Neural_Nets_Doing_Simple_Tasks / PARITY-calculatingNN_Schmidhuber_V2.0.py
MartialTerran's picture
Upload PARITY-calculatingNN_Schmidhuber_V2.0.py
c8b58ee verified
# Filename: PARITY-calculatingNN_Supercoded_V2.0.py
# Description: An advanced, highly configurable PyTorch script to train a neural network on the N-bit parity problem.
#
# Supercoding LLM Recode of PARITY-calculatingNN_Schmidhuber_V1.0.py
#
# This version introduces significant enhancements for robust experimentation and analysis:
# - Full Hyperparameterization: All key parameters are exposed via command-line arguments for easy tuning.
# - Flexible Problem Definition: The concept of "parity" can be switched between 'even', 'odd', or 'majority' rule,
# allowing the network to be tested on different but related logical problems.
# - Advanced Visualization: Generates a suite of high-quality plots (saved to disk and shown in popups) inspired
# by analytical scientific scripts, including:
# - Detailed Training History (Loss & Accuracy)
# - Confusion Matrix Heatmap
# - Prediction Margin Distribution Histogram
# - Raw Output vs. True Label Scatter Plot
# - Comprehensive Reporting: Automatically generates a run-specific folder containing the plots, a detailed text report,
# the trained model weights, and a log of hard-to-learn data samples.
# - Interactive Mode: Allows the user to test the trained model with custom binary inputs.
# ==============================================================================
# === LIBRARY IMPORTS ===
# ==============================================================================
print("Initializing... Loading libraries.")
import torch
import torch.nn as nn
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.metrics import confusion_matrix, classification_report
import random
import os
import argparse
import datetime
import json
print("Libraries loaded successfully.\n")
# ==============================================================================
# === CORE FUNCTIONS ===
# ==============================================================================
def setup_directories(args):
"""Creates a unique, timestamped directory for the current run to store all outputs."""
timestamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
run_name = f"run_{timestamp}_N{args.n_bits}_L{args.l_layers}_H{args.hidden_size}_{args.parity_type}"
base_dir = "parity_nn_results"
run_dir = os.path.join(base_dir, run_name)
plots_dir = os.path.join(run_dir, "plots")
os.makedirs(plots_dir, exist_ok=True)
print(f"Created output directory: {run_dir}")
return run_dir, plots_dir
def generate_data(num_samples, num_bits, parity_type='even'):
"""
Generates input data and corresponding labels based on the specified parity rule.
Args:
num_samples (int): The number of data samples to generate.
num_bits (int): The bit-width of each data sample.
parity_type (str): The rule for calculating the label.
'even': Standard even parity bit (1 if odd number of 1s, 0 if even).
'odd': Standard odd parity bit (1 if even number of 1s, 0 if odd).
'majority': Label is 1 if the count of 1s > N/2, else 0.
Returns:
tuple: A tuple containing torch tensors for data and labels.
"""
data = []
labels = []
for _ in range(num_samples):
bits = [random.randint(0, 1) for _ in range(num_bits)]
sum_of_bits = sum(bits)
if parity_type == 'even':
label = sum_of_bits % 2
elif parity_type == 'odd':
label = (sum_of_bits + 1) % 2
elif parity_type == 'majority':
label = 1 if sum_of_bits > num_bits / 2 else 0
else:
raise ValueError(f"Unknown parity_type: {parity_type}")
data.append(bits)
labels.append(label)
return torch.tensor(data, dtype=torch.float32), torch.tensor(labels, dtype=torch.float32).reshape(-1, 1)
class ParityNet(nn.Module):
"""A flexible feed-forward neural network model."""
def __init__(self, input_size, hidden_size, num_hidden_layers, output_size, activation_func='relu'):
super(ParityNet, self).__init__()
if activation_func.lower() == 'relu':
activation = nn.ReLU()
elif activation_func.lower() == 'tanh':
activation = nn.Tanh()
elif activation_func.lower() == 'leakyrelu':
activation = nn.LeakyReLU()
else:
raise ValueError("Unsupported activation function. Choose 'relu', 'tanh', or 'leakyrelu'.")
