Skip to content

Class swarmauri_embedding_mlm.MlmEmbedding.MlmEmbedding

swarmauri_embedding_mlm.MlmEmbedding.MlmEmbedding

MlmEmbedding(**kwargs)

Bases: EmbeddingBase

EmbeddingBase implementation that fine-tunes a Masked Language Model (MLM).

Source code in swarmauri_embedding_mlm/MlmEmbedding.py
33
34
35
36
37
38
39
40
41
def __init__(self, **kwargs):
    super().__init__(**kwargs)
    self._tokenizer = AutoTokenizer.from_pretrained(self.embedding_name)
    self._model = AutoModelForMaskedLM.from_pretrained(self.embedding_name)
    self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    self._model.to(self._device)
    self._mask_token_id = self._tokenizer.convert_tokens_to_ids(
        [self._tokenizer.mask_token]
    )[0]

embedding_name class-attribute instance-attribute

embedding_name = 'bert-base-uncased'

batch_size class-attribute instance-attribute

batch_size = 32

learning_rate class-attribute instance-attribute

learning_rate = 5e-05

masking_ratio class-attribute instance-attribute

masking_ratio = 0.15

randomness_ratio class-attribute instance-attribute

randomness_ratio = 0.1

epochs class-attribute instance-attribute

epochs = 0

add_new_tokens class-attribute instance-attribute

add_new_tokens = False

type class-attribute instance-attribute

type = 'MlmEmbedding'

model_config class-attribute instance-attribute

model_config = ConfigDict(
    extra="allow", arbitrary_types_allowed=True
)

id class-attribute instance-attribute

id = Field(default_factory=generate_id)

members class-attribute instance-attribute

members = None

owners class-attribute instance-attribute

owners = None

host class-attribute instance-attribute

host = None

default_logger class-attribute

default_logger = None

logger class-attribute instance-attribute

logger = None

name class-attribute instance-attribute

name = None

resource class-attribute instance-attribute

resource = Field(default=EMBEDDING.value, frozen=True)

version class-attribute instance-attribute

version = '0.1.0'

extract_features

extract_features()

Extracts the tokens from the vocabulary of the fine-tuned MLM.

Returns: - List[str]: A list of token strings in the model's vocabulary.

Source code in swarmauri_embedding_mlm/MlmEmbedding.py
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
def extract_features(self) -> List[str]:
    """
    Extracts the tokens from the vocabulary of the fine-tuned MLM.

    Returns:
    - List[str]: A list of token strings in the model's vocabulary.
    """
    # Get the vocabulary size
    vocab_size = len(self._tokenizer)

    # Retrieve the token strings for each id in the vocabulary
    token_strings = [
        self._tokenizer.convert_ids_to_tokens(i) for i in range(vocab_size)
    ]

    return token_strings

fit

fit(documents)
Source code in swarmauri_embedding_mlm/MlmEmbedding.py
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
def fit(self, documents: List[Union[str, Any]]):
    # Check if we need to add new tokens
    if self.add_new_tokens:
        new_tokens = self.find_new_tokens(documents)
        if new_tokens:
            num_added_toks = self._tokenizer.add_tokens(new_tokens)
            if num_added_toks > 0:
                logging.info(f"Added {num_added_toks} new tokens.")
                self.model.resize_token_embeddings(len(self._tokenizer))

    encodings = self._tokenizer(
        documents,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=512,
    )
    input_ids, attention_mask, labels = self._mask_tokens(encodings)
    optimizer = AdamW(self._model.parameters(), lr=self.learning_rate)
    dataset = TensorDataset(input_ids, attention_mask, labels)
    data_loader = DataLoader(dataset, batch_size=self.batch_size, shuffle=True)

    self._model.train()
    for batch in data_loader:
        batch = {
            k: v.to(self._device)
            for k, v in zip(["input_ids", "attention_mask", "labels"], batch)
        }
        outputs = self._model(**batch)
        loss = outputs.loss
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
    self.epochs += 1
    logging.info(f"Epoch {self.epochs} complete. Loss {loss.item()}")

find_new_tokens

find_new_tokens(documents)
Source code in swarmauri_embedding_mlm/MlmEmbedding.py
141
142
143
144
145
146
147
148
149
def find_new_tokens(self, documents):
    # Identify unique words in documents that are not in the tokenizer's vocabulary
    unique_words = set()
    for doc in documents:
        tokens = set(doc.split())  # Simple whitespace tokenization
        unique_words.update(tokens)
    existing_vocab = set(self._tokenizer.get_vocab().keys())
    new_tokens = list(unique_words - existing_vocab)
    return new_tokens if new_tokens else None

transform

transform(documents)

Generates embeddings for a list of documents using the fine-tuned MLM.

