flowyml 1.7.1__py3-none-any.whl → 1.7.2__py3-none-any.whl

flowyml/assets/model.py

@@ -1,20 +1,755 @@
-"""Model Asset - Represents ML models with metadata and lineage."""
+"""Model Asset - Represents ML models with automatic metadata extraction."""
 
 from typing import Any
+import contextlib
+import logging
+
 from flowyml.assets.base import Asset
 
+logger = logging.getLogger(__name__)
+
+
+class ModelInspector:
+    """Utility class for extracting model metadata from various frameworks.
+
+    This class provides robust, fail-safe extraction of model metadata.
+    It will never raise exceptions - if extraction fails, it returns
+    partial results with whatever could be extracted.
+
+    Supported frameworks with rich extraction:
+        - Keras/TensorFlow: Full architecture, optimizer, loss, metrics
+        - PyTorch: Architecture, parameters, layer info
+        - Scikit-learn: Hyperparameters, feature importance
+        - XGBoost/LightGBM/CatBoost: Trees, hyperparameters
+
+    All other model types are supported with basic metadata (type name, etc.)
+    """
+
+    @staticmethod
+    def detect_framework(model: Any) -> str | None:
+        """Detect the ML framework of a model.
+
+        Returns one of: 'keras', 'tensorflow', 'pytorch', 'sklearn', 'xgboost',
+        'lightgbm', 'catboost', 'huggingface', 'onnx', 'custom', or None
+
+        This method is safe to call on any object and will never raise.
+        """
+        if model is None:
+            return None
+
+        try:
+            type_name = type(model).__name__
+            module_name = type(model).__module__
+
+            # Keras/TensorFlow - check multiple indicators
+            if any(x in module_name.lower() for x in ["keras", "tf.keras"]):
+                return "keras"
+            if type_name in ("Sequential", "Functional", "Model") and "tensorflow" in module_name.lower():
+                return "keras"
+            if "tensorflow" in module_name.lower():
+                return "tensorflow"
+
+            # PyTorch - check for nn.Module inheritance (handles user-defined models)
+            # Check if any base class is from torch
+            try:
+                for base in type(model).__mro__:
+                    base_module = base.__module__
+                    if "torch" in base_module.lower():
+                        return "pytorch"
+            except Exception:
+                pass
+
+            # Also check module name directly (for torch tensors, etc.)
+            if "torch" in module_name.lower():
+                return "pytorch"
+
+            # Check for PyTorch-specific attributes
+            if hasattr(model, "forward") and hasattr(model, "parameters") and hasattr(model, "state_dict"):
+                # Likely a PyTorch model
+                return "pytorch"
+
+            # Scikit-learn - check for common base classes
+            if "sklearn" in module_name.lower():
+                return "sklearn"
+
+            # XGBoost
+            if "xgboost" in module_name.lower() or type_name.startswith("XGB"):
+                return "xgboost"
+
+            # LightGBM
+            if "lightgbm" in module_name.lower() or type_name.startswith("LGB"):
+                return "lightgbm"
+
+            # CatBoost
+            if "catboost" in module_name.lower():
+                return "catboost"
+
+            # Hugging Face Transformers
+            if "transformers" in module_name.lower():
+                return "huggingface"
+
+            # ONNX
+            if "onnx" in module_name.lower():
+                return "onnx"
+
+            # JAX/Flax
+            if "flax" in module_name.lower() or "jax" in module_name.lower():
+                return "jax"
+
+            # Check for common ML model attributes
+            if hasattr(model, "predict") and hasattr(model, "fit"):
+                return "sklearn"  # Sklearn-like API
+
+            return "custom"
+
+        except Exception as e:
+            logger.debug(f"Error detecting framework: {e}")
+            return "unknown"
+
+    @staticmethod
+    def extract_keras_info(model: Any) -> dict[str, Any]:
+        """Extract metadata from a Keras/TensorFlow model.
+
+        This method is robust and will extract as much as possible,
+        even if some attributes are not available (e.g., uncompiled model).