layers = []
# Input layer
layers.append(nn.Linear(input_size, hidden_size))
layers.append(activation)
# Hidden layers
for _ in range(num_hidden_layers - 1):
layers.append(nn.Linear(hidden_size, hidden_size))
layers.append(activation)
# Output layer
layers.append(nn.Linear(hidden_size, output_size))
layers.append(nn.Sigmoid()) # For binary classification
self.layers = nn.Sequential(*layers)
def forward(self, x):
return self.layers(x)
def train_model(model, train_data, train_labels, test_data, test_labels, args, run_dir):
"""Trains the model and logs performance and difficult samples."""
criterion = nn.BCELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
history = {'epoch': [], 'train_loss': [], 'val_acc': []}
hard_samples = {}
print("\n" + "="*50)
print("=== STARTING TRAINING ===")
print(f"Epochs: {args.epochs}, LR: {args.learning_rate}, Stop Threshold: {args.min_loss_threshold}")
print("="*50)
# Interactive plot setup
plt.ion()
fig, ax = plt.subplots(figsize=(10, 6))
for epoch in range(args.epochs):
model.train()
outputs = model(train_data)
loss = criterion(outputs, train_labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (epoch + 1) % args.plot_update_freq == 0:
model.eval()
with torch.no_grad():
val_outputs = model(test_data)
predicted = (val_outputs > 0.5).float()
accuracy = (predicted == test_labels).sum().float() / len(test_labels)
# Log hard samples (misclassified during this validation check)
misclassified_mask = (predicted != test_labels).flatten()
for i, is_misclassified in enumerate(misclassified_mask):
if is_misclassified:
sample_str = ''.join(map(str, test_data[i].int().tolist()))
hard_samples[sample_str] = hard_samples.get(sample_str, 0) + 1
history['epoch'].append(epoch + 1)
history['train_loss'].append(loss.item())
history['val_acc'].append(accuracy.item())
print(f'Epoch [{epoch+1}/{args.epochs}], Loss: {loss.item():.5f}, Validation Accuracy: {accuracy.item():.4f}')
# Update real-time plot
ax.clear()
ax.plot(history['epoch'], history['train_loss'], 'r-', label='Training Loss')
ax.set_xlabel(f"Epoch (x{args.plot_update_freq})")
ax.set_ylabel("Loss", color='r')
ax.tick_params(axis='y', labelcolor='r')
ax2 = ax.twinx()
ax2.plot(history['epoch'], history['val_acc'], 'b-', label='Validation Accuracy')
ax2.set_ylabel("Accuracy", color='b')
ax2.tick_params(axis='y', labelcolor='b')
ax2.set_ylim(0, 1.05)
fig.suptitle("Live Training Progress")
fig.legend(loc="upper center", bbox_to_anchor=(0.5, 0.95), ncol=2)
plt.grid(True)
plt.draw()
plt.pause(0.01)
if loss.item() < args.min_loss_threshold:
print(f"\nReached minimum loss threshold of {args.min_loss_threshold} at epoch {epoch+1}. Stopping training.")
break
plt.ioff()
print("\n" + "="*50)
print("=== TRAINING COMPLETE ===")
print("="*50 + "\n")
# Save hard samples log
if hard_samples:
hard_samples_path = os.path.join(run_dir, "hard_samples.json")
sorted_hard_samples = sorted(hard_samples.items(), key=lambda item: item[1], reverse=True)
with open(hard_samples_path, 'w') as f:
json.dump(dict(sorted_hard_samples), f, indent=4)
print(f"Logged {len(hard_samples)} unique hard-to-learn samples to {hard_samples_path}")
return model, history
def evaluate_model(model, test_data, test_labels):
"""Evaluates the final model and returns a dictionary of metrics."""
