test_rbs_debug.py

#!/usr/bin/env python3
"""
Quick test script to verify RBS debug logging works correctly.
This loads cached logits and runs a quick calibration test.
"""
import sys
import os
import logging
import numpy as np
import torch
import torch.nn.functional as F

# Add paths
sys.path.append('/hdd/haolan')
sys.path.append('/hdd/haolan/SMART')

from calibrator.Component.utils.utils import get_all_metrics
from calibrator.Component.model.temperature_scaling import TemperatureScalingCalibrator

# Set up logging to see debug messages
logging.basicConfig(level=logging.INFO, format='%(name)s - %(levelname)s - %(message)s')

print("="*80)
print("Testing RBS Debug Logging")
print("="*80)

# Load some cached logits for testing
cache_dir = "/hdd/haolan/SMART/cache/cifar10_resnet50_seed1"
test_logits = np.load(f"{cache_dir}/test_logits.npy")
test_labels = np.load(f"{cache_dir}/test_labels.npy")
val_logits = np.load(f"{cache_dir}/val_logits.npy")
val_labels = np.load(f"{cache_dir}/val_labels.npy")

# Convert to tensors
test_logits_tensor = torch.tensor(test_logits, dtype=torch.float32)
test_labels_tensor = torch.tensor(test_labels, dtype=torch.long)
val_logits_tensor = torch.tensor(val_logits, dtype=torch.float32)
val_labels_tensor = torch.tensor(val_labels, dtype=torch.long)

print(f"\nLoaded data:")
print(f"  Test logits shape: {test_logits_tensor.shape}")
print(f"  Test labels shape: {test_labels_tensor.shape}")

# Test 1: Uncalibrated
print("\n" + "="*80)
print("TEST 1: Uncalibrated Logits")
print("="*80)
uncal_metrics = get_all_metrics(
    labels=test_labels_tensor,
    logits=test_logits_tensor,
    probs=None,
    n_bins=15
)
print(f"Uncalibrated RBS: {uncal_metrics['rbs']:.6f}")
print(f"Uncalibrated NLL: {uncal_metrics['nll']:.6f}")
print(f"Uncalibrated ECE: {uncal_metrics['ece']:.6f}")

# Test 2: Temperature Scaling
print("\n" + "="*80)
print("TEST 2: Temperature Scaling")
print("="*80)

# Train temperature scaling
ts_calibrator = TemperatureScalingCalibrator(loss_type='CE')
ts_calibrator.fit(val_logits_tensor, val_labels_tensor)
optimal_temp = ts_calibrator.temperature.item()
print(f"Optimal temperature: {optimal_temp:.6f}")

# Get calibrated logits
calibrated_logits = ts_calibrator.calibrate(test_logits_tensor, return_logits=True)
print(f"Calibrated logits stats:")
print(f"  Mean: {calibrated_logits.mean().item():.6f}")
print(f"  Std: {calibrated_logits.std().item():.6f}")
print(f"  Sample [0, :5]: {calibrated_logits[0, :5].tolist()}")

# Compute metrics with calibrated logits
cal_metrics = get_all_metrics(
    labels=test_labels_tensor,
    logits=calibrated_logits,
    probs=None,
    n_bins=15
)
print(f"\nTemperature Scaling RBS: {cal_metrics['rbs']:.6f}")
print(f"Temperature Scaling NLL: {cal_metrics['nll']:.6f}")
print(f"Temperature Scaling ECE: {cal_metrics['ece']:.6f}")

# Test 3: Using probabilities directly
print("\n" + "="*80)
print("TEST 3: Direct Probabilities (simulating SMART)")
print("="*80)

# Convert calibrated logits to probabilities
ts_probs = F.softmax(calibrated_logits, dim=1)
print(f"Probability stats:")
print(f"  Mean: {ts_probs.mean().item():.6f}")
print(f"  Std: {ts_probs.std().item():.6f}")
print(f"  Sample [0, :5]: {ts_probs[0, :5].tolist()}")
print(f"  Sum [0]: {ts_probs[0].sum().item():.6f}")

# Compute metrics with probabilities
probs_metrics = get_all_metrics(
    labels=test_labels_tensor,
    logits=None,
    probs=ts_probs,
    n_bins=15
)
print(f"\nDirect Probs RBS: {probs_metrics['rbs']:.6f}")
print(f"Direct Probs NLL: {probs_metrics['nll']:.6f}")
print(f"Direct Probs ECE: {probs_metrics['ece']:.6f}")

# Compare results
print("\n" + "="*80)
print("COMPARISON")
print("="*80)
print(f"Method                    RBS          NLL          ECE")
print(f"Uncalibrated:          {uncal_metrics['rbs']:.6f}     {uncal_metrics['nll']:.6f}     {uncal_metrics['ece']:.6f}")
print(f"TS (logits):           {cal_metrics['rbs']:.6f}     {cal_metrics['nll']:.6f}     {cal_metrics['ece']:.6f}")
print(f"TS (probs):            {probs_metrics['rbs']:.6f}     {probs_metrics['nll']:.6f}     {probs_metrics['ece']:.6f}")
print(f"\nRBS Change (uncal→TS):  {(cal_metrics['rbs'] - uncal_metrics['rbs']):.6f} ({100*(cal_metrics['rbs'] - uncal_metrics['rbs'])/uncal_metrics['rbs']:.2f}%)")
print(f"ECE Change (uncal→TS):  {(cal_metrics['ece'] - uncal_metrics['ece']):.6f} ({100*(cal_metrics['ece'] - uncal_metrics['ece'])/uncal_metrics['ece']:.2f}%)")
print(f"\nNote: TS (logits) and TS (probs) should give identical results!")
print(f"Match: {abs(cal_metrics['rbs'] - probs_metrics['rbs']) < 1e-6}")

print("\n" + "="*80)
print("Test complete! Check the debug logs above.")
print("="*80)