+        """
+        result = {"framework": "keras", "_auto_extracted": True}
+
+        # Parameter count - handle both built and unbuilt models
+        try:
+            if hasattr(model, "count_params"):
+                with contextlib.suppress(Exception):
+                    result["parameters"] = model.count_params()
+        except Exception as e:
+            logger.debug(f"Error getting param count: {e}")
+
+        # Trainable parameters
+        try:
+            if hasattr(model, "trainable_weights") and model.trainable_weights:
+                trainable = sum(int(w.numpy().size) if hasattr(w, "numpy") else 0 for w in model.trainable_weights)
+                if trainable > 0:
+                    result["trainable_parameters"] = trainable
+        except Exception as e:
+            logger.debug(f"Error getting trainable params: {e}")
+
+        # Architecture name
+        try:
+            if hasattr(model, "name") and model.name:
+                result["architecture"] = model.name
+            result["model_class"] = type(model).__name__
+        except Exception as e:
+            logger.debug(f"Error getting architecture: {e}")
+
+        # Layer info - handle models without layers attribute
+        try:
+            if hasattr(model, "layers"):
+                layers = model.layers
+                if layers:
+                    result["num_layers"] = len(layers)
+                    result["layer_types"] = list({type(layer).__name__ for layer in layers})
+        except Exception as e:
+            logger.debug(f"Error getting layer info: {e}")
+
+        # Input/Output shapes - handle unbuilt models
+        with contextlib.suppress(Exception):
+            if hasattr(model, "input_shape"):
+                input_shape = model.input_shape
+                if input_shape is not None:
+                    result["input_shape"] = str(input_shape)
+
+        with contextlib.suppress(Exception):
+            if hasattr(model, "output_shape"):
+                output_shape = model.output_shape
+                if output_shape is not None:
+                    result["output_shape"] = str(output_shape)
+
+        # Optimizer info (only for compiled models)
+        try:
+            if hasattr(model, "optimizer") and model.optimizer is not None:
+                opt = model.optimizer
+                result["optimizer"] = type(opt).__name__
+
+                # Get learning rate - handle different Keras versions
+                if hasattr(opt, "learning_rate"):
+                    lr = opt.learning_rate
+                    if hasattr(lr, "numpy"):
+                        lr = float(lr.numpy())
+                    elif callable(lr):
+                        # Learning rate schedule
+                        result["lr_schedule"] = type(lr).__name__
+                    else:
+                        lr = float(lr)
+                    if isinstance(lr, (int, float)):
+                        result["learning_rate"] = lr
+                elif hasattr(opt, "lr"):
+                    # Older Keras versions
+                    result["learning_rate"] = float(opt.lr)
+        except Exception as e:
+            logger.debug(f"Error getting optimizer info: {e}")
+
+        # Loss function (only for compiled models)
+        try:
+            if hasattr(model, "loss") and model.loss is not None:
+                loss = model.loss
+                if isinstance(loss, str):
+                    result["loss_function"] = loss
+                elif hasattr(loss, "__name__"):
+                    result["loss_function"] = loss.__name__
+                elif hasattr(loss, "name"):
+                    result["loss_function"] = loss.name
+                elif hasattr(loss, "__class__"):
+                    result["loss_function"] = type(loss).__name__
+        except Exception as e:
+            logger.debug(f"Error getting loss function: {e}")
+
+        # Metrics (only for compiled models)
+        try:
+            if hasattr(model, "metrics_names") and model.metrics_names:
+                result["metrics"] = list(model.metrics_names)
+            elif hasattr(model, "metrics") and model.metrics:
+                result["metrics"] = [m.name if hasattr(m, "name") else str(m) for m in model.metrics]
+        except Exception as e:
+            logger.debug(f"Error getting metrics: {e}")
+
+        # Check if model is compiled
+        with contextlib.suppress(Exception):
+            result["is_compiled"] = hasattr(model, "optimizer") and model.optimizer is not None
+
+        # Check if model is built
+        with contextlib.suppress(Exception):
+            if hasattr(model, "built"):
+                result["is_built"] = model.built
+
+        return result
+
+    @staticmethod
+    def extract_pytorch_info(model: Any) -> dict[str, Any]:
+        """Extract metadata from a PyTorch model.
+
+        This method is robust and handles:
+        - nn.Module models
+        - Custom modules
+        - Pretrained models from torchvision, transformers, etc.
