Testers - PyTorch Metric Learning (original) (raw)

Testers take your model and dataset, and compute nearest-neighbor based accuracy metrics. Note that the testers require the faiss package, which you can install with conda.

In general, testers are used as follows:

`from pytorch_metric_learning import testers t = testers.SomeTestingFunction(*args, **kwargs) dataset_dict = {"train": train_dataset, "val": val_dataset} all_accuracies = tester.test(dataset_dict, epoch, model)

Or if your model is composed of a trunk + embedder

all_accuracies = tester.test(dataset_dict, epoch, trunk, embedder) `

You can perform custom actions by writing an end-of-testing hook (see the documentation for BaseTester), and you can access the test results directly via the all_accuracies attribute:

def end_of_testing_hook(tester): print(tester.all_accuracies)

This will print out a dictionary of accuracy metrics, per dataset split. You'll see something like this:

{"train": {"AMI_level0": 0.53, ...}, "val": {"AMI_level0": 0.44, ...}}

Each of the accuracy metric names is appended with level0, which refers to the 0th label hierarchy level (see the documentation for BaseTester). This is only relevant if you're dealing with multi-label datasets.

For an explanation of the default accuracy metrics, see the AccuracyCalculator documentation.

Testing splits¶

By default, every dataset in dataset_dict will be evaluated using itself as the query and reference (on which to find nearest neighbors). More flexibility is allowed with the optional argument splits_to_eval taken by tester.test().splits_to_eval is a list of (query_split, [list_of_reference_splits]) tuples.

For example, let's say your dataset_dict has two keys: "dataset_a" and "train".

• The default splits_to_eval = None is equivalent to:
  splits_to_eval = [('dataset_a', ['dataset_a']), ('train', ['train'])]
• dataset_a as the query, and train as the reference:
  splits_to_eval = [('dataset_a', ['train'])]
• dataset_a as the query, and dataset_a + train as the reference:
  splits_to_eval = [('dataset_a', ['dataset_a', 'train'])]

BaseTester¶

All trainers extend this class and therefore inherit its __init__ arguments.

testers.BaseTester(normalize_embeddings=True, use_trunk_output=False, batch_size=32, dataloader_num_workers=2, pca=None, data_device=None, dtype=None, data_and_label_getter=None, label_hierarchy_level=0, end_of_testing_hook=None, dataset_labels=None, set_min_label_to_zero=False, accuracy_calculator=None, visualizer=None, visualizer_hook=None,)

Parameters:

• normalize_embeddings: If True, embeddings will be normalized to Euclidean norm of 1 before nearest neighbors are computed.
• use_trunk_output: If True, the output of the trunk_model will be used to compute nearest neighbors, i.e. the output of the embedder model will be ignored.
• batch_size: How many dataset samples to process at each iteration when computing embeddings.
• dataloader_num_workers: How many processes the dataloader will use.
• pca: The number of dimensions that your embeddings will be reduced to, using PCA. The default is None, meaning PCA will not be applied.
• data_device: Which gpu to use for the loaded dataset samples. If None, then the gpu or cpu will be used (whichever is available).
• dtype: The type that the dataset output will be converted to, e.g. torch.float16. If set to None, then no type casting will be done.
• data_and_label_getter: A function that takes the output of your dataset's __getitem__ function, and returns a tuple of (data, labels). If None, then it is assumed that __getitem__ returns (data, labels).
• label_hierarchy_level: If each sample in your dataset has multiple labels, then this integer argument can be used to select which "level" to use. This assumes that your labels are "2-dimensional" with shape (num_samples, num_hierarchy_levels). Leave this at the default value, 0, if your data does not have multiple labels per sample.
• end_of_testing_hook: This is an optional function that has one input argument (the tester object) and performs some action (e.g. logging data) at the end of testing.
  • You'll probably want to access the accuracy metrics, which are stored in tester.all_accuracies. This is a nested dictionary with the following format: tester.all_accuracies[split_name][metric_name] = metric_value
  • If you want ready-to-use hooks, take a look at the logging_presets module.
• dataset_labels: The labels for your dataset. Can be 1-dimensional (1 label per datapoint) or 2-dimensional, where each row represents a datapoint, and the columns are the multiple labels that the datapoint has. Labels can be integers or strings. This option needs to be specified only if set_min_label_to_zero is True.
• set_min_label_to_zero: If True, labels will be mapped such that they represent their rank in the label set. For example, if your dataset has labels 5, 10, 12, 13, then at each iteration, these would become 0, 1, 2, 3. You should also set this to True if you want to use string labels. In that case, 'dog', 'cat', 'monkey' would get mapped to 1, 0, 2. If True, you must pass in dataset_labels (see above). The default is False.
• accuracy_calculator: Optional. An object that extends AccuracyCalculator. This will be used to compute the accuracy of your model. By default, AccuracyCalculator is used.
• visualizer: Optional. An object that has implemented the fit_transform method, as done by UMAP and many scikit-learn functions. For example, you can set visualizer = umap.UMAP(). The object's fit_transform function should take in a 2D array of embeddings, and reduce the dimensionality, such that calling visualizer.fit_transform(embeddings) results in a 2D array of size (N, 2).
• visualizer_hook: Optional. This function will be passed the following args. You can do whatever you want in this function, but the reason it exists is to allow you to save a plot of the embeddings etc.
  • visualizer: The visualizer object that you passed in.
  • embeddings: The dimensionality reduced embeddings.
  • label: The corresponding labels for each embedding.
  • split_name: The name of the split (train, val, etc.)
  • keyname: The name of the dictionary key where the embeddings and labels are stored.
  • epoch: The epoch for which the embeddings are being computed.

Functions:

• tester.test

Call this to test your model on a dataset dict. It returns a dictionary of accuracies.

all_accuracies = tester.test( dataset_dict, # dictionary mapping strings to datasets epoch, # used for logging trunk_model, # your model embedder_model=None, # by default this will be a no-op splits_to_eval=None, collate_fn=None # custom collate_fn for the dataloader )

• tester.get_all_embeddings

Returns all the embeddings and labels for the input dataset and model.

embeddings, labels = tester.get_all_embeddings( dataset, # Any pytorch dataset trunk_model, # your model embedder_model=None, # by default this will be a no-op collate_fn=None, # custom collate_fn for the dataloader eval=True, # set models to eval mode return_as_numpy=False )

GlobalEmbeddingSpaceTester¶

Computes nearest neighbors by looking at all points in the embedding space (rather than a subset). This is probably the tester you are looking for. To see it in action, check one of the example notebooks

testers.GlobalEmbeddingSpaceTester(*args, **kwargs)

WithSameParentLabelTester¶

This assumes there is a label hierarchy. For each sample, the search space is narrowed by only looking at sibling samples, i.e. samples with the same parent label. For example, consider a dataset with 4 fine-grained classes {cat, dog, car, truck}, and 2 coarse-grained classes {animal, vehicle}. The nearest neighbor search for cats and dogs will consist of animals, and the nearest-neighbor search for cars and trucks will consist of vehicles.

testers.WithSameParentLabelTester(*args, **kwargs)

GlobalTwoStreamEmbeddingSpaceTester¶

This is the corresponding tester for TwoStreamMetricLoss. The supplied dataset must return (anchor, positive, label).

testers.GlobalTwoStreamEmbeddingSpaceTester(*args, **kwargs)

Requirements:

This tester only supports the default value for splits_to_eval: each split is used for both query and reference