The Prompt Guided Neural Architecture Search (PG-NAS) is a PyTorch-based framework designed to generate neural network architectures based on textual prompts. It utilizes a prompt encoder to interpret the input prompt and proposes a corresponding architecture with shape tracking. The model supports various layer types, including convolutional, linear, and attention layers.
- Prompt-based Architecture Generation: Generate neural network architectures from natural language prompts.
- Layer Type Support: Supports convolutional, linear, and attention layers.
- Shape Tracking: Keeps track of input and output shapes throughout the architecture.
- Weight Initialization: Automatically initializes weights for each layer based on the proposed architecture.
- Model Saving and Loading: Save and load model artifacts, including architecture and configuration.
- Python 3.6+
- PyTorch
- Transformers
- Loguru
- Other dependencies as specified in the code