+        - Models in eval or train mode
+        """
+        result = {"framework": "pytorch", "_auto_extracted": True}
+
+        # Model class name
+        try:
+            result["model_class"] = type(model).__name__
+            result["architecture"] = type(model).__name__
+        except Exception:
+            pass
+
+        # Parameter count
+        try:
+            if hasattr(model, "parameters"):
+                params = list(model.parameters())
+                if params:
+                    total_params = sum(p.numel() for p in params)
+                    trainable_params = sum(p.numel() for p in params if p.requires_grad)
+                    result["parameters"] = total_params
+                    result["trainable_parameters"] = trainable_params
+                    result["frozen_parameters"] = total_params - trainable_params
+        except Exception as e:
+            logger.debug(f"Error getting PyTorch param count: {e}")
+
+        # Layer info from modules
+        try:
+            if hasattr(model, "modules"):
+                modules = list(model.modules())
+                if modules:
+                    # Skip the first module (the model itself)
+                    layer_modules = modules[1:] if len(modules) > 1 else modules
+                    result["num_layers"] = len(layer_modules)
+
+                    # Get unique layer types
+                    layer_types = set()
+                    for m in layer_modules:
+                        layer_type = type(m).__name__
+                        # Skip container modules
+                        if layer_type not in ("Sequential", "ModuleList", "ModuleDict"):
+                            layer_types.add(layer_type)
+                    result["layer_types"] = list(layer_types)
+        except Exception as e:
+            logger.debug(f"Error getting PyTorch layer info: {e}")
+
+        # Named modules for architecture insights
+        try:
+            if hasattr(model, "named_modules"):
+                named = dict(model.named_modules())
+                if named:
+                    result["num_named_modules"] = len(named)
+        except Exception:
+            pass
+
+        # State dict info
+        try:
+            if hasattr(model, "state_dict"):
+                state_dict = model.state_dict()
+                if state_dict:
+                    result["num_tensors"] = len(state_dict)
+                    # Get tensor shapes for key layers
+                    tensor_shapes = {}
+                    for name, tensor in list(state_dict.items())[:10]:  # First 10 only
+                        tensor_shapes[name] = list(tensor.shape)
+                    result["tensor_shapes_sample"] = tensor_shapes
+        except Exception as e:
+            logger.debug(f"Error getting PyTorch state dict: {e}")
+
+        # Training mode
+        try:
+            if hasattr(model, "training"):
+                result["training_mode"] = model.training
+        except Exception:
+            pass
+
+        # Device info
+        try:
+            if hasattr(model, "parameters"):
+                first_param = next(model.parameters(), None)
+                if first_param is not None:
+                    result["device"] = str(first_param.device)
+                    result["dtype"] = str(first_param.dtype)
+        except Exception:
+            pass
+
+        # Check for common PyTorch model attributes
+        try:
+            # Input features (common in many models)
+            for attr in ["in_features", "in_channels", "input_size", "num_features"]:
+                if hasattr(model, attr):
+                    val = getattr(model, attr, None)
+                    if val is not None:
+                        result[attr] = val
+                        break
+
+            # Output features
+            for attr in ["out_features", "out_channels", "output_size", "num_classes"]:
+                if hasattr(model, attr):
+                    val = getattr(model, attr, None)
+                    if val is not None:
+                        result[attr] = val
+                        break
+        except Exception:
+            pass
+
+        return result
+
+    @staticmethod
+    def extract_sklearn_info(model: Any) -> dict[str, Any]:
+        """Extract metadata from a scikit-learn model.
+
+        Handles all sklearn estimators including:
+        - Classifiers, regressors, transformers
+        - Ensemble methods (RandomForest, GradientBoosting, etc.)
+        - Linear models
+        - Pipelines
+        """
+        result = {"framework": "sklearn", "_auto_extracted": True}
+
+        # Model type
+        try:
+            result["model_class"] = type(model).__name__
+            result["architecture"] = type(model).__name__
+        except Exception:
+            pass
+
+        # Get parameters (safe extraction)
+        try:
+            if hasattr(model, "get_params"):
+                params = model.get_params(deep=False)  # Shallow to avoid recursion
+                # Filter to serializable values
+                filtered_params = {}
+                for k, v in params.items():
+                    if v is None:
+                        continue
+                    if isinstance(v, (str, int, float, bool)):
+                        filtered_params[k] = v
+                    elif isinstance(v, (list, tuple)) and len(v) < 10:
+                        # Small lists/tuples of primitives
+                        if all(isinstance(x, (str, int, float, bool, type(None))) for x in v):
+                            filtered_params[k] = list(v)
+                if filtered_params:
+                    result["hyperparameters"] = filtered_params
+        except Exception as e:
+            logger.debug(f"Error getting sklearn params: {e}")
+
+        # Feature importance (tree-based models)
+        with contextlib.suppress(Exception):
+            if hasattr(model, "feature_importances_"):
+                importances = model.feature_importances_
+                result["has_feature_importances"] = True
+                result["num_features"] = len(importances)
+                # Store top 10 feature importances
