migration

[jamilgafur]

1. Adversarial Attacks (Attacks.py)

• Signature & Logic Simplifications:
  • gradient_map: Replaced complex backpropagation options (guided, relu) with a simpler implementation utilizing a new compute_gradients helper function.
  • gradient_ascent: Significantly altered the signature. It now maximizes a target_class on a single input image rather than maximizing a specific target_neuron across a batch of data.
  • fgsm_attack: Modified to return a list of tensors rather than a stacked tensor array (torch.stack).
• Code Cleanup: Removed several unused utility functions, including visualize_adversarial_examples, log_metrics, saliency_map, and the Config class.

---

2. Particle Logic (BirdParticle.py)

• Model Decoupling: The BirdParticle initialization no longer requires PyTorch-specific arguments like model, input_data, target_class, or device. It now only tracks its own position, velocity, and clamping boundaries.
• Method Refactoring: * The fitness() method was replaced by evaluate(costFunc, model), delegating the evaluation logic to an external cost function rather than calculating softmax probabilities internally.
  • update_velocity now takes a simple list/tensor for the global best position (pos_best_g) instead of recalculating based on the model state.
  • Clamping constraints were simplified to use minclamp and maxclamp parameters instead of clip_value_position.

---

3. Swarm Logic (Swarm.py)

• Class Renaming: ParticleSwarm was renamed to PSO.
• Core Architectural Shift: * Like BirdParticle, the swarm initialization was heavily stripped down. It now accepts starting_positions and an explicit cost_func rather than hardcoding image classification concepts like starting_class and target_class.
  • The internal execution loop (run() with tqdm progress bars) and internal logging (setup_logging, log_progress) were removed. The swarm now relies on a step() function, giving the user control over the training loop.
• Data Handling & Output: * Added a get_history() method that outputs the swarm's positional history as a pandas.DataFrame, and a save_history() method to export it to a CSV.
  • getBest() and getPoints() were updated to return torch.Tensor objects rather than unhooked NumPy arrays.
  • The reduce_excess_perturbations method was completely removed.

---

4. Unit Tests (tests/)

• test_adversarial_attacks.py: Updated assertions to account for the attacks returning lists instead of single multi-dimensional tensors, and updated test_gradient_ascent to pass target_class instead of target_neuron.
• test_bird_particle.py: Removed the PyTorch model fixture dependency, as BirdParticle no longer requires it for initialization. Updated parameter names to match minclamp and maxclamp.
• test_pso_optimization.py: * Updated references from ParticleSwarm to PSO.
  • Implemented a dummy_cost_func fixture to fulfill the new PSO initialization requirements.
  • Updated type-checking assertions, as extraction methods now return torch.Tensor instead of np.ndarray.
  • Updated the logging test to verify the creation of the new pandas-driven history.csv file rather than a standard .log file.