- Malignant Classes:
MEL, BCC, AKIEC
- Benign Classes:
NV, BKL, DF
- Benign / Malignant:
VASC
-
L-way n-shot: L classes are randomly sampled from support domain into the support set. For each class, n images are chosen and placed into the support set. A different parameter, m determines the number of images per class that will be chosen for the query set.
-
Prototypical Networks: During training, episodes consisting of a support set S and a query set Q are sampled as described earlier. Then, a prototype for each class k is computed as the mean embedding of the samples
from the support set belonging to that class.
-
Reference: Episodic Learning
- All classes are part of both, train and test sets
- Split randomly using
src/split_train_test in 70-30 ratio (ratio is per-class)
- Use the function
split_data_all_classes()
TRAINING CONFIGURATION
Class Names: ([ MEL, NV, BCC, AKIEC, BKL, DF, VASC ])
Class Distribution: tensor([ 835, 5029, 386, 246, 825, 87, 107])
TESTING CONFIGURATION
Class Names: ([ MEL, NV, BCC, AKIEC, BKL, DF, VASC ])
Class Distribution: ([ 278, 1676, 128, 81, 274, 28, 35])
Support Domain: tensor([0, 1, 2, 3, 4, 5, 6])
Forced Support: tensor([])
Query Domain: tensor([0, 1, 2, 3, 4, 5, 6])
- Most mis-classified classes are removed from the train set, and moved to the test set
- Specifically,
MEL, NV and BCC are made the test set, and the model is trained only using the other 4 classes.
- Split deterministically using
src/split_train_test in 70-30 ratio (ratio is per-class)
- Use the function
split_test_classes()
- This tests the strength of the similarity function learnt by the model
TRAINING CONFIGURATION
Class Names: ([ AKIEC, BKL, DF, VASC ])
Class Distribution: tensor([ 327, 1099, 115, 142])
TESTING CONFIGURATION
Class Names: ([ MEL, NV, BCC ])
Class Distribution: ([1113, 6705, 514])
Support Domain: tensor([0, 1, 2])
Forced Support: tensor([])
Query Domain: tensor([0, 1, 2])
- Complete data is split in 70-30 ratio (per-class ratio) into train and test sets
- Most mis-classified classes are removed from the train set, and moved to isolation (these are never seen by the trainer)
- Specifically,
MEL, NV and BCC are moved to isolation set from test set, and the model is trained only using the other 4 classes
- This tests the strength of the similarity function learnt by the model
- Split randomly using
src/split_train_test in 70-30 ratio (ratio is per-class)
- Use the functions
split_data_all_classes() and split_test_classes(), in succession
TRAINING CONFIGURATION
Class Names: ([ AKIEC, BKL, DF, VASC ])
Class Distribution: tensor([229, 770, 81, 100])
TESTING CONFIGURATION
Class Names: ([ MEL, NV, BCC, AKIEC, BKL, DF, VASC ])
Class Distribution: ([ 333, 2011, 154, 98, 329, 34, 42])
Support Domain: tensor([0, 1, 2, 3, 4, 5, 6])
Forced Support: tensor([])
Query Domain: tensor([0, 1, 2])
- 2-way, 3-shot testing
- The test set contains all 7 classes
- When sampling, query set is populated only with one of
MEL, NV and BCC. Support set, however, can contain any one of the seven classes
TRAINING CONFIGURATION
Class Names: ([ AKIEC, BKL, DF, VASC ])
Class Distribution: tensor([229, 770, 81, 100])
TESTING CONFIGURATION
Class Names: ([ MEL, NV, BCC, AKIEC, BKL, DF, VASC ])
Class Distribution: ([ 333, 2011, 154, 98, 329, 34, 42])
Support Domain: tensor([0, 1, 2, 3, 4, 5, 6])
Forced Support: tensor([])
Query Domain: tensor([0, 1, 2])
- 3-way, 3-shot testing
- The test set contains all 7 classes
- When sampling, query set is populated only with one of
MEL, NV and BCC. Support set, however, can contain any one of the seven classes
TRAINING CONFIGURATION
Class Names: ([ AKIEC, BKL, DF, VASC ])
Class Distribution: tensor([229, 770, 81, 100])
TESTING CONFIGURATION
Class Names: ([ MEL, NV, BCC, AKIEC, BKL, DF, VASC ])
Class Distribution: ([ 333, 2011, 154, 98, 329, 34, 42])
Support Domain: tensor([0, 1, 2, 3, 4, 5, 6])
Forced Support: tensor([ 1 ])
Query Domain: tensor([0, 1, 2])
- 3-way, 3-shot testing
- The test set contains all 7 classes
- When sampling, query set is populated only with one of
MEL, NV and BCC. Support set, however, can contain any one of the seven classes
- As an additional constraint, to study the impact of misclassification induced by
NV, it is always included in the support set
- Complete data is split in 70-30 ratio (per-class ratio) into train and test sets
- Classes with the least number of images-per-class are removed from the train set, and moved to isolation (these are never seen by the trainer)
- Specifically,
AKIEC, VASC and DF are moved to isolation set from test set, and the model is trained only using the other 4 classes
- This tests the ability of the model to learn the "similarity" function from a large dataset, and successfully use it to distinguish between classes that have limited annotated data available
- Split randomly using
src/split_train_test in 70-30 ratio (ratio is per-class)
- Use the functions
split_data_all_classes() and split_test_classes(), in succession
TRAINING CONFIGURATION
Class Names: ([ MEL, NV, BCC, BKL ])
Class Distribution: tensor([ 780, 4694, 360, 770])
TESTING CONFIGURATION
Class Names: ([ MEL, NV, BCC, AKIEC, BKL, DF, VASC ])
Class Distribution: ([ 333, 2011, 154, 98, 329, 34, 42])
Support Domain: tensor([0, 1, 2, 3, 4, 5, 6])
Forced Support: tensor([ 1 ])
Query Domain: tensor([3, 5, 6])
- 2-way, 3-shot testing
- The test set contains all 7 classes
- When sampling, query set is populated only with one of
MEL, NV and BCC. Support set, however, can contain any one of the seven classes
- As an additional constraint, to study the impact of misclassification induced by
NV, it is always included in the support set
TRAINING CONFIGURATION
Class Names: ([ MEL, NV, BCC, BKL ])
Class Distribution: tensor([ 780, 4694, 360, 770])
TESTING CONFIGURATION
Class Names: ([ MEL, NV, BCC, AKIEC, BKL, DF, VASC ])
Class Distribution: ([ 333, 2011, 154, 98, 329, 34, 42])
Support Domain: tensor([0, 1, 2, 3, 4, 5, 6])
Forced Support: tensor([ 1 ])
Query Domain: tensor([3, 5, 6])
- 3-way, 3-shot testing
- The test set contains all 7 classes
- When sampling, query set is populated only with one of
MEL, NV and BCC. Support set, however, can contain any one of the seven classes
- As an additional constraint, to study the impact of misclassification induced by
NV, it is always included in the support set