+                if len(importances) <= 20:
+                    result["feature_importances"] = list(importances)
+
+        # Coefficients (linear models)
+        with contextlib.suppress(Exception):
+            if hasattr(model, "coef_"):
+                coef = model.coef_
+                if hasattr(coef, "shape"):
+                    result["coef_shape"] = str(coef.shape)
+                if hasattr(coef, "size") and coef.size <= 100:
+                    result["num_coefficients"] = int(coef.size)
+
+        # Intercept
+        with contextlib.suppress(Exception):
+            if hasattr(model, "intercept_"):
+                intercept = model.intercept_
+                if hasattr(intercept, "tolist"):
+                    intercept = intercept.tolist()
+                if isinstance(intercept, (int, float)) or (isinstance(intercept, list) and len(intercept) <= 10):
+                    result["intercept"] = intercept
+
+        # Classes (classifiers)
+        try:
+            if hasattr(model, "classes_"):
+                classes = model.classes_
+                result["num_classes"] = len(classes)
+                if len(classes) <= 20:
+                    # Convert to list if numpy array
+                    if hasattr(classes, "tolist"):
+                        classes = classes.tolist()
+                    result["classes"] = list(classes)
+        except Exception:
+            pass
+
+        # Number of estimators (ensemble models)
+        try:
+            if hasattr(model, "n_estimators"):
+                result["n_estimators"] = model.n_estimators
+            if hasattr(model, "estimators_"):
+                result["num_estimators_fitted"] = len(model.estimators_)
+        except Exception:
+            pass
+
+        # Tree-specific attributes
+        try:
+            if hasattr(model, "max_depth"):
+                result["max_depth"] = model.max_depth
+            if hasattr(model, "n_features_in_"):
+                result["n_features_in"] = model.n_features_in_
+        except Exception:
+            pass
+
+        # Check if fitted
+        try:
+            # Common sklearn pattern for checking if fitted
+            from sklearn.utils.validation import check_is_fitted
+
+            check_is_fitted(model)
+            result["is_fitted"] = True
+        except Exception:
+            result["is_fitted"] = False
+
+        return result
+
+    @staticmethod
+    def extract_xgboost_info(model: Any) -> dict[str, Any]:
+        """Extract metadata from an XGBoost model."""
+        result = {"framework": "xgboost", "_auto_extracted": True}
+
+        try:
+            result["model_class"] = type(model).__name__
+            result["architecture"] = "XGBoost"
+        except Exception:
+            pass
+
+        # Get hyperparameters
+        try:
+            if hasattr(model, "get_params"):
+                params = model.get_params()
+                result["hyperparameters"] = {
+                    k: v for k, v in params.items() if v is not None and isinstance(v, (str, int, float, bool))
+                }
+        except Exception as e:
+            logger.debug(f"Error getting XGBoost params: {e}")
+
+        # Booster info
+        try:
+            if hasattr(model, "get_booster"):
+                booster = model.get_booster()
+                if hasattr(booster, "num_trees"):
+                    result["num_trees"] = booster.num_trees()
+                if hasattr(booster, "num_features"):
+                    result["num_features"] = booster.num_features()
+            elif hasattr(model, "n_estimators"):
+                result["n_estimators"] = model.n_estimators
+        except Exception:
+            pass
+
+        # Feature importance
+        try:
+            if hasattr(model, "feature_importances_"):
+                result["num_features"] = len(model.feature_importances_)
+                result["has_feature_importances"] = True
+        except Exception:
+            pass
+
+        # Best iteration (for early stopping)
+        try:
+            if hasattr(model, "best_iteration"):
+                result["best_iteration"] = model.best_iteration
+            if hasattr(model, "best_score"):
+                result["best_score"] = model.best_score
+        except Exception:
+            pass
+
+        return result
+
+    @staticmethod
+    def extract_lightgbm_info(model: Any) -> dict[str, Any]:
+        """Extract metadata from a LightGBM model."""
+        result = {"framework": "lightgbm", "_auto_extracted": True}
+
+        try:
+            result["model_class"] = type(model).__name__
+            result["architecture"] = "LightGBM"
+        except Exception:
+            pass
+
+        # Get hyperparameters
+        try:
+            if hasattr(model, "get_params"):
+                params = model.get_params()
+                result["hyperparameters"] = {
+                    k: v for k, v in params.items() if v is not None and isinstance(v, (str, int, float, bool))
+                }
+        except Exception:
+            pass
+
+        # Booster info
+        try:
+            if hasattr(model, "booster_"):
+                booster = model.booster_
+                if hasattr(booster, "num_trees"):
+                    result["num_trees"] = booster.num_trees()
+            elif hasattr(model, "n_estimators"):
+                result["n_estimators"] = model.n_estimators
+        except Exception:
+            pass
+
+        # Feature importance
+        try:
+            if hasattr(model, "feature_importances_"):
+                result["num_features"] = len(model.feature_importances_)
+                result["has_feature_importances"] = True
+        except Exception:
+            pass
+
+        # Best iteration
+        try:
+            if hasattr(model, "best_iteration_"):
+                result["best_iteration"] = model.best_iteration_
+            if hasattr(model, "best_score_"):
+                result["best_score"] = model.best_score_
+        except Exception:
+            pass
+
+        return result
+
+    @staticmethod
+    def extract_huggingface_info(model: Any) -> dict[str, Any]:
+        """Extract metadata from a Hugging Face Transformers model."""
