scripts/analyze_layers.py

import torch
import os
from safetensors.torch import load_file

def count_layers(state_dict, exclude_prefixes=None):
    """
    Counts unique layers in a state dict. 
    Groups parameters by their module prefix (everything before the last dot).
    """
    if exclude_prefixes is None:
        exclude_prefixes = []
    
    total_layers = set()
    custom_layers = set()
    
    for key in state_dict.keys():
        parts = key.split('.')
        if len(parts) > 1:
            module_name = '.'.join(parts[:-1])
        else:
            module_name = key
            
        total_layers.add(module_name)
        
        # Check if this module is pretrained
        is_pretrained = any(key.startswith(p + '.') or key == p for p in exclude_prefixes)
        if not is_pretrained:
            custom_layers.add(module_name)
            
    return len(total_layers), len(custom_layers)

def count_parameters(state_dict, exclude_prefixes=None):
    if exclude_prefixes is None:
        exclude_prefixes = []
        
    total_params = 0
    custom_params = 0
    
    for name, param in state_dict.items():
        p_count = param.numel()
        total_params += p_count
        
        is_pretrained = any(name.startswith(p + '.') or name == p for p in exclude_prefixes)
        if not is_pretrained:
            custom_params += p_count
            
    return total_params, custom_params

# Define models and their pretrained prefixes
models_config = {
    "Bass": {
        "file": "bass_sota.pth",
        "exclude": ["audio_encoder"]
    },
    "Drums": {
        "file": "drums.safetensors",
        "exclude": ["wavlm"]
    },
    "Vocals": {
        "file": "vocals.pt",
        "exclude": []
    }
}

print(f"{'='*100}")
print(f"{'MODEL':<10} | {'TOTAL LAYERS':<15} | {'CUSTOM LAYERS':<15} | {'CUSTOM PARAMS':<15} | {'FILE'}")
print(f"{'='*100}")

for model_name, cfg in models_config.items():
    filename = cfg["file"]
    exclude = cfg["exclude"]
    
    if not os.path.exists(filename):
        print(f"{model_name:<10} | {'MISSING':<15} | {'N/A':<15} | {'N/A':<15} | {filename}")
        continue

    try:
        if filename.endswith(".safetensors"):
            data = load_file(filename, device='cpu')
        else:
            data = torch.load(filename, map_location='cpu', weights_only=False)
            
            # Handle cases where model is wrapped in a dict
            if isinstance(data, dict):
                if "model" in data:
                    data = data["model"]
                elif "model_state_dict" in data:
                    data = data["model_state_dict"]
                elif "state_dict" in data:
                    data = data["state_dict"]

        total_l, custom_l = count_layers(data, exclude)
        total_p, custom_p = count_parameters(data, exclude)
        
        print(f"{model_name:<10} | {total_l:<15} | {custom_l:<15} | {custom_p:<15,} | {filename}")
        
    except Exception as e:
        print(f"{model_name:<10} | {'ERROR':<15} | {'N/A':<15} | {'N/A':<15} | {filename} - {str(e)[:30]}...")

print(f"{'='*100}")