selectors

This module is aimed at How to select samples to provide more information for training.

method	description
BaseSelector	Base class.
DefaultSelector	Do nothing.
DenseTripletSelector	Select all triples.
DensePairSelector	Select all pairs.
DistanceWeightedSelector	Distance weighted selector
SemiHardSelector	Semi-hard selector
RandomTripletSelector	Randomly select triplets
HardSelector	Hardest hard sample

Class

BaseSelector

class gedml.core.selectors.base_selector.BaseSelector(**kwargs)

Bases: gedml.core.modules.with_recorder.WithRecorder

Base class of selectors.

forward(metric_mat, row_labels, col_labels, is_same_source=False) → tuple

Parameters

• metric_mat (torch.Tensor) – Metric matrix.

• row_labels (torch.Tensor) – Labels of rows.

• col_labels (torch.Tensor) – Labels of columns.

• is_same_source (bool) – Whether the two data streams are from the same source.

Returns

Five type of elements:

1. metric_mat (torch.Tensor): Metric matrix.

2. labels_row (torch.Tensor): Labels of rows.

3. labels_col (torch.Tensor): Labels of columns.

4. is_same_source (bool): Whether the two tensors are from the same source.

5. indices_tuple (dict): Dict that has two key: “tuples” and “flags”

6. weights (torch.Tensor): Weights.

Return type

tuple

DefaultSelector

class gedml.core.selectors.default_selector.DefaultSelector(**kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

Do nothing selector.

forward(metric_mat, row_labels, col_labels, is_same_source=False) → tuple

Do nothing.

DenseTripletSelector

class gedml.core.selectors.dense_triplet_selector.DenseTripletSelector(**kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

Select all triplets.

forward(metric_mat, row_labels, col_labels, is_same_source=False) → tuple

Select all triplets.

DensePairSelector

class gedml.core.selectors.dense_pair_selector.DensePairSelector(**kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

Select all pairs.

forward(metric_mat, row_labels, col_labels, is_same_source=False) → tuple

Select all pairs.

DistanceWeightedSelector

class gedml.core.selectors.distance_weighted_selector.DistanceWeightedSelector(lower_cutoff=0.5, upper_cutoff=1.4, embedding_dim=512, **kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

Distance weighted sampling method, euclidean distance metric is required.

paper: Sampling Matters in Deep Embedding Learning

forward(metric_mat, row_labels, col_labels, is_same_source=False) → tuple

Randomly select a positive sample for anchor sample and select a negative sample for anchor sample according to the distance weighted probability.

The distribution of pairwise distances follows:

\(q(d) \propto d^{n-2} [1 - \frac{1}{4} d^2 ]^{\frac{n-3}{2}}\)

SemiHardSelector

class gedml.core.selectors.semi_hard_selector.SemiHardSelector(margin=0.2, **kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

Semi-hard sampling method, euclidean distance metric is required.

forward(metric_mat, row_labels, col_labels, is_same_source=False)

Randomly select a positive sample and select a negative sample holds:

\(d_p < d_n < d_p + margin\)

RandomTripletSelector

class gedml.core.selectors.random_triplet_selector.RandomTripletSelector(**kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

Semi-hard sampling method, euclidean distance metric is required.

forward(metric_mat, row_labels, col_labels, is_same_source=False)

Randomly select a positive sample and a negative sample

HardSelector

class gedml.core.selectors.hard_selector.HardSelector(hardneg_cutoff=0.5, is_similarity=False, **kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

A self-defined selector according to distance weighted sampling method

forward(metric_mat, row_labels, col_labels, is_same_source=False) → tuple

Parameters

• metric_mat (torch.Tensor) – Metric matrix.

• row_labels (torch.Tensor) – Labels of rows.

• col_labels (torch.Tensor) – Labels of columns.

• is_same_source (bool) – Whether the two data streams are from the same source.

Returns

Five type of elements:

1. metric_mat (torch.Tensor): Metric matrix.

2. labels_row (torch.Tensor): Labels of rows.

3. labels_col (torch.Tensor): Labels of columns.

4. is_same_source (bool): Whether the two tensors are from the same source.

5. indices_tuple (dict): Dict that has two key: “tuples” and “flags”

6. weights (torch.Tensor): Weights.

Return type

tuple

HardPairSelector

class gedml.core.selectors.hard_pair_selector.HardPairSelector(hardneg_cutoff=0.5, is_similarity=False, **kwargs)

Bases: gedml.core.selectors.base_selector.BaseSelector

A self-defined selector according to distance weighted sampling method

forward(metric_mat, row_labels, col_labels, is_same_source=False) → tuple

Parameters

• metric_mat (torch.Tensor) – Metric matrix.

• row_labels (torch.Tensor) – Labels of rows.

• col_labels (torch.Tensor) – Labels of columns.

• is_same_source (bool) – Whether the two data streams are from the same source.

Returns

Five type of elements:

1. metric_mat (torch.Tensor): Metric matrix.

2. labels_row (torch.Tensor): Labels of rows.

3. labels_col (torch.Tensor): Labels of columns.

4. is_same_source (bool): Whether the two tensors are from the same source.

5. indices_tuple (dict): Dict that has two key: “tuples” and “flags”

6. weights (torch.Tensor): Weights.

Return type

tuple