dragon-ml-toolbox 19.13.0__py3-none-any.whl → 20.0.0__py3-none-any.whl

ml_tools/ML_utilities/_inspection.py

@@ -0,0 +1,325 @@
+import pandas as pd
+from pathlib import Path
+from typing import Union, Optional
+import torch
+from torch import nn
+
+from ..utilities import load_dataframe
+from ..IO_tools import save_list_strings, custom_logger
+
+from ..path_manager import make_fullpath, list_subdirectories
+from .._core import get_logger
+from ..keys._keys import DatasetKeys, SHAPKeys, UtilityKeys, PyTorchCheckpointKeys
+
+
+_LOGGER = get_logger("ML Inspection")
+
+
+__all__ = [
+    "get_model_parameters",
+    "inspect_model_architecture",
+    "inspect_pth_file",
+    "select_features_by_shap"
+]
+
+
+def get_model_parameters(model: nn.Module, save_dir: Optional[Union[str,Path]]=None) -> dict[str, int]:
+    """
+    Calculates the total and trainable parameters of a PyTorch model.
+
+    Args:
+        model (nn.Module): The PyTorch model to inspect.
+        save_dir: Optional directory to save the output as a JSON file.
+
+    Returns:
+        Dict[str, int]: A dictionary containing:
+            - "total_params": The total number of parameters.
+            - "trainable_params": The number of trainable parameters (where requires_grad=True).
+    """
+    total_params = sum(p.numel() for p in model.parameters())
+    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    
+    report = {
+        UtilityKeys.TOTAL_PARAMS: total_params,
+        UtilityKeys.TRAINABLE_PARAMS: trainable_params
+    }
+    
+    if save_dir is not None:
+        output_dir = make_fullpath(save_dir, make=True, enforce="directory")
+        custom_logger(data=report,
+                      save_directory=output_dir,
+                      log_name=UtilityKeys.MODEL_PARAMS_FILE,
+                      add_timestamp=False,
+                      dict_as="json")
+
+    return report
+
+
+def inspect_model_architecture(
+    model: nn.Module,
+    save_dir: Union[str, Path]
+) -> None:
+    """
+    Saves a human-readable text summary of a model's instantiated
+    architecture, including parameter counts.
+
+    Args:
+        model (nn.Module): The PyTorch model to inspect.
+        save_dir (str | Path): Directory to save the text file.
+    """
+    # --- 1. Validate path ---
+    output_dir = make_fullpath(save_dir, make=True, enforce="directory")
+    architecture_filename = UtilityKeys.MODEL_ARCHITECTURE_FILE + ".txt"
+    filepath = output_dir / architecture_filename
+
+    # --- 2. Get parameter counts from existing function ---
+    try:
+        params_report = get_model_parameters(model) # Get dict, don't save
+        total = params_report.get(UtilityKeys.TOTAL_PARAMS, 'N/A')
+        trainable = params_report.get(UtilityKeys.TRAINABLE_PARAMS, 'N/A')
+        header = (
+            f"Model: {model.__class__.__name__}\n"
+            f"Total Parameters: {total:,}\n"
+            f"Trainable Parameters: {trainable:,}\n"
+            f"{'='*80}\n\n"
+        )
+    except Exception as e:
+        _LOGGER.warning(f"Could not get model parameters: {e}")
+        header = f"Model: {model.__class__.__name__}\n{'='*80}\n\n"
+
+    # --- 3. Get architecture string ---
+    arch_string = str(model)
+
+    # --- 4. Write to file ---
+    try:
+        with open(filepath, 'w', encoding='utf-8') as f:
+            f.write(header)
+            f.write(arch_string)
+        _LOGGER.info(f"Model architecture summary saved to '{filepath.name}'")
+    except Exception as e:
+        _LOGGER.error(f"Failed to write model architecture file: {e}")
+        raise
+
+
+def inspect_pth_file(
+    pth_path: Union[str, Path],
+    save_dir: Union[str, Path],
+) -> None:
+    """
+    Inspects a .pth file (e.g., checkpoint) and saves a human-readable
+    JSON summary of its contents.
+
+    Args:
+        pth_path (str | Path): The path to the .pth file to inspect.
+        save_dir (str | Path): The directory to save the JSON report.
+
+    Returns:
+        Dict (str, Any): A dictionary containing the inspection report.
+
+    Raises:
+        ValueError: If the .pth file is empty or in an unrecognized format.