+        result = {"framework": "huggingface", "_auto_extracted": True}
+
+        try:
+            result["model_class"] = type(model).__name__
+            result["architecture"] = type(model).__name__
+        except Exception:
+            pass
+
+        # Config info
+        try:
+            if hasattr(model, "config"):
+                config = model.config
+                result["model_type"] = getattr(config, "model_type", None)
+                result["hidden_size"] = getattr(config, "hidden_size", None)
+                result["num_attention_heads"] = getattr(config, "num_attention_heads", None)
+                result["num_hidden_layers"] = getattr(config, "num_hidden_layers", None)
+                result["vocab_size"] = getattr(config, "vocab_size", None)
+                # Clean up None values
+                result = {k: v for k, v in result.items() if v is not None}
+        except Exception:
+            pass
+
+        # Parameter count
+        try:
+            if hasattr(model, "num_parameters"):
+                result["parameters"] = model.num_parameters()
+            elif hasattr(model, "parameters"):
+                result["parameters"] = sum(p.numel() for p in model.parameters())
+        except Exception:
+            pass
+
+        # Device
+        try:
+            if hasattr(model, "device"):
+                result["device"] = str(model.device)
+        except Exception:
+            pass
+
+        return result
+
+    @staticmethod
+    def extract_generic_info(model: Any) -> dict[str, Any]:
+        """Extract basic metadata from any model type.
+
+        This is the fallback for unknown/custom models.
+        """
+        result = {"framework": "custom", "_auto_extracted": True}
+
+        try:
+            result["model_class"] = type(model).__name__
+            result["module"] = type(model).__module__
+        except Exception:
+            pass
+
+        # Check for common model attributes
+        try:
+            # Fit/predict API (sklearn-like)
+            result["has_fit"] = hasattr(model, "fit")
+            result["has_predict"] = hasattr(model, "predict")
+            result["has_transform"] = hasattr(model, "transform")
+
+            # Parameters
+            if hasattr(model, "get_params"):
+                result["has_get_params"] = True
+
+            # State dict (PyTorch-like)
+            if hasattr(model, "state_dict"):
+                result["has_state_dict"] = True
+        except Exception:
+            pass
+
+        return result
+
+    @staticmethod
+    def extract_info(model: Any) -> dict[str, Any]:
+        """Auto-detect framework and extract model metadata.
+
+        This method is the main entry point for model metadata extraction.
+        It is designed to NEVER fail - if extraction fails for any reason,
+        it returns a minimal result with whatever could be extracted.
+
+        Args:
+            model: Any model object from any framework
+
+        Returns:
+            Dict with extracted metadata. Always includes:
+            - framework: Detected framework name
+            - _auto_extracted: Whether extraction succeeded
+            - model_class: Class name of the model (if available)
+        """
+        # Handle None model
+        if model is None:
+            return {"framework": None, "_auto_extracted": False, "error": "Model is None"}
+
+        # Detect framework
+        framework = ModelInspector.detect_framework(model)
+
+        # Extract based on framework
+        try:
+            if framework == "keras":
+                return ModelInspector.extract_keras_info(model)
+            elif framework == "tensorflow":
+                # TensorFlow models that aren't Keras
+                result = ModelInspector.extract_keras_info(model)
+                result["framework"] = "tensorflow"
+                return result
+            elif framework == "pytorch":
+                return ModelInspector.extract_pytorch_info(model)
+            elif framework == "sklearn":
+                return ModelInspector.extract_sklearn_info(model)
+            elif framework == "xgboost":
+                return ModelInspector.extract_xgboost_info(model)
+            elif framework == "lightgbm":
+                return ModelInspector.extract_lightgbm_info(model)
+            elif framework == "catboost":
+                # CatBoost is similar to XGBoost
+                result = ModelInspector.extract_xgboost_info(model)
+                result["framework"] = "catboost"
+                return result
+            elif framework == "huggingface":
+                return ModelInspector.extract_huggingface_info(model)
+            elif framework in ("jax", "onnx"):
+                # Basic extraction for JAX/ONNX
+                result = ModelInspector.extract_generic_info(model)
+                result["framework"] = framework
+                return result
+            else:
+                # Unknown or custom framework - use generic extraction
+                return ModelInspector.extract_generic_info(model)
+
+        except Exception as e:
+            # If all else fails, return minimal info
+            logger.debug(f"Error extracting model info: {e}")
+            return {
+                "framework": framework or "unknown",
+                "model_class": type(model).__name__,
+                "_auto_extracted": False,
+                "_extraction_error": str(e),
+            }
+
+    @staticmethod
+    def extract_training_history_from_callback(callback: Any) -> dict[str, list] | None:
+        """Extract training history from a FlowyML callback or Keras History object."""