model.eval()
with torch.no_grad():
outputs = model(test_data)
predicted = (outputs > 0.5).float()
accuracy = (predicted == test_labels).sum().float() / len(test_labels)
# Separate margins for predictions of 1 and 0
margins_ones = outputs[predicted == 1] - 0.5
margins_zeros = 0.5 - outputs[predicted == 0]
results = {
"accuracy": accuracy.item(),
"raw_outputs": outputs.flatten().numpy(),
"predictions": predicted.flatten().numpy(),
"labels": test_labels.flatten().numpy(),
"margins_ones": margins_ones.numpy(),
"margins_zeros": margins_zeros.numpy(),
"conf_matrix": confusion_matrix(test_labels.numpy(), predicted.numpy()),
"class_report": classification_report(test_labels.numpy(), predicted.numpy(), output_dict=True, zero_division=0)
}
return results
def generate_plots(history, results, args, plots_dir):
"""Generates and saves a suite of analytical plots."""
print("Generating and saving analysis plots...")
plt.style.use('seaborn-v0_8-whitegrid')
# --- 1. Training History Plot ---
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(18, 6))
fig.suptitle(f'Training History for N={args.n_bits} {args.parity_type.title()} Parity', fontsize=16)
ax1.plot(history['epoch'], history['train_loss'], label='Training Loss', color='crimson')
ax1.set_title('Training Loss over Epochs')
ax1.set_xlabel('Epoch')
ax1.set_ylabel('Binary Cross-Entropy Loss')
ax1.legend()
ax2.plot(history['epoch'], history['val_acc'], label='Validation Accuracy', color='royalblue')
ax2.set_title('Validation Accuracy over Epochs')
ax2.set_xlabel('Epoch')
ax2.set_ylabel('Accuracy')
ax2.set_ylim(0, 1.05)
ax2.legend()
plt.tight_layout(rect=[0, 0.03, 1, 0.95])
plt.savefig(os.path.join(plots_dir, "01_training_history.png"))
plt.show()
# --- 2. Confusion Matrix Heatmap ---
plt.figure(figsize=(8, 6))
sns.heatmap(results['conf_matrix'], annot=True, fmt='d', cmap='Blues',
xticklabels=['Predicted 0', 'Predicted 1'],
yticklabels=['Actual 0', 'Actual 1'])
plt.title('Confusion Matrix on Test Set', fontsize=14)
plt.ylabel('True Label')
plt.xlabel('Predicted Label')
plt.savefig(os.path.join(plots_dir, "02_confusion_matrix.png"))
plt.show()
# --- 3. Prediction Margin Distribution ---
plt.figure(figsize=(12, 6))
plt.hist(results['margins_zeros'], bins=20, alpha=0.7, color='coral', label='Margins for "0" Predictions')
plt.hist(results['margins_ones'], bins=20, alpha=0.7, color='teal', label='Margins for "1" Predictions')
plt.title('Distribution of Prediction Margins (Confidence)', fontsize=14)
plt.xlabel('Margin (Distance from 0.5 threshold)')
plt.ylabel('Frequency')
plt.legend()
plt.savefig(os.path.join(plots_dir, "03_prediction_margins.png"))
plt.show()
# --- 4. Raw Outputs vs. Labels ---
plt.figure(figsize=(10, 7))
jitter = np.random.normal(0, 0.015, size=len(results['labels'])) # For better visualization
colors = ['coral' if l == 0 else 'teal' for l in results['labels']]
plt.scatter(results['labels'] + jitter, results['raw_outputs'], c=colors, alpha=0.6)
plt.axhline(y=0.5, color='r', linestyle='--', label='Decision Boundary (0.5)')
plt.title('Model Raw Output vs. True Labels', fontsize=14)
plt.xlabel('True Label (with jitter)')
plt.ylabel('Sigmoid Output (Probability)')
plt.xticks([0, 1], ['Class 0', 'Class 1'])
plt.legend()
plt.savefig(os.path.join(plots_dir, "04_outputs_vs_labels.png"))
plt.show()
print("All plots saved.")
def generate_report(args, results, run_dir):
"""Generates and saves a comprehensive text report of the run."""
