diff options
Diffstat (limited to 'utils')
| -rw-r--r-- | utils/configuration.py | 1 | ||||
| -rw-r--r-- | utils/sampler.py | 35 | ||||
| -rw-r--r-- | utils/triplet_loss.py | 9 |
3 files changed, 39 insertions, 6 deletions
diff --git a/utils/configuration.py b/utils/configuration.py index 8ee08f2..8dcae07 100644 --- a/utils/configuration.py +++ b/utils/configuration.py @@ -8,6 +8,7 @@ class SystemConfiguration(TypedDict): CUDA_VISIBLE_DEVICES: str save_dir: str image_log_on: bool + val_size: int class DatasetConfiguration(TypedDict): diff --git a/utils/sampler.py b/utils/sampler.py index 0977f94..581d7a2 100644 --- a/utils/sampler.py +++ b/utils/sampler.py @@ -15,7 +15,18 @@ class TripletSampler(data.Sampler): ): super().__init__(data_source) self.metadata_labels = data_source.metadata['labels'] + metadata_conditions = data_source.metadata['conditions'] + self.subsets = {} + for condition in metadata_conditions: + pre, _ = condition.split('-') + if self.subsets.get(pre, None) is None: + self.subsets[pre] = [] + self.subsets[pre].append(condition) + self.num_subsets = len(self.subsets) + self.num_seq = {pre: len(seq) for (pre, seq) in self.subsets.items()} + self.min_num_seq = min(self.num_seq.values()) self.labels = data_source.labels + self.conditions = data_source.conditions self.length = len(self.labels) self.indexes = np.arange(0, self.length) (self.pr, self.k) = batch_size @@ -26,15 +37,31 @@ class TripletSampler(data.Sampler): # Sample pr subjects by sampling labels appeared in dataset sampled_subjects = random.sample(self.metadata_labels, k=self.pr) for label in sampled_subjects: - clips_from_subject = self.indexes[self.labels == label].tolist() + mask = self.labels == label + # Fix unbalanced datasets + if self.num_subsets > 1: + condition_mask = np.zeros(self.conditions.shape, dtype=bool) + for num, conditions_ in zip( + self.num_seq.values(), self.subsets.values() + ): + if num > self.min_num_seq: + conditions = random.sample( + conditions_, self.min_num_seq + ) + else: + conditions = conditions_ + for condition in conditions: + condition_mask |= self.conditions == condition + mask &= condition_mask + clips = self.indexes[mask].tolist() # Sample k clips from the subject without replacement if # have enough clips, k more clips will sampled for # disentanglement k = self.k * 2 - if len(clips_from_subject) >= k: - _sampled_indexes = random.sample(clips_from_subject, k=k) + if len(clips) >= k: + _sampled_indexes = random.sample(clips, k=k) else: - _sampled_indexes = random.choices(clips_from_subject, k=k) + _sampled_indexes = random.choices(clips, k=k) sampled_indexes += _sampled_indexes yield sampled_indexes diff --git a/utils/triplet_loss.py b/utils/triplet_loss.py index 03fff21..5e3a97a 100644 --- a/utils/triplet_loss.py +++ b/utils/triplet_loss.py @@ -28,6 +28,7 @@ class BatchTripletLoss(nn.Module): else: # is_all positive_negative_dist = self._all_distance(dist, y, p, n) + non_zero_counts = None if self.margin: losses = F.relu(self.margin + positive_negative_dist).view(p, -1) non_zero_counts = (losses != 0).sum(1).float() @@ -35,14 +36,18 @@ class BatchTripletLoss(nn.Module): loss_metric = self._none_zero_mean(losses, non_zero_counts) else: # is_sum loss_metric = losses.sum(1) - return loss_metric, flat_dist, non_zero_counts else: # Soft margin losses = F.softplus(positive_negative_dist).view(p, -1) if self.is_mean: loss_metric = losses.mean(1) else: # is_sum loss_metric = losses.sum(1) - return loss_metric, flat_dist, None + + return { + 'loss': loss_metric, + 'dist': flat_dist, + 'counts': non_zero_counts + } @staticmethod def _batch_distance(x): |