Embedding Models

taxotagger.models

ModelFactory

Factory class to get the embedding model for the given model identifier.

get_model staticmethod

get_model(
    model_id: str, config: ProjectConfig
) -> EmbedModelBase

Get the embedding model for the given model identifier.

Parameters:

• model_id (str) –

  The identifier of the model to load.

• config (ProjectConfig) –

  The configurations for the project.

Returns:

• EmbedModelBase –

  The embedding model instance for the given model identifier.

Examples:

>>> config = ProjectConfig()
>>> model = ModelFactory.get_model("MycoAI-CNN", config)

Source code in src/taxotagger/models.py

@staticmethod
def get_model(model_id: str, config: ProjectConfig) -> EmbedModelBase:
    """Get the embedding model for the given model identifier.

    Args:
        model_id: The identifier of the model to load.
        config: The configurations for the project.

    Returns:
        The embedding model instance for the given model identifier.

    Examples:
        >>> config = ProjectConfig()
        >>> model = ModelFactory.get_model("MycoAI-CNN", config)
    """
    if model_id == "MycoAI-CNN":
        return MycoAICNNEmbedModel(config)
    # Add more embedding models here if needed
    else:
        raise ValueError(
            f"Invalid model id {model_id}. Valid models are {PRETRAINED_MODELS.keys()}"
        )

MycoAICNNEmbedModel

MycoAICNNEmbedModel(config: ProjectConfig)

Bases: EmbedModelBase

Embedding model for the pretrained MycoAI-CNN.

Source code in src/taxotagger/models.py

def __init__(self, config: ProjectConfig) -> None:
    self._config = config
    self.model = load_model(self.name, config)

name class-attribute instance-attribute

name = 'MycoAI-CNN'

model instance-attribute

model = load_model(name, config)

embed

embed(fasta_file: str) -> dict[str, list[dict[str, Any]]]

Calculate the embeddings for the given FASTA file.

Parameters:

• fasta_file (str) –

  The path to the FASTA file to embed.

Returns:

• dict[str, list[dict[str, Any]]] –

  A dictionary of embeddings for each taxonomy level. The dictionary keys are the taxonomy levels, and the values are lists of dictionaries containing the id, embeddings and metadata for each sequence.

  The shape of the list is (n_samples), where n_samples is the number of sequences.

  The keys of the inside dictionaries are: id, vector, and the taxonomy levels (e.g. phylum, class, order, family, genus, species) and other metadata fields present in the FASTA header.

  The shape of the vector is (n_features), where n_features is the number of features in the embedding. The number of features for each taxonomy level is:

  - phylum: 18
  - class: 70
  - order: 231
  - family: 791
  - genus: 3695
  - species: 14742
  

  The returned data looks like:

  {
  "phylum": [{"id": "seq1", "vector": [0.1, 0.2, ...], "phylum": "Basidiomycota", ...}, ...],
  "class": [{"id": "seq1", "vector": [0.5, 0.6, ...], "class": "Agaricomycetes", ...}, ...],
  "order": [{"id": "seq1", "vector": [0.9, 0.8, ...], "order": "Corticiales", ...}, ...],
  "family": [{"id": "seq1", "vector": [0.3, 0.4, ...], "family": "Corticiaceae", ...}, ...],
  "genus": [{"id": "seq1", "vector": [0.7, 0.8, ...], "genus": "Waitea", ...}, ...],
  "species": [{"id": "seq1", "vector": [0.5, 0.6, ...], "species": "Circinata", ...}, ...]
  }
  

Examples:

>>> config = ProjectConfig()
>>> model = MycoAICNNEmbedModel(config)
>>> embeddings = model.embed("dna1.fasta")

Source code in src/taxotagger/models.py

def embed(self, fasta_file: str) -> dict[str, list[dict[str, Any]]]:
    """Calculate the embeddings for the given FASTA file.

    Args:
        fasta_file: The path to the FASTA file to embed.

    Returns:
        A dictionary of embeddings for each taxonomy level.
            The dictionary keys are the taxonomy levels, and the values are lists of dictionaries
            containing the id, embeddings and metadata for each sequence.

            The shape of the list is `(n_samples)`, where `n_samples` is the number of sequences.