+        try:
+            # FlowyML callback
+            if hasattr(callback, "get_training_history"):
+                return callback.get_training_history()
+
+            # Keras History object
+            if hasattr(callback, "history"):
+                history = callback.history
+                if isinstance(history, dict):
+                    # Add epochs if not present
+                    if "epochs" not in history and history:
+                        first_key = next(iter(history.keys()))
+                        history["epochs"] = list(range(1, len(history[first_key]) + 1))
+                    return history
+
+            # Training history dict directly
+            if isinstance(callback, dict) and any(isinstance(v, list) for v in callback.values()):
+                return callback
+
+        except Exception as e:
+            logger.debug(f"Error extracting training history: {e}")
+
+        return None
+
 
 class Model(Asset):
-    """Model asset with training metadata and lineage.
+    """Model asset with automatic metadata extraction and training history.
+
+    The Model class automatically extracts metadata from various ML frameworks,
+    reducing boilerplate code and improving UX. It also captures training history
+    for visualization in the FlowyML dashboard.
+
+    Supported frameworks:
+        - Keras/TensorFlow: Auto-extracts layers, parameters, optimizer, loss
+        - PyTorch: Auto-extracts modules, parameters, training mode
+        - Scikit-learn: Auto-extracts hyperparameters, feature importance
+        - XGBoost/LightGBM: Auto-extracts trees, hyperparameters
 
     Example:
-        >>> model = Model(
-        ...     name="resnet50_v1",
-        ...     version="v1.0.0",
-        ...     data=trained_model,
-        ...     architecture="resnet50",
-        ...     framework="pytorch",
-        ...     properties={"params": 25_557_032},
+        >>> # Minimal usage - properties auto-extracted!
+        >>> model_asset = Model.create(
+        ...     data=trained_keras_model,
+        ...     name="my_model",
+        ... )
+        >>> print(model_asset.parameters)  # Auto-extracted
+        >>> print(model_asset.framework)  # Auto-detected
+
+        >>> # With FlowyML callback - training history auto-captured
+        >>> callback = FlowymlKerasCallback(experiment_name="demo")
+        >>> model.fit(X, y, callbacks=[callback])
+        >>> model_asset = Model.create(
+        ...     data=model,
+        ...     name="trained_model",
+        ...     flowyml_callback=callback,  # Auto-extracts training history!
         ... )
     """
 
@@ -31,35 +766,302 @@ class Model(Asset):
         parent: Asset | None = None,
         tags: dict[str, str] | None = None,
         properties: dict[str, Any] | None = None,
+        training_history: dict[str, list] | None = None,
+        auto_extract: bool = True,
     ):
+        """Initialize Model with automatic metadata extraction.
+
+        Args:
+            name: Model name
+            version: Version string
+            data: The model object (Keras, PyTorch, sklearn, etc.)
+            architecture: Architecture name (auto-detected if not provided)
+            framework: Framework name (auto-detected if not provided)
+            input_shape: Input shape (auto-detected for Keras)
+            output_shape: Output shape (auto-detected for Keras)
+            trained_on: Dataset this model was trained on
+            parent: Parent asset for lineage
+            tags: Metadata tags
+            properties: Additional properties (merged with auto-extracted)
+            training_history: Training metrics per epoch
+            auto_extract: Whether to auto-extract model metadata
+        """
+        # Initialize properties
+        final_properties = properties.copy() if properties else {}
+
+        # Auto-extract model metadata if enabled
+        if auto_extract and data is not None:
+            extracted = ModelInspector.extract_info(data)
+            # Merge - user-provided values take precedence
+            for key, value in extracted.items():
+                if key not in final_properties:
+                    final_properties[key] = value
+
+            # Set framework from extracted if not provided
+            if framework is None and "framework" in extracted:
+                framework = extracted["framework"]
+
+            # Set architecture from extracted if not provided
+            if architecture is None and "architecture" in extracted:
+                architecture = extracted["architecture"]
+
         super().__init__(
             name=name,
             version=version,
             data=data,
             parent=parent,
             tags=tags,
-            properties=properties,
+            properties=final_properties,
         )
 
         self.architecture = architecture
         self.framework = framework
         self.input_shape = input_shape
         self.output_shape = output_shape
+        self.training_history = training_history
 
         # Track training dataset
         if trained_on:
             self.parents.append(trained_on)
             trained_on.children.append(self)
 
-        # Add model-specific properties
+        # Add model-specific properties (explicit ones override extracted)
         if architecture:
             self.metadata.properties["architecture"] = architecture
         if framework:
             self.metadata.properties["framework"] = framework
         if input_shape:
-            self.metadata.properties["input_shape"] = input_shape
+            self.metadata.properties["input_shape"] = str(input_shape)
         if output_shape:
-            self.metadata.properties["output_shape"] = output_shape
+            self.metadata.properties["output_shape"] = str(output_shape)
+
+    @classmethod
+    def create(
+        cls,
+        data: Any,
+        name: str | None = None,
+        version: str | None = None,
+        parent: "Asset | None" = None,
+        flowyml_callback: Any = None,
+        keras_history: Any = None,
+        auto_extract: bool = True,
+        **kwargs: Any,
+    ) -> "Model":
+        """Create a Model asset with automatic metadata extraction.