+    """
+    # --- 1. Validate paths ---
+    pth_file = make_fullpath(pth_path, enforce="file")
+    output_dir = make_fullpath(save_dir, make=True, enforce="directory")
+    pth_name = pth_file.stem
+
+    # --- 2. Load data ---
+    try:
+        # Load onto CPU to avoid GPU memory issues
+        loaded_data = torch.load(pth_file, map_location=torch.device('cpu'))
+    except Exception as e:
+        _LOGGER.error(f"Failed to load .pth file '{pth_file}': {e}")
+        raise
+
+    # --- 3. Initialize Report ---
+    report = {
+        "top_level_type": str(type(loaded_data)),
+        "top_level_summary": {},
+        "model_state_analysis": None,
+        "notes": []
+    }
+
+    # --- 4. Parse loaded data ---
+    if isinstance(loaded_data, dict):
+        # --- Case 1: Loaded data is a dictionary (most common case) ---
+        # "main loop" that iterates over *everything* first.
+        for key, value in loaded_data.items():
+            key_summary = {}
+            val_type = str(type(value))
+            key_summary["type"] = val_type
+            
+            if isinstance(value, torch.Tensor):
+                key_summary["shape"] = list(value.shape)
+                key_summary["dtype"] = str(value.dtype)
+            elif isinstance(value, dict):
+                key_summary["key_count"] = len(value)
+                key_summary["key_preview"] = list(value.keys())[:5]
+            elif isinstance(value, (int, float, str, bool)):
+                key_summary["value_preview"] = str(value)
+            elif isinstance(value, (list, tuple)):
+                 key_summary["value_preview"] = str(value)[:100]
+            
+            report["top_level_summary"][key] = key_summary
+
+        # Now, try to find the model state_dict within the dict
+        if PyTorchCheckpointKeys.MODEL_STATE in loaded_data and isinstance(loaded_data[PyTorchCheckpointKeys.MODEL_STATE], dict):
+            report["notes"].append(f"Found standard checkpoint key: '{PyTorchCheckpointKeys.MODEL_STATE}'. Analyzing as model state_dict.")
+            state_dict = loaded_data[PyTorchCheckpointKeys.MODEL_STATE]
+            report["model_state_analysis"] = _generate_weight_report(state_dict)
+        
+        elif all(isinstance(v, torch.Tensor) for v in loaded_data.values()):
+            report["notes"].append("File dictionary contains only tensors. Analyzing entire dictionary as model state_dict.")
+            state_dict = loaded_data
+            report["model_state_analysis"] = _generate_weight_report(state_dict)
+        
+        else:
+            report["notes"].append("Could not identify a single model state_dict. See top_level_summary for all contents. No detailed weight analysis will be performed.")
+
+    elif isinstance(loaded_data, nn.Module):
+        # --- Case 2: Loaded data is a full pickled model ---
+        # _LOGGER.warning("Loading a full, pickled nn.Module is not recommended. Inspecting its state_dict().")
+        report["notes"].append("File is a full, pickled nn.Module. This is not recommended. Extracting state_dict() for analysis.")
+        state_dict = loaded_data.state_dict()
+        report["model_state_analysis"] = _generate_weight_report(state_dict)
+
+    else:
+        # --- Case 3: Unrecognized format (e.g., single tensor, list) ---
+        _LOGGER.error(f"Could not parse .pth file. Loaded data is of type {type(loaded_data)}, not a dict or nn.Module.")
+        raise ValueError()
+
+    # --- 5. Save Report ---
+    custom_logger(data=report,
+                  save_directory=output_dir,
+                  log_name=UtilityKeys.PTH_FILE + pth_name,
+                  add_timestamp=False,
+                  dict_as="json")
+
+
+def _generate_weight_report(state_dict: dict) -> dict:
+    """
+    Internal helper to analyze a state_dict and return a structured report.
+    
+    Args:
+        state_dict (dict): The model state_dict to analyze.
+
+    Returns:
+        dict: A report containing total parameters and a per-parameter breakdown.
+    """
+    weight_report = {}
+    total_params = 0
+    if not isinstance(state_dict, dict):
+        _LOGGER.warning(f"Attempted to generate weight report on non-dict type: {type(state_dict)}")
+        return {"error": "Input was not a dictionary."}
+
+    for key, tensor in state_dict.items():
+        if not isinstance(tensor, torch.Tensor):
+             _LOGGER.warning(f"Skipping key '{key}' in state_dict: value is not a tensor (type: {type(tensor)}).")
