PyTorch Sequence Fusion — LSTM + Transformer Attention Toolkit

Build and experiment with sequence models using a configurable PyTorch architecture that combines LSTM, Transformer self-attention, learned positional encoding, flexible pooling, and an MLP output head.

Sequence Fusion is a source-code toolkit for Python developers and machine learning practitioners who want a clear starting point for sequence-modeling experiments and custom ML projects.

WHAT'S INCLUDED

• Installable Python package — reusable model components with a clean import path.

• Editable JSON configuration — customize dimensions, LSTM direction, dropout, attention heads, pooling, and classifier layers.

• Quick-start example — see how to initialize the model, prepare input tensors, pass Boolean masks, and run a forward pass.

• Synthetic training demo — explore a basic training loop on generated toy data.

• Smoke tests — check supported pooling modes, masking, dimension validation, and bidirectional operation.

• Bilingual documentation — installation, configuration, examples, and troubleshooting in English and Russian.

• License file — usage terms for integrating the toolkit into personal or commercial projects.

MODEL ARCHITECTURE

1. Learned positional encoding adds trainable position information to the sequence.

2. LSTM processes sequential features and supports configurable bidirectional operation.

3. Transformer encoder applies multi-head self-attention and a feed-forward block.

4. Flexible pooling supports attention, max, mean, or combined max + mean pooling.

5. MLP output head produces task-specific raw logits.

POTENTIAL USE CASES

Adapt the architecture for experiments involving:

• Sequence classification

• Time-series feature modeling

• Sensor and signal sequences

• Custom machine learning research and prototyping

Suitability and predictive quality depend on your data, preprocessing, training procedure, and evaluation. The toolkit is a starting point for development, not a validated production model.

REQUIREMENTS

• Python 3.10 or newer

• PyTorch 2.0 or newer

• CPU is sufficient for the included examples

• A compatible GPU is optional for larger experiments

QUICK START

1. Download and extract the ZIP archive.

2. Open a terminal in the Sequence_Fusion_PyTorch_Toolkit folder.

3. Create and activate a virtual environment.

4. Install the dependencies and package.

5. Run the example scripts.

Windows PowerShell

py -3 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt
pip install -e .
python examples\basic_usage.py

macOS / Linux

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements.txt
pip install -e .
python examples/basic_usage.py

To run the training demonstration:

python examples/train_toy.py

The basic example uses input shaped (4, 24, 8) and returns output logits shaped (4, 3).

IMPORTANT — PLEASE READ

This download does not include pretrained weights. The model starts with randomly initialized parameters. You must train and evaluate it on data appropriate for your use case before interpreting its predictions.

Also not included:

• A real-world dataset or validated benchmark results

• A ready-made forecasting or trading strategy

• Guaranteed prediction accuracy or financial returns

• Hosted inference, managed services, or ongoing support

LICENSE

The included license permits one individual or one organization to use, modify, and integrate the source code into personal or commercial projects. Reselling or redistributing the raw toolkit as a standalone source-code product is prohibited. Third-party dependencies remain subject to their own licenses.

Support and future updates are not included unless explicitly stated in the listing.

DIGITAL DOWNLOAD

You will receive a downloadable ZIP archive containing the Python source code, configuration, examples, tests, and documentation. No physical item will be shipped.