Source code in swarmauri_embedding_mlm/MlmEmbedding.py
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
def transform(self, documents: List[Union[str, Any]]) -> List[Vector]:
    """
    Generates embeddings for a list of documents using the fine-tuned MLM.
    """
    self._model.eval()
    embedding_list = []

    for document in documents:
        inputs = self._tokenizer(
            document,
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=512,
        )
        inputs = {k: v.to(self._device) for k, v in inputs.items()}
        with torch.no_grad():
            outputs = self._model(**inputs)
        # Extract embedding (for simplicity, averaging the last hidden states)
        if hasattr(outputs, "last_hidden_state"):
            embedding = outputs.last_hidden_state.mean(1)
        else:
            # Fallback or corrected attribute access
            embedding = outputs["logits"].mean(1)
        embedding = embedding.cpu().numpy()
        embedding_list.append(Vector(value=embedding.squeeze().tolist()))

    return embedding_list

fit_transform

fit_transform(documents, **kwargs)

Fine-tunes the MLM and generates embeddings for the provided documents.

Source code in swarmauri_embedding_mlm/MlmEmbedding.py
180
181
182
183
184
185
def fit_transform(self, documents: List[Union[str, Any]], **kwargs) -> List[Vector]:
    """
    Fine-tunes the MLM and generates embeddings for the provided documents.
    """
    self.fit(documents, **kwargs)
    return self.transform(documents)

infer_vector

infer_vector(data, *args, **kwargs)

Generates an embedding for the input data.

  • data (Union[str, Any]): The input data, expected to be a textual representation. Could be a single string or a batch of strings.
Source code in swarmauri_embedding_mlm/MlmEmbedding.py
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
def infer_vector(self, data: Union[str, Any], *args, **kwargs) -> Vector:
    """
    Generates an embedding for the input data.

    Parameters:
    - data (Union[str, Any]): The input data, expected to be a textual representation.
                              Could be a single string or a batch of strings.
    """
    # Tokenize the input data and ensure the tensors are on the correct device.
    self._model.eval()
    inputs = self._tokenizer(
        data, return_tensors="pt", padding=True, truncation=True, max_length=512
    )
    inputs = {k: v.to(self._device) for k, v in inputs.items()}

    # Generate embeddings using the model
    with torch.no_grad():
        outputs = self._model(**inputs)

    if hasattr(outputs, "last_hidden_state"):
        # Access the last layer and calculate the mean across all tokens (simple pooling)
        embedding = outputs.last_hidden_state.mean(dim=1)
    else:
        embedding = outputs["logits"].mean(1)
    # Move the embeddings back to CPU for compatibility with downstream tasks if necessary
    embedding = embedding.cpu().numpy()

    return Vector(value=embedding.squeeze().tolist())

save_model

save_model(path)

Saves the model and tokenizer to the specified directory.

Source code in swarmauri_embedding_mlm/MlmEmbedding.py
216
217
218
219
220
221
def save_model(self, path: str) -> None:
    """
    Saves the model and tokenizer to the specified directory.
    """
    self._model.save_pretrained(path)
    self._tokenizer.save_pretrained(path)

load_model

load_model(path)

Loads the model and tokenizer from the specified directory.

Source code in swarmauri_embedding_mlm/MlmEmbedding.py
223
224
225
226
227
228
229
def load_model(self, path: str) -> None:
    """
    Loads the model and tokenizer from the specified directory.
    """
    self._model = AutoModelForMaskedLM.from_pretrained(path)
    self._tokenizer = AutoTokenizer.from_pretrained(path)
    self._model.to(self._device)  # Ensure the model is loaded to the correct device

register_model classmethod

register_model()

Decorator to register a base model in the unified registry.

