.debug/test.py

import numpy as np
import json
import torch

# Load stats file
stats_file = "../meta/stats.json"
with open(stats_file, 'r') as f:
    stats = json.load(f)

print("Keys in stats file:", list(stats.keys()))

# Check each field more thoroughly by simulating the torch conversion
problem_found = False
for key in stats:
    print(f"\nExamining key: {key}")
    for stat_type in ["mean", "std"]:
        if stat_type in stats[key]:
            value = stats[key][stat_type]
            print(f"  - {stat_type} type: {type(value)}")
            
            try:
                # Try converting to numpy array (this is what the code does)
                numpy_val = np.array(value)
                print(f"    NumPy array shape: {numpy_val.shape}, dtype: {numpy_val.dtype}")
                
                # Test if this can be converted to torch tensor
                try:
                    torch_val = torch.from_numpy(numpy_val).to(dtype=torch.float32)
                    print(f"    ✓ Successfully converted to torch tensor")
                except Exception as e:
                    print(f"    ✗ ERROR converting to torch tensor: {type(e).__name__}: {e}")
                    print(f"    Value: {numpy_val}")
                    problem_found = True
            except Exception as e:
                print(f"    ✗ ERROR converting to NumPy array: {type(e).__name__}: {e}")
                print(f"    Value: {value}")
                problem_found = True

if not problem_found:
    print("\nNo problems found in the stats file. The error might be happening in another part of the code.")
    print("Try fixing the stats file with a preprocessing step:")

    # Create a fixed stats file
    fixed_stats = {}
    for key in stats:
        fixed_stats[key] = {}
        for field, value in stats[key].items():
            if field in ["mean", "std"]:
                # Try to safely convert to float array
                try:
                    # First convert to numpy array
                    arr = np.array(value)
                    # If object type, replace with safe values
                    if arr.dtype == np.dtype('O'):
                        arr = np.ones_like(arr, dtype=np.float32)
                    else:
                        # Make sure it's float32
                        arr = arr.astype(np.float32)
                    # Convert back to list for JSON
                    fixed_stats[key][field] = arr.tolist()
                except Exception as e:
                    print(f"Fixing {key}.{field}: {e}")
                    # Just use a safe default
                    fixed_stats[key][field] = [1.0]
            else:
                # Keep other fields unchanged
                fixed_stats[key][field] = value
    
    # Save the fixed stats
    with open("../meta/stats_fixed.json", 'w') as f:
        json.dump(fixed_stats, f, indent=2)
    
    print("Created fixed stats file: ../meta/stats_fixed.json")
    print("You can replace the original stats.json with this file.")