+             weight_report[key] = {
+                 "type": str(type(tensor)),
+                 "value_preview": str(tensor)[:50] # Show a preview
+             }
+             continue
+        weight_report[key] = {
+            "shape": list(tensor.shape),
+            "dtype": str(tensor.dtype),
+            "requires_grad": tensor.requires_grad,
+            "num_elements": tensor.numel()
+        }
+        total_params += tensor.numel()
+
+    return {
+        "total_parameters": total_params,
+        "parameter_key_count": len(weight_report),
+        "parameters": weight_report
+    }
+
+
+def select_features_by_shap(
+    root_directory: Union[str, Path],
+    shap_threshold: float,
+    log_feature_names_directory: Optional[Union[str, Path]],
+    verbose: bool = True) -> list[str]:
+    """
+    Scans subdirectories to find SHAP summary CSVs, then extracts feature
+    names whose mean absolute SHAP value meets a specified threshold.
+
+    This function is useful for automated feature selection based on feature
+    importance scores aggregated from multiple models.
+
+    Args:
+        root_directory (str | Path):
+            The path to the root directory that contains model subdirectories.
+        shap_threshold (float):
+            The minimum mean absolute SHAP value for a feature to be included
+            in the final list.
+        log_feature_names_directory (str | Path | None):
+            If given, saves the chosen feature names as a .txt file in this directory.
+
+    Returns:
+        list[str]:
+            A single, sorted list of unique feature names that meet the
+            threshold criteria across all found files.
+    """
+    if verbose:
+        _LOGGER.info(f"Starting feature selection with SHAP threshold >= {shap_threshold}")
+    root_path = make_fullpath(root_directory, enforce="directory")
+
+    # --- Step 2: Directory and File Discovery ---
+    subdirectories = list_subdirectories(root_dir=root_path, verbose=False, raise_on_empty=True)
+    
+    shap_filename = SHAPKeys.SAVENAME + ".csv"
+
+    valid_csv_paths = []
+    for dir_name, dir_path in subdirectories.items():
+        expected_path = dir_path / shap_filename
+        if expected_path.is_file():
+            valid_csv_paths.append(expected_path)
+        else:
+            _LOGGER.warning(f"No '{shap_filename}' found in subdirectory '{dir_name}'.")
+    
+    if not valid_csv_paths:
+        _LOGGER.error(f"Process halted: No '{shap_filename}' files were found in any subdirectory.")
+        return []
+
+    if verbose:
+        _LOGGER.info(f"Found {len(valid_csv_paths)} SHAP summary files to process.")
+
+    # --- Step 3: Data Processing and Feature Extraction ---
+    master_feature_set = set()
+    for csv_path in valid_csv_paths:
+        try:
+            df, _ = load_dataframe(csv_path, kind="pandas", verbose=False)
+            
+            # Validate required columns
+            required_cols = {SHAPKeys.FEATURE_COLUMN, SHAPKeys.SHAP_VALUE_COLUMN}
+            if not required_cols.issubset(df.columns):
+                _LOGGER.warning(f"Skipping '{csv_path}': missing required columns.")
+                continue
+
+            # Filter by threshold and extract features
+            filtered_df = df[df[SHAPKeys.SHAP_VALUE_COLUMN] >= shap_threshold]
+            features = filtered_df[SHAPKeys.FEATURE_COLUMN].tolist()
+            master_feature_set.update(features)
+
+        except (ValueError, pd.errors.EmptyDataError):
+            _LOGGER.warning(f"Skipping '{csv_path}' because it is empty or malformed.")
+            continue
+        except Exception as e:
+            _LOGGER.error(f"An unexpected error occurred while processing '{csv_path}': {e}")
+            continue
+
+    # --- Step 4: Finalize and Return ---
+    final_features = sorted(list(master_feature_set))
+    if verbose:
+        _LOGGER.info(f"Selected {len(final_features)} unique features across all files.")
