class VideoTop1Metric(Metric):
def __init__(self):
self.am = AverageMeter()
def update(self, prediction, target):
if target.dim() == 3:
target = target.max(dim=1)[0]
prediction_video = prediction.max(dim=0, keepdim=True)[0]
target_video = target.max(dim=0, keepdim=True)[0]
prec1, = accuracy(prediction_video, target_video, topk=(1, ))
self.am.update(prec1.item())
def __repr__(self):
return 'VideoPrec@1 {am.val:.3f} ({am.avg:.3f})'.format(am=self.am)
def compute(self):
return ('videotop1', self.am.avg)
class Top1Metric(Metric):
def __init__(self):
self.am = AverageMeter()
def update(self, prediction, target):
if isinstance(prediction, dict):
prediction = prediction['class_prediction']
if isinstance(target, dict):
target = target['class_target']
if len(target) > 0:
prec1, = accuracy(prediction, target, topk=(1, ))
self.am.update(prec1.item())
def __repr__(self):
return 'Prec@1 {am.val:.3f} ({am.avg:.3f})'.format(am=self.am)
def compute(self):
return ('top1', self.am.avg)
def __init__(self):
self.am = AverageMeter()
self.predictions = []
self.targets = []
this_dir = os.path.dirname(__file__)
top60path = '../external/ActivityNet/Evaluation/ava/ava_action_list_v2.1_for_activitynet_2018.pbtxt.txt'
top60path = os.path.join(this_dir, top60path)
with open(top60path) as f:
self.top60 = [int(x) for x in re.findall('[0-9]+', f.read())]
class TripletAccuracyMetric(Metric):
def __init__(self):
self.am = AverageMeter()
def update(self, prediction, target):
if isinstance(prediction, dict):
prediction = prediction['triplet_prediction']
if isinstance(target, dict):
target = target['triplet_target']
if len(target) > 0:
prec1 = self.triplet_accuracy(prediction, target)
self.am.update(prec1.item())
def __repr__(self):
return 'triplet_accuracy_metric {am.val:.3f} ({am.avg:.3f})'.format(
am=self.am)
def compute(self):
return ('triplet_accuracy_metric', self.am.avg)
def triplet_accuracy(self, output, target, weights=None):
"""
if target>0 then first output should be smaller than right output
optional weighted average
"""
if type(output) is not list:
output = [(x.item(), y.item()) for x, y in zip(*output)]
correct = [
x < y if t > 0 else y < x for (x, y), t in zip(output, target)
]
if weights is None:
return np.mean(correct)
else:
weights = weights.numpy()
weights = weights / (1e-5 + np.sum(weights))
return np.sum(np.array(correct).astype(float) * weights)
def __init__(self):
self.am = AverageMeter()
def __init__(self):
self.am = AverageMeter()
self.predictions = []
self.targets = []