+
+        This is the preferred way to create Model objects. Metadata is
+        automatically extracted from the model, and training history can
+        be captured from FlowyML callbacks.
+
+        Args:
+            data: The model object (Keras, PyTorch, sklearn, etc.)
+            name: Asset name (auto-generated if not provided)
+            version: Asset version
+            parent: Parent asset for lineage
+            flowyml_callback: FlowymlKerasCallback for auto-capturing training history
+            keras_history: Keras History object from model.fit()
+            auto_extract: Whether to auto-extract model metadata
+            **kwargs: Additional parameters including:
+                - training_history: Dict of training metrics per epoch
+                - architecture: Model architecture name
+                - framework: ML framework (keras, pytorch, etc.)
+                - properties: Additional properties
+                - tags: Metadata tags
+
+        Returns:
+            New Model instance with auto-extracted metadata
+
+        Example:
+            >>> # Simple usage - everything auto-extracted
+            >>> model_asset = Model.create(data=model, name="my_model")
+
+            >>> # With FlowyML callback
+            >>> callback = FlowymlKerasCallback(experiment_name="demo")
+            >>> model.fit(X, y, callbacks=[callback])
+            >>> model_asset = Model.create(
+            ...     data=model,
+            ...     name="trained_model",
+            ...     flowyml_callback=callback,
+            ... )
+
+            >>> # With Keras History
+            >>> history = model.fit(X, y)
+            >>> model_asset = Model.create(
+            ...     data=model,
+            ...     name="trained_model",
+            ...     keras_history=history,
+            ... )
+        """
+        from datetime import datetime
+
+        asset_name = name or f"Model_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
+
+        # Extract Model-specific parameters
+        training_history = kwargs.pop("training_history", None)
+        architecture = kwargs.pop("architecture", None)
+        framework = kwargs.pop("framework", None)
+        input_shape = kwargs.pop("input_shape", None)
+        output_shape = kwargs.pop("output_shape", None)
+        trained_on = kwargs.pop("trained_on", None)
+
+        # Auto-extract training history from callback or history object
+        if training_history is None:
+            if flowyml_callback is not None:
+                training_history = ModelInspector.extract_training_history_from_callback(
+                    flowyml_callback,
+                )
+            elif keras_history is not None:
+                training_history = ModelInspector.extract_training_history_from_callback(
+                    keras_history,
+                )
+
+        # Extract tags and properties
+        tags = kwargs.pop("tags", {})
+        props = kwargs.pop("properties", {})
+        # Merge remaining kwargs into properties
+        props.update(kwargs)
+
+        return cls(
+            name=asset_name,
+            version=version,
+            data=data,
+            architecture=architecture,
+            framework=framework,
+            input_shape=input_shape,
+            output_shape=output_shape,
+            trained_on=trained_on,
+            parent=parent,
+            tags=tags,
+            properties=props,
+            training_history=training_history,
+            auto_extract=auto_extract,
+        )
+
+    @classmethod
+    def from_keras(
+        cls,
+        model: Any,
+        name: str | None = None,
+        callback: Any = None,
+        history: Any = None,
+        **kwargs: Any,
+    ) -> "Model":
+        """Create a Model asset from a Keras model with full auto-extraction.
+
+        Args:
+            model: Keras model object
+            name: Asset name
+            callback: FlowymlKerasCallback for training history
+            history: Keras History object from model.fit()
+            **kwargs: Additional properties
+
+        Returns:
+            Model asset with auto-extracted Keras metadata
+        """
+        return cls.create(
+            data=model,
+            name=name,
+            framework="keras",
+            flowyml_callback=callback,
+            keras_history=history,
+            **kwargs,
+        )
+
+    @classmethod
+    def from_pytorch(
+        cls,
+        model: Any,
+        name: str | None = None,
+        training_history: dict | None = None,
+        **kwargs: Any,
+    ) -> "Model":
+        """Create a Model asset from a PyTorch model with full auto-extraction.