+        
+    if log_feature_names_directory is not None:
+        save_names_path = make_fullpath(log_feature_names_directory, make=True, enforce="directory")
+        save_list_strings(list_strings=final_features,
+                          directory=save_names_path,
+                          filename=DatasetKeys.FEATURE_NAMES,
+                          verbose=verbose)
+    
+    return final_features
+

ml_tools/ML_utilities/_train_tools.py

@@ -0,0 +1,205 @@
+from pathlib import Path
+from typing import Union, Any, Iterable
+from torch import nn
+
+from ..serde import serialize_object_filename
+
+from .._core import get_logger
+
+
+_LOGGER = get_logger("Torch Utilities")
+
+
+__all__ = [
+    "build_optimizer_params",
+    "set_parameter_requires_grad",
+    "save_pretrained_transforms",
+]
+
+
+def build_optimizer_params(model: nn.Module, weight_decay: float = 0.01) -> list[dict[str, Any]]:
+    """
+    Groups model parameters to apply weight decay only to weights (matrices/embeddings),
+    while excluding biases and normalization parameters (scales/shifts).
+
+    This function uses a robust hybrid strategy:
+    1. It excludes parameters matching standard names (e.g., "bias", "norm").
+    2. It excludes any parameter with < 2 dimensions (vector parameters), which 
+       automatically catches unnamed BatchNorm/LayerNorm weights in Sequential containers.
+
+    Args:
+        model (nn.Module): 
+            The PyTorch model.
+        weight_decay (float): 
+            The L2 regularization coefficient for the weights. 
+            (Default: 0.01)
+
+    Returns:
+        List[Dict[str, Any]]: A list of parameter groups formatted for PyTorch optimizers.
+            - Group 0: 'params' = Weights (decay applied)
+            - Group 1: 'params' = Biases/Norms (decay = 0.0)
+    """
+    # 1. Hard-coded strings for explicit safety
+    no_decay_strings = {"bias", "LayerNorm", "BatchNorm", "GroupNorm", "norm.weight"}
+    
+    decay_params = []
+    no_decay_params = []
+    
+    # 2. Iterate only over trainable parameters
+    for name, param in model.named_parameters():
+        if not param.requires_grad:
+            continue
+            
+        # Check 1: Name match
+        is_blacklisted_name = any(nd in name for nd in no_decay_strings)
+        
+        # Check 2: Dimensionality (Robust fallback)
+        # Weights/Embeddings are 2D+, Biases/Norm Scales are 1D
+        is_1d = param.ndim < 2
+
+        if is_blacklisted_name or is_1d:
+            no_decay_params.append(param)
+        else:
+            decay_params.append(param)
+            
+    _LOGGER.info(f"Weight decay configured:\n    Decaying parameters: {len(decay_params)}\n    Non-decaying parameters: {len(no_decay_params)}")
+
+    return [
+        {
+            'params': decay_params,
+            'weight_decay': weight_decay,
+        },
+        {
+            'params': no_decay_params,
+            'weight_decay': 0.0,
+        }
+    ]
+
+
+def set_parameter_requires_grad(
+    model: nn.Module,
+    unfreeze_last_n_params: int,
+) -> int:
+    """
+    Freezes or unfreezes parameters in a model based on unfreeze_last_n_params.
+
+    - N = 0: Freezes ALL parameters.
+    - N > 0 and N < total: Freezes ALL parameters, then unfreezes the last N.
+    - N >= total: Unfreezes ALL parameters.
+
+    Note: 'N' refers to individual parameter tensors (e.g., `layer.weight`
+    or `layer.bias`), not modules or layers. For example, to unfreeze
+    the final nn.Linear layer, you would use N=2 (for its weight and bias).
+
+    Args:
+        model (nn.Module): The model to modify.
+        unfreeze_last_n_params (int):
+            The number of parameter tensors to unfreeze, starting from
+            the end of the model.
+
+    Returns:
+        int: The total number of individual parameters (elements) that were set to `requires_grad=True`.
+    """
+    if unfreeze_last_n_params < 0:
+        _LOGGER.error(f"unfreeze_last_n_params must be >= 0, but got {unfreeze_last_n_params}")
+        raise ValueError()
+
+    # --- Step 1: Get all parameter tensors ---
+    all_params = list(model.parameters())
+    total_param_tensors = len(all_params)
+
+    # --- Case 1: N = 0 (Freeze ALL parameters) ---
+    # early exit for the "freeze all" case.
+    if unfreeze_last_n_params == 0:
+        params_frozen = _set_params_grad(all_params, requires_grad=False)
+        _LOGGER.warning(f"Froze all {total_param_tensors} parameter tensors ({params_frozen} total elements).")