RETURNS DESCRIPTION
Callable

A decorator function that registers the model class.

TYPE: Callable[[Type[BaseModel]], Type[BaseModel]]

Source code in swarmauri_base/DynamicBase.py
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
@classmethod
def register_model(cls) -> Callable[[Type[BaseModel]], Type[BaseModel]]:
    """
    Decorator to register a base model in the unified registry.

    Returns:
        Callable: A decorator function that registers the model class.
    """

    def decorator(model_cls: Type[BaseModel]):
        """Register ``model_cls`` as a base model."""
        model_name = model_cls.__name__
        if model_name in cls._registry:
            glogger.warning(
                "Model '%s' is already registered; skipping duplicate.", model_name
            )
            return model_cls

        cls._registry[model_name] = {"model_cls": model_cls, "subtypes": {}}
        glogger.debug("Registered base model '%s'.", model_name)
        DynamicBase._recreate_models()
        return model_cls

    return decorator

register_type classmethod

register_type(resource_type=None, type_name=None)

Decorator to register a subtype under one or more base models in the unified registry.

PARAMETER DESCRIPTION
resource_type

The base model(s) under which to register the subtype. If None, all direct base classes (except DynamicBase) are used.

TYPE: Optional[Union[Type[T], List[Type[T]]]] DEFAULT: None

type_name

An optional custom type name for the subtype.

TYPE: Optional[str] DEFAULT: None

RETURNS DESCRIPTION
Callable

A decorator function that registers the subtype.

TYPE: Callable[[Type[DynamicBase]], Type[DynamicBase]]

Source code in swarmauri_base/DynamicBase.py
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
@classmethod
def register_type(
    cls,
    resource_type: Optional[Union[Type[T], List[Type[T]]]] = None,
    type_name: Optional[str] = None,
) -> Callable[[Type["DynamicBase"]], Type["DynamicBase"]]:
    """
    Decorator to register a subtype under one or more base models in the unified registry.

    Parameters:
        resource_type (Optional[Union[Type[T], List[Type[T]]]]):
            The base model(s) under which to register the subtype. If None, all direct base classes (except DynamicBase)
            are used.
        type_name (Optional[str]): An optional custom type name for the subtype.

    Returns:
        Callable: A decorator function that registers the subtype.
    """

    def decorator(subclass: Type["DynamicBase"]):
        """Register ``subclass`` as a subtype."""
        if resource_type is None:
            resource_types = [
                base for base in subclass.__bases__ if base is not cls
            ]
        elif not isinstance(resource_type, list):
            resource_types = [resource_type]
        else:
            resource_types = resource_type

        for rt in resource_types:
            if not issubclass(subclass, rt):
                raise TypeError(
                    f"'{subclass.__name__}' must be a subclass of '{rt.__name__}'."
                )
            final_type_name = type_name or getattr(
                subclass, "_type", subclass.__name__
            )
            base_model_name = rt.__name__

            if base_model_name not in cls._registry:
                cls._registry[base_model_name] = {"model_cls": rt, "subtypes": {}}
                glogger.debug(
                    "Created new registry entry for base model '%s'.",
                    base_model_name,
                )

            subtypes_dict = cls._registry[base_model_name]["subtypes"]
            if final_type_name in subtypes_dict:
                glogger.warning(
                    "Type '%s' already exists under '%s'; skipping duplicate.",
                    final_type_name,
                    base_model_name,
                )
                continue

            subtypes_dict[final_type_name] = subclass
            glogger.debug(
                "Registered '%s' as '%s' under '%s'.",
                subclass.__name__,
                final_type_name,
                base_model_name,
            )

        DynamicBase._recreate_models()
        return subclass

    return decorator

model_validate_toml classmethod

model_validate_toml(toml_data)

Validate a model from a TOML string.