+
+        Args:
+            model: PyTorch model object (nn.Module)
+            name: Asset name
+            training_history: Training metrics dict
+            **kwargs: Additional properties
+
+        Returns:
+            Model asset with auto-extracted PyTorch metadata
+        """
+        return cls.create(
+            data=model,
+            name=name,
+            framework="pytorch",
+            training_history=training_history,
+            **kwargs,
+        )
+
+    @classmethod
+    def from_sklearn(
+        cls,
+        model: Any,
+        name: str | None = None,
+        **kwargs: Any,
+    ) -> "Model":
+        """Create a Model asset from a scikit-learn model with full auto-extraction.
+
+        Args:
+            model: Scikit-learn model object
+            name: Asset name
+            **kwargs: Additional properties
+
+        Returns:
+            Model asset with auto-extracted sklearn metadata
+        """
+        return cls.create(
+            data=model,
+            name=name,
+            framework="sklearn",
+            **kwargs,
+        )
+
+    @property
+    def parameters(self) -> int | None:
+        """Get number of model parameters (auto-extracted)."""
+        return self.metadata.properties.get("parameters") or self.metadata.properties.get("params")
+
+    @property
+    def trainable_parameters(self) -> int | None:
+        """Get number of trainable parameters (auto-extracted)."""
+        return self.metadata.properties.get("trainable_parameters")
+
+    @property
+    def num_layers(self) -> int | None:
+        """Get number of layers (auto-extracted)."""
+        return self.metadata.properties.get("num_layers")
+
+    @property
+    def layer_types(self) -> list[str] | None:
+        """Get list of layer types (auto-extracted)."""
+        return self.metadata.properties.get("layer_types")
+
+    @property
+    def optimizer(self) -> str | None:
+        """Get optimizer name (auto-extracted from Keras)."""
+        return self.metadata.properties.get("optimizer")
+
+    @property
+    def learning_rate(self) -> float | None:
+        """Get learning rate (auto-extracted from Keras)."""
+        return self.metadata.properties.get("learning_rate")
+
+    @property
+    def loss_function(self) -> str | None:
+        """Get loss function (auto-extracted from Keras)."""
+        return self.metadata.properties.get("loss_function")
+
+    @property
+    def metrics(self) -> list[str] | None:
+        """Get metrics (auto-extracted from Keras)."""
+        return self.metadata.properties.get("metrics")
+
+    @property
+    def hyperparameters(self) -> dict | None:
+        """Get hyperparameters (auto-extracted from sklearn/xgboost)."""
+        return self.metadata.properties.get("hyperparameters")
 
     def get_training_datasets(self):
         """Get all datasets this model was trained on."""
@@ -69,7 +1071,7 @@ class Model(Asset):
 
     def get_parameters_count(self) -> int | None:
         """Get number of model parameters if available."""
-        return self.metadata.properties.get("params") or self.metadata.properties.get("parameters")
+        return self.parameters
 
     def get_architecture_info(self) -> dict[str, Any]:
         """Get architecture information."""
@@ -78,5 +1080,40 @@ class Model(Asset):
             "framework": self.framework,
             "input_shape": self.input_shape,
             "output_shape": self.output_shape,
-            "parameters": self.get_parameters_count(),
+            "parameters": self.parameters,
+            "trainable_parameters": self.trainable_parameters,
+            "num_layers": self.num_layers,
+            "layer_types": self.layer_types,
+        }
+
+    def get_training_info(self) -> dict[str, Any]:
+        """Get training information."""
+        result = {
+            "optimizer": self.optimizer,
+            "learning_rate": self.learning_rate,
+            "loss_function": self.loss_function,
+            "metrics": self.metrics,
         }
+
+        if self.training_history:
+            epochs = self.training_history.get("epochs", [])
+            result["epochs_trained"] = len(epochs)
+
+            # Get final metrics
+            for key, values in self.training_history.items():
+                if key != "epochs" and values:
+                    result[f"final_{key}"] = values[-1]
+
+        return {k: v for k, v in result.items() if v is not None}
+
+    def __repr__(self) -> str:
+        """String representation with key info."""
+        parts = [f"Model(name='{self.name}'"]
+        if self.framework:
+            parts.append(f"framework='{self.framework}'")
+        if self.parameters:
+            parts.append(f"params={self.parameters:,}")
+        if self.training_history:
+            epochs = len(self.training_history.get("epochs", []))
+            parts.append(f"epochs={epochs}")
+        return ", ".join(parts) + ")"