report_path = os.path.join(run_dir, "report.txt")
# Margin stats
def get_margin_stats(margins):
if len(margins) == 0: return "N/A", "N/A", "N/A"
return f"{np.min(margins):.4f}", f"{np.max(margins):.4f}", f"{np.mean(margins):.4f}"
min_m0, max_m0, avg_m0 = get_margin_stats(results['margins_zeros'])
min_m1, max_m1, avg_m1 = get_margin_stats(results['margins_ones'])
tn, fp, fn, tp = results['conf_matrix'].ravel() if results['conf_matrix'].size == 4 else (0,0,0,0)
report_content = f"""
# ==========================================================
# == PARITY NEURAL NETWORK EXPERIMENT REPORT
# ==========================================================
# Run Directory: {run_dir}
# Report Time: {datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")}
#
# ==========================================================
# == HYPERPARAMETERS
# ==========================================================
# Problem Type: {args.parity_type.title()}
# Input Bits (N): {args.n_bits}
# Hidden Layers (L): {args.l_layers}
# Neurons per Hidden Layer: {args.hidden_size}
# Activation Function: {args.activation.upper()}
#
# --- Training Configuration ---
# Epochs: {args.epochs}
# Learning Rate: {args.learning_rate}
# Train Samples: {args.num_train_samples}
# Test Samples: {args.num_test_samples}
# Loss Stop Threshold: {args.min_loss_threshold}
#
# ==========================================================
# == EVALUATION RESULTS
# ==========================================================
# Final Test Accuracy: {results['accuracy']:.4f}
#
# --- Confusion Matrix ---
# True Positives (1->1): {tp}
# True Negatives (0->0): {tn}
# False Positives (0->1): {fp}
# False Negatives (1->0): {fn}
#
# --- Prediction Margin Analysis (Confidence) ---
# | Min | Max | Average
# ----------------------------------------------------------
# Margin (Zeros): | {min_m0:<7} | {max_m0:<7} | {avg_m0:<7}
# Margin (Ones): | {min_m1:<7} | {max_m1:<7} | {avg_m1:<7}
#
# ==========================================================
# == CONCLUSION
# ==========================================================
# The model was trained to solve the {args.n_bits}-bit '{args.parity_type}' problem.
# With a final test accuracy of {results['accuracy']:.2%}, the network has demonstrated
# {'a high degree of success' if results['accuracy'] > 0.95 else 'a moderate level of success' if results['accuracy'] > 0.7 else 'difficulty'}
# in learning the underlying logical rule.
#
# The margin analysis indicates the model's confidence. Larger average margins
# suggest a more robust and decisive model. The generated plots provide
# further visual insight into the training process and final performance.
#
# This experiment explores the capability of a simple Feed-Forward Network
# to learn complex, non-linear functions like parity, a task often cited
# as a challenge for non-recurrent architectures but clearly achievable
# with sufficient network capacity and training.
#
"""
print("\n" + report_content)
with open(report_path, "w") as f:
f.write(report_content)
print(f"Report saved to {report_path}")
def user_inference_loop(model, args):
"""An interactive loop for the user to test the model. replicate the label-calculation logic directly within the user_inference_loop. This isolates the calculation and uses the user's provided bit string, fixing the crash and ensuring the "True Label" is accurate for the given input. With this correction, the interactive mode will now function as intended, correctly comparing the model's prediction against the true calculated label for your input. System integrity restored."""
print("\n" + "="*50)
print("=== INTERACTIVE INFERENCE MODE ===")
print(f"Enter a {args.n_bits}-bit binary string (e.g., {'10101'[:args.n_bits]}) or 'q' to quit.")
print("="*50)
model.eval()
while True:
user_input = input(f"Input ({args.n_bits} bits) > ")
if user_input.lower() == 'q':
break
if len(user_input) != args.n_bits or not all(c in '01' for c in user_input):
print(f"Error: Please enter exactly {args.n_bits} bits (0s and 1s).")