            The keys of the inside dictionaries are: `id`, `vector`, and the taxonomy levels
            (e.g. `phylum`, `class`, `order`, `family`, `genus`, `species`) and other metadata
            fields present in the FASTA header.

            The shape of the `vector` is `(n_features)`, where `n_features` is the number of
            features in the embedding. The number of features for each taxonomy level is:

                - phylum: 18
                - class: 70
                - order: 231
                - family: 791
                - genus: 3695
                - species: 14742

            The returned data looks like:
            ```python
            {
            "phylum": [{"id": "seq1", "vector": [0.1, 0.2, ...], "phylum": "Basidiomycota", ...}, ...],
            "class": [{"id": "seq1", "vector": [0.5, 0.6, ...], "class": "Agaricomycetes", ...}, ...],
            "order": [{"id": "seq1", "vector": [0.9, 0.8, ...], "order": "Corticiales", ...}, ...],
            "family": [{"id": "seq1", "vector": [0.3, 0.4, ...], "family": "Corticiaceae", ...}, ...],
            "genus": [{"id": "seq1", "vector": [0.7, 0.8, ...], "genus": "Waitea", ...}, ...],
            "species": [{"id": "seq1", "vector": [0.5, 0.6, ...], "species": "Circinata", ...}, ...]
            }
            ```

    Examples:
        >>> config = ProjectConfig()
        >>> model = MycoAICNNEmbedModel(config)
        >>> embeddings = model.embed("dna1.fasta")
    """
    headers, encoded_data = self.parse_and_encode_fasta(fasta_file)
    # headers shape (n_samples, n_headers), e.g.
    # [['id1', 'phylum1', 'class1', 'order1', 'family1', 'genus1', 'species1', 'SH_id1'], ...]

    embeddings = []
    dataloader = DataLoader(encoded_data, shuffle=False)
    with torch.no_grad():
        for x, _ in dataloader:  # (encoded data, labels)
            y_pred = self.model(x.to(self._config.device))
            embeddings.append(y_pred)
    # embeddings shape (n_samples, n_taxonomies, (1, n_features)), where n_taxonomies is 6, e.g.
    # [[phylumTensor1, classTensor1, orderTensor1, familyTensor1, genusTensor1, speciesTensor1], ...]
    # the shape of each tensor is (1, n_features), n_features are different for each taxonomy level

    data_collections = {}
    for i, taxo_level in enumerate(TAXONOMY_LEVELS):
        data_list = [
            {
                "id": headers[j][0],
                # squeeze to remove the batch dimension: (1, n_features) -> (n_features)
                "vector": embeddings[j][i].squeeze().numpy(),
                taxo_level: headers[j][i + 2],
                "SH_id": headers[j][-1],
            }
            for j in range(len(embeddings))
        ]
        data_collections[taxo_level] = data_list
    return data_collections

parse_and_encode_fasta

parse_and_encode_fasta(
    fasta_file: str,
) -> tuple[list[list[str]], TensorData]

Parse headers and encode the sequences in the given FASTA file.

The sequences are encoded using the encoders defined in the pretrained model.

Parameters:

• fasta_file (str) –

  The path to the FASTA file.

Returns:

• tuple[list[list[str]], TensorData] –

  A tuple containing the headers and the encoded data for the sequences in the FASTA file.

  The shape of the headers is (n_samples, n_headers), where n_samples is the number of sequences and n_headers is the 9 metadata fields parsed from the header.

Source code in src/taxotagger/models.py

def parse_and_encode_fasta(self, fasta_file: str) -> tuple[list[list[str]], data.TensorData]:
    """Parse headers and encode the sequences in the given FASTA file.

    The sequences are encoded using the encoders defined in the pretrained model.

    Args:
        fasta_file: The path to the FASTA file.

    Returns:
        A tuple containing the headers and the encoded data for the sequences in the FASTA file.

            The shape of the headers is `(n_samples, n_headers)`, where `n_samples` is the
            number of sequences and `n_headers` is the 9 metadata fields parsed from the header.
    """
    input_data = data.Data(fasta_file, tax_parser=None, allow_duplicates=False)
    # Using custom parser to parse the FASTA headers
    headers = [parse_unite_fasta_header(header) for header in input_data.data["id"].values]
    encoded_data = input_data.encode_dataset(self.model.dna_encoder, self.model.tax_encoder)
    return headers, encoded_data