secmlt.models package

Subpackages

• secmlt.models.data_processing package
  • Submodules
  • secmlt.models.data_processing.data_processing module
    • DataProcessing
      • DataProcessing.invert()
  • secmlt.models.data_processing.identity_data_processing module
    • IdentityDataProcessing
      • IdentityDataProcessing.invert()
  • Module contents
• secmlt.models.hugging_face package
  • Submodules
  • secmlt.models.hugging_face.base_hf_lm module
    • HFCausalLM
      • HFCausalLM.decode()
      • HFCausalLM.encode()
      • HFCausalLM.generate()
      • HFCausalLM.hidden_states()
      • HFCausalLM.load()
      • HFCausalLM.logprobs()
      • HFCausalLM.model
      • HFCausalLM.predict()
      • HFCausalLM.tokenizer
  • Module contents
• secmlt.models.pytorch package
  • Submodules
  • secmlt.models.pytorch.base_pytorch_nn module
    • BasePyTorchClassifier
      • BasePyTorchClassifier.model
      • BasePyTorchClassifier.predict()
      • BasePyTorchClassifier.train()
  • secmlt.models.pytorch.base_pytorch_trainer module
    • BasePyTorchTrainer
      • BasePyTorchTrainer.train()
  • Module contents

Submodules

secmlt.models.base_language_model module

Basic wrapper for generic language model.

class secmlt.models.base_language_model.BaseLanguageModel

Bases: ABC

Abstract base class defining the common interface for language models.

abstract decode(ids: LongTensor, **kwargs) → str | list[str]

Convert token IDs back into text.

Parameters:

ids (torch.LongTensor) – Token IDs tensor.

Returns:

Decoded text(s).

Return type:

str or list of str

abstract encode(text: str | list[str], **kwargs) → LongTensor

Convert input text into token IDs.

Parameters:

text (str or list of str) – Input text(s) to tokenize.

Returns:

Token IDs tensor.

Return type:

torch.LongTensor

abstract generate(prompts: list[str], **kwargs) → list[str]

Generate text continuations from given prompts.

Parameters:

prompts (list of str) – List of prompt strings.

Returns:

Generated text outputs.

Return type:

list of str

abstract hidden_states(input_ids: LongTensor, **kwargs) → list[Tensor]

Return hidden states of the model for given input.

Parameters:

input_ids (torch.LongTensor) – Input token IDs.

Returns:

Hidden representations per layer.

Return type:

list of torch.Tensor

abstract logprobs(prompts: list[str], targets: list[str], **kwargs) → list[Tensor]

Compute log-probabilities for each token in the target continuations.

Parameters:

• prompts (list of str) – Conditioning prompts.

• targets (list of str) – Target continuations.

Returns:

List of log-probability tensors, one per sample in the batch. Each tensor has shape [target_len_i].

Return type:

list of torch.Tensor

abstract predict(input_ids: LongTensor, **kwargs) → Tensor

Predict logits for the next token given an input sequence.

Parameters:

input_ids (torch.LongTensor) – Input token IDs.

Returns:

Logits for the next token (shape [batch, vocab_size]).

Return type:

torch.Tensor

secmlt.models.base_model module

Basic wrapper for generic model.

class secmlt.models.base_model.BaseModel(preprocessing: DataProcessing | None = None, postprocessing: DataProcessing | None = None)

Bases: ABC

Basic model wrapper.

decision_function(x: Tensor, *args, **kwargs) → Tensor

Return the decision function from the model.

Requires override to specify custom args and kwargs passing.

Parameters:

x (torch.Tensor) – Input damples.

Returns:

Model output scores.

Return type:

torch.Tensor

abstract predict(x: Tensor, *args, **kwargs) → Tensor

Return output predictions for given model.

Parameters:

x (torch.Tensor) – Input samples.

Returns:

Predictions from the model.

Return type:

torch.Tensor

abstract train(dataloader: DataLoader) → BaseModel

Train the model with the given dataloader.

Parameters:

dataloader (DataLoader) – Train data loader.

secmlt.models.base_trainer module

Model trainers.

class secmlt.models.base_trainer.BaseTrainer

Bases: object

Abstract class for model trainers.

abstract train(model: BaseModel, dataloader: DataLoader) → BaseModel

Train a model with the given dataloader.

Parameters:

• model (BaseModel) – Model to train.

• dataloader (DataLoader) – Training dataloader.

Returns:

The trained model.

Return type:

BaseModel

Module contents

Machine learning models and wrappers.