+        return 0  # 0 parameters unfrozen
+
+    # --- Case 2: N >= total (Unfreeze ALL parameters) ---
+    if unfreeze_last_n_params >= total_param_tensors:
+        if unfreeze_last_n_params > total_param_tensors:
+            _LOGGER.warning(f"Requested to unfreeze {unfreeze_last_n_params} params, but model only has {total_param_tensors}. Unfreezing all.")
+        
+        params_unfrozen = _set_params_grad(all_params, requires_grad=True)
+        _LOGGER.info(f"Unfroze all {total_param_tensors} parameter tensors ({params_unfrozen} total elements) for training.")
+        return params_unfrozen
+
+    # --- Case 3: 0 < N < total (Standard: Freeze all, unfreeze last N) ---
+    # Freeze ALL
+    params_frozen = _set_params_grad(all_params, requires_grad=False)
+    _LOGGER.info(f"Froze {params_frozen} parameters.")
+
+    # Unfreeze the last N
+    params_to_unfreeze = all_params[-unfreeze_last_n_params:]
+    
+    # these are all False, so the helper will set them to True
+    params_unfrozen = _set_params_grad(params_to_unfreeze, requires_grad=True)
+
+    _LOGGER.info(f"Unfroze the last {unfreeze_last_n_params} parameter tensors ({params_unfrozen} total elements) for training.")
+
+    return params_unfrozen
+
+
+def _set_params_grad(
+    params: Iterable[nn.Parameter], 
+    requires_grad: bool
+) -> int:
+    """
+    A helper function to set the `requires_grad` attribute for an iterable
+    of parameters and return the total number of elements changed.
+    """
+    params_changed = 0
+    for param in params:
+        if param.requires_grad != requires_grad:
+            param.requires_grad = requires_grad
+            params_changed += param.numel()
+    return params_changed
+
+
+def save_pretrained_transforms(model: nn.Module, output_dir: Union[str, Path]):
+    """
+    Checks a model for the 'self._pretrained_default_transforms' attribute, if found,
+    serializes the returned transform object as a .joblib file.
+    
+    Used for wrapper vision models when initialized with pre-trained weights.
+
+    This saves the callable transform object itself for
+    later use, such as passing it directly to the 'transform_source'
+    argument of the PyTorchVisionInferenceHandler.
+
+    Args:
+        model (nn.Module): The model instance to check.
+        output_dir (str | Path): The directory where the transform file will be saved.
+    """
+    output_filename = "pretrained_model_transformations"
+
+    # 1. Check for the "secret attribute"
+    if not hasattr(model, '_pretrained_default_transforms'):
+        _LOGGER.warning(f"Model of type {type(model).__name__} does not have the required attribute. No transformations saved.")
+        return
+
+    # 2. Get the transform object
+    try:
+        transform_obj = model._pretrained_default_transforms
+    except Exception as e:
+        _LOGGER.error(f"Error calling the required attribute on model: {e}")
+        return
+
+    # 3. Check if the object is actually there
+    if transform_obj is None:
+        _LOGGER.warning(f"Model {type(model).__name__} has the required attribute but returned None. No transforms saved.")
+        return
+
+    # 4. Serialize and save using serde
+    try:
+        serialize_object_filename(
+            obj=transform_obj,
+            save_dir=output_dir,
+            filename=output_filename,
+            verbose=True,
+            raise_on_error=True
+        )
+        # _LOGGER.info(f"Successfully saved pretrained transforms to '{output_dir}'.")
+    except Exception as e:
+        _LOGGER.error(f"Failed to serialize transformations: {e}")
+        raise

ml_tools/{ML_vision_transformers.py → ML_vision_transformers/__init__.py}

@@ -1,18 +1,21 @@
-from ._core._ML_vision_transformers import (
-    TRANSFORM_REGISTRY,
+from ._core_transforms import (
     ResizeAspectFill,
     LetterboxResize,
     HistogramEqualization,
     RandomHistogramEqualization,
-    create_offline_augmentations,
-    info
 )
 
+from ._offline_augmentation import create_offline_augmentations
+
+from ._imprimir import info
+
+
 __all__ = [
-    "TRANSFORM_REGISTRY",
+    # Custom Transforms
     "ResizeAspectFill",
     "LetterboxResize",
     "HistogramEqualization",
     "RandomHistogramEqualization",
-    "create_offline_augmentations"
+    # Offline Augmentation
+    "create_offline_augmentations",
 ]