continue
bits = [int(c) for c in user_input]
data_tensor = torch.tensor(bits, dtype=torch.float32).reshape(1, -1)
with torch.no_grad():
output = model(data_tensor)
prediction = (output > 0.5).int().item()
confidence = output.item()
print(f" Model Output: {confidence:.4f}")
print(f" -> Predicted Label: {prediction}")
# --- FIX START ---
# The original code incorrectly tried to pass the user's data back into
# the generate_data function. The fix is to calculate the true label
# directly here using the same logic from the data generation process.
sum_of_bits = sum(bits)
true_label = -1 # Default/error value
if args.parity_type == 'even':
true_label = sum_of_bits % 2
elif args.parity_type == 'odd':
true_label = (sum_of_bits + 1) % 2
elif args.parity_type == 'majority':
true_label = 1 if sum_of_bits > args.n_bits / 2 else 0
# --- FIX END ---
print(f" -> True Label ({args.parity_type}): {true_label} {'(Correct)' if prediction == true_label else '(Incorrect)'}\n")
# ==============================================================================
# === MAIN EXECUTION BLOCK ===
# ==============================================================================
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Train a Neural Network for the N-Bit Parity Problem.")
# --- Model Architecture ---
parser.add_argument('-n', '--n_bits', type=int, default=5, help="Number of input bits (N).")
parser.add_argument('-l', '--l_layers', type=int, default=2, help="Number of hidden layers (L).")
parser.add_argument('-hs', '--hidden_size', type=int, default=10, help="Number of neurons per hidden layer.")
parser.add_argument('-a', '--activation', type=str, default='relu', choices=['relu', 'tanh', 'leakyrelu'], help="Activation function for hidden layers.")
# --- Training Parameters ---
parser.add_argument('-e', '--epochs', type=int, default=10000, help="Maximum number of training epochs.")
parser.add_argument('-lr', '--learning_rate', type=float, default=0.003, help="Optimizer learning rate.")
parser.add_argument('-loss', '--min_loss_threshold', type=float, default=0.01, help="Loss threshold to stop training early.")
parser.add_argument('-puf', '--plot_update_freq', type=int, default=100, help="Frequency (in epochs) to update the live plot.")
# --- Data and Problem Type ---
parser.add_argument('-pt', '--parity_type', type=str, default='even', choices=['even', 'odd', 'majority'], help="The logical rule to learn.")
parser.add_argument('-nts', '--num_train_samples', type=int, default=2000, help="Number of samples for the training dataset.")
parser.add_argument('-ntests', '--num_test_samples', type=int, default=500, help="Number of samples for the test dataset.")
args = parser.parse_args()
# 1. Setup
run_dir, plots_dir = setup_directories(args)
# Save args for reproducibility
with open(os.path.join(run_dir, 'hyperparameters.json'), 'w') as f:
json.dump(vars(args), f, indent=4)
# 2. Generate Data
print(f"\nGenerating {args.num_train_samples} training and {args.num_test_samples} test samples for {args.n_bits}-bit '{args.parity_type}' parity...")
train_data, train_labels = generate_data(args.num_train_samples, args.n_bits, args.parity_type)
test_data, test_labels = generate_data(args.num_test_samples, args.n_bits, args.parity_type)
print("Data generation complete.")
# 3. Create Model
model = ParityNet(args.n_bits, args.hidden_size, args.l_layers, 1, args.activation)
print("\nModel Architecture:")
print(model)
# 4. Train Model
trained_model, history = train_model(model, train_data, train_labels, test_data, test_labels, args, run_dir)
# 5. Save Model Weights
model_path = os.path.join(run_dir, f"parity_nn_N{args.n_bits}_{args.parity_type}.pth")
torch.save(trained_model.state_dict(), model_path)
print(f"Trained model weights saved to: {model_path}")
# 6. Evaluate and Report
if len(history['epoch']) > 0: # Ensure training ran for at least one update cycle
final_results = evaluate_model(trained_model, test_data, test_labels)
generate_report(args, final_results, run_dir)
generate_plots(history, final_results, args, plots_dir)
else:
print("\nTraining was too short to generate a full report and plots.")
# 7. Interactive Mode
user_inference_loop(trained_model, args)
print("\nSupercoding LLM task complete. System returning to normal operation.")