Evaluate NN accuracy dependence on training data size

[k-yoshimi]

Summary

Add functionality to evaluate and visualize how NN accuracy (MSE, etc.) depends on the number of training samples.

Background

• BoxData.generate_learning_data(n_samples) can generate datasets of varying sizes
• We want to train the same NN architecture (models/encoder_1/encoder.py) with different data sizes and plot the learning curve (data size vs. accuracy)

Tasks

• Script to automatically run data generation → training → evaluation for multiple data sizes (e.g., 100, 500, 1000, 5000, 10000, 20000)
• Plot learning curve: data size vs. MSE (train/validation/test)
• Save results (CSV, etc.)

Related files

• src/seap/prediction/datasets.py — data generation
• models/encoder_1/encoder.py — NN model and training loop
• src/seap/prediction/utils.py — plot_learning_curve() and other utilities