Source code in swarmauri_base/TomlMixin.py
12
13
14
15
16
17
18
19
20
21
22
23
24
@classmethod
def model_validate_toml(cls, toml_data: str):
    """Validate a model from a TOML string."""
    try:
        # Parse TOML into a Python dictionary
        toml_content = tomllib.loads(toml_data)

        # Convert the dictionary to JSON and validate using Pydantic
        return cls.model_validate_json(json.dumps(toml_content))
    except tomllib.TOMLDecodeError as e:
        raise ValueError(f"Invalid TOML data: {e}")
    except ValidationError as e:
        raise ValueError(f"Validation failed: {e}")

model_dump_toml

model_dump_toml(
    fields_to_exclude=None, api_key_placeholder=None
)

Return a TOML representation of the model.

Source code in swarmauri_base/TomlMixin.py
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
def model_dump_toml(self, fields_to_exclude=None, api_key_placeholder=None):
    """Return a TOML representation of the model."""
    if fields_to_exclude is None:
        fields_to_exclude = []

    # Load the JSON string into a Python dictionary
    json_data = json.loads(self.model_dump_json())

    # Function to recursively remove specific keys and handle api_key placeholders
    def process_fields(data, fields_to_exclude):
        """Recursively filter fields and apply placeholders."""
        if isinstance(data, dict):
            return {
                key: (
                    api_key_placeholder
                    if key == "api_key" and api_key_placeholder is not None
                    else process_fields(value, fields_to_exclude)
                )
                for key, value in data.items()
                if key not in fields_to_exclude
            }
        elif isinstance(data, list):
            return [process_fields(item, fields_to_exclude) for item in data]
        else:
            return data

    # Filter the JSON data
    filtered_data = process_fields(json_data, fields_to_exclude)

    # Convert the filtered data into TOML
    return toml.dumps(filtered_data)

model_validate_yaml classmethod

model_validate_yaml(yaml_data)

Validate a model from a YAML string.

Source code in swarmauri_base/YamlMixin.py
11
12
13
14
15
16
17
18
19
20
21
22
23
@classmethod
def model_validate_yaml(cls, yaml_data: str):
    """Validate a model from a YAML string."""
    try:
        # Parse YAML into a Python dictionary
        yaml_content = yaml.safe_load(yaml_data)

        # Convert the dictionary to JSON and validate using Pydantic
        return cls.model_validate_json(json.dumps(yaml_content))
    except yaml.YAMLError as e:
        raise ValueError(f"Invalid YAML data: {e}")
    except ValidationError as e:
        raise ValueError(f"Validation failed: {e}")

model_dump_yaml

model_dump_yaml(
    fields_to_exclude=None, api_key_placeholder=None
)

Return a YAML representation of the model.

Source code in swarmauri_base/YamlMixin.py
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
def model_dump_yaml(self, fields_to_exclude=None, api_key_placeholder=None):
    """Return a YAML representation of the model."""
    if fields_to_exclude is None:
        fields_to_exclude = []

    # Load the JSON string into a Python dictionary
    json_data = json.loads(self.model_dump_json())

    # Function to recursively remove specific keys and handle api_key placeholders
    def process_fields(data, fields_to_exclude):
        """Recursively filter fields and apply placeholders."""
        if isinstance(data, dict):
            return {
                key: (
                    api_key_placeholder
                    if key == "api_key" and api_key_placeholder is not None
                    else process_fields(value, fields_to_exclude)
                )
                for key, value in data.items()
                if key not in fields_to_exclude
            }
        elif isinstance(data, list):
            return [process_fields(item, fields_to_exclude) for item in data]
        else:
            return data

    # Filter the JSON data
    filtered_data = process_fields(json_data, fields_to_exclude)

    # Convert the filtered data into YAML using safe mode
    return yaml.safe_dump(filtered_data, default_flow_style=False)

model_post_init

model_post_init(logger=None)

Assign a logger instance after model initialization.

Source code in swarmauri_base/LoggerMixin.py
23
24
25
26
27
28
def model_post_init(self, logger: Optional[FullUnion[LoggerBase]] = None) -> None:
    """Assign a logger instance after model initialization."""

    # Directly assign the provided FullUnion[LoggerBase] or fallback to the
    # class-level default.
    self.logger = self.logger or logger or self.default_logger