recall) if p >= self.p_threshold
)
except ValueError:
value = 0
return expected.new_tensor(value, dtype=torch.float)
@registry.register_metric("r@k_retrieval")
class RecallAtK_ret(BaseMetric):
def __init__(self, name="recall@k"):
super().__init__(name)
def _get_RatK_multi(
self, correlations: Tensor, labels: Tensor, k: int, factor: int
):
_, top_k_ids = torch.topk(correlations, k, dim=1)
hits = (
torch.logical_and(
labels[:, None] <= top_k_ids, top_k_ids < labels[:, None] + factor
)
.long()
.max(dim=1)[0]
)
return hits
def calculate(
self,
sample_list: Dict[str, Tensor],
model_output: Dict[str, Tensor],
k: int,
flip=False,
*args,
**kwargs,
):
# calculate image to text retrieval recalls
# correlations shape is either BxB or Bx(5B)
# when flip=True, calculate text to image
image_embeddings = model_output["scores"]
text_embeddings = model_output["targets"]
correlations = image_embeddings @ text_embeddings.t() # B x B or Bx5B
assert correlations.shape[1] % correlations.shape[0] == 0
batch_size = correlations.shape[0]
factor = correlations.shape[1] // correlations.shape[0]
labels = torch.arange(batch_size, device=image_embeddings.device) * factor
if flip:
correlations = correlations.t() # 5B x B
labels = torch.arange(batch_size, device=image_embeddings.device)
labels = labels[:, None].expand(-1, factor).flatten()
factor = 1
hits = self._get_RatK_multi(correlations, labels, k, factor)
ratk = hits.sum().float() / hits.shape[0]
return ratk
@registry.register_metric("r@1_retrieval")
class RecallAt1_ret(RecallAtK_ret):
def __init__(self):
super().__init__("r@1")
def calculate(
self,
sample_list: Dict[str, Tensor],
model_output: Dict[str, Tensor],
*args,
**kwargs,
):
ratk = super().calculate(sample_list, model_output, 1)
return ratk
@registry.register_metric("r@1_rev_retrieval")
class RecallAt1_rev_ret(RecallAtK_ret):
def __init__(self):
super().__init__("r@1_rev")
def calculate(
self,
sample_list: Dict[str, Tensor],
model_output: Dict[str, Tensor],
*args,
**kwargs,
):
ratk = super().calculate(sample_list, model_output, 1, flip=True)
return ratk
@registry.register_metric("r@5_retrieval")
class RecallAt5_ret(RecallAtK_ret):
def __init__(self):
super().__init__("r@5")
def calculate(
self,
sample_list: Dict[str, Tensor],
model_output: Dict[str, Tensor],
*args,
**kwargs,
):
ratk = super().calculate(sample_list, model_output, 5)
return ratk
@registry.register_metric("r@5_rev_retrieval")
class RecallAt5_rev_ret(RecallAtK_ret):
def __init__(self):
super().__init__("r@5_rev")
def calculate(
self,
sample_list: Dict[str, Tensor],
model_output: Dict[str, Tensor],
*args,
**kwargs,
):
ratk = super().calculate(sample_list, model_output, 5, flip=True)
return ratk
@registry.register_metric("r@10_retrieval")
class RecallAt10_ret(RecallAtK_ret):
def __init__(self):
super().__init__("r@10")
def calculate(
self,
sample_list: Dict[str, Tensor],
model_output: Dict[str, Tensor],
*args,
**kwargs,
):
ratk = super().calculate(sample_list, model_output, 10)
return ratk
@registry.register_metric("r@10_rev_retrieval")
class RecallAt10_rev_ret(RecallAtK_ret):
def __init__(self):
super().__init__("r@10_rev")
def calculate(
self,
sample_list: Dict[str, Tensor],
model_output: Dict[str, Tensor],
*args,
**kwargs,
):
ratk = super().calculate(sample_list, model_output, 10, flip=True)
return ratk
@registry.register_metric("detection_mean_ap")
class DetectionMeanAP(BaseMetric):
"""Metric for calculating the detection mean average precision (mAP) using the COCO
evaluation toolkit, returning the default COCO-style mAP@IoU=0.50:0.95
**Key:** ``detection_mean_ap``
"""
def __init__(self, dataset_json_files, *args, **kwargs):
"""Initialization function detection mean AP (mAP)
Args:
dataset_json_files (Dict): paths to the dataset (instance) json files
for each dataset type and dataset name in the following format:
``{'val/detection_coco': '/path/to/instances_val2017.json', ...}``
"""
super().__init__("detection_mean_ap")
self.required_params = ["__prediction_report__"]
self.dataset_json_files = dataset_json_files
def calculate(
self, sample_list, model_output, execute_on_master_only=True, *args, **kwargs
):
"""Calculate detection mean AP (mAP) from the prediction list and the dataset
annotations. The function returns COCO-style mAP@IoU=0.50:0.95.
Args:
sample_list (SampleList): SampleList provided by DataLoader for
current iteration.
model_output (Dict): Dict returned by model. This should contain
"prediction_report" field, which is a list of
detection predictions from the model.
execute_on_master_only (bool): Whether to only run mAP evaluation on the
master node over the gathered detection prediction
(to avoid wasting computation and CPU OOM).
Default: True (only run mAP evaluation on master).
Returns:
torch.FloatTensor: COCO-style mAP@IoU=0.50:0.95.
"""
# as the detection mAP metric is run on the entire dataset-level predictions,
# which are *already* gathered from all notes, the evaluation should only happen
# in one node and broadcasted to other nodes (to avoid CPU OOM due to concurrent
# mAP evaluation)
from mmf.utils.distributed import broadcast_tensor, is_master
from mmf.utils.general import get_current_device
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
device = get_current_device()
if execute_on_master_only and not is_master():
# dummy mAP to be override in boardcasting
mAP = torch.tensor(-1, dtype=torch.float, device=device)
else:
predictions = model_output.prediction_report
cocoGt = COCO(
self.dataset_json_files[sample_list.dataset_name][
sample_list.dataset_type
]
)
cocoDt = cocoGt.loadRes(predictions)
cocoEval = COCOeval(cocoGt, cocoDt, "bbox")
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
mAP = torch.tensor(cocoEval.stats[0], dtype=torch.float, device=device)
if execute_on_master_only:
mAP = broadcast_tensor(mAP, src=0)
return mAP
