Pipeline

takepod/__init__.py

@@ -22,7 +22,8 @@
            "models",
            "preproc",
            "storage",
-           "validation"]
+           "validation",
+           "pipeline"]
 
 
 # Reference for initialization of logging scikit-learn

takepod/examples/experiment_example.py

@@ -9,10 +9,11 @@
 from takepod.datasets.impl.pauza_dataset import PauzaHRDataset
 from takepod.models.impl.fc_model import ScikitMLPClassifier
 from takepod.models.impl.simple_trainers import SimpleTrainer
-from takepod.models import Experiment
+from takepod.models import Experiment, FeatureTransformer
 from takepod.validation import k_fold_classification_metrics
 from takepod.model_selection import grid_search
 from sklearn.metrics import accuracy_score
+from sklearn.preprocessing import StandardScaler
 
 
 def numericalize_pauza_rating(rating):
@@ -70,15 +71,16 @@ def train_iterator_provider(dataset):
     embedding_matrix = vectorizer.get_embedding_matrix(
         fields["Text"].vocab)
 
-    feature_transform = partial(feature_transform_mean_fun,
-                                embedding_matrix=embedding_matrix)
+    feature_transform_fn = partial(feature_transform_mean_fun,
+                                   embedding_matrix=embedding_matrix)
+
+    feature_transformer = FeatureTransformer(feature_transform_fn, StandardScaler())
 
     experiment = Experiment(ScikitMLPClassifier,
-                            trainer,
-                            train_iterator_provider,
-                            None,
-                            feature_transform,
-                            label_transform_fun)
+                            trainer=trainer,
+                            training_iterator_callable=train_iterator_provider,
+                            feature_transformer=feature_transformer,
+                            label_transform_fun=label_transform_fun)
 
     _, model_params, train_params = \
         grid_search(experiment,

takepod/models/__init__.py

@@ -2,9 +2,10 @@
 
 from .model import AbstractFrameworkModel, AbstractSupervisedModel
 from .batch_transform_functions import default_feature_transform, default_label_transform
+from .transformers import FeatureTransformer, TensorTransformer
 from .experiment import Experiment
 from .trainer import AbstractTrainer
 
 __all__ = ["AbstractFrameworkModel", "AbstractSupervisedModel",
            "default_feature_transform", "default_label_transform", "Experiment",
-           "AbstractTrainer"]
+           "AbstractTrainer", "FeatureTransformer", "TensorTransformer"]

takepod/models/experiment.py

@@ -1,26 +1,29 @@
 """Modules defines an experiment - class used to combine iteration over data,
 model training and prediction."""
-from typing import Callable, NamedTuple, Dict, Type
+from typing import Callable, NamedTuple, Dict, Type, Union
+from inspect import isclass
+import logging
 
 import numpy as np
 
 from takepod.datasets.dataset import Dataset
 from takepod.datasets.iterator import Iterator, SingleBatchIterator
-from takepod.models import AbstractSupervisedModel,\
-    default_feature_transform, default_label_transform
+from takepod.models import AbstractSupervisedModel, \
+    default_feature_transform, default_label_transform, FeatureTransformer
 from takepod.models.trainer import AbstractTrainer
 
+_LOGGER = logging.getLogger(__name__)
+
 
 class Experiment:
     """Class used to streamline model fitting and prediction."""
 
     def __init__(self,
-                 model_class: Type[AbstractSupervisedModel],
-                 trainer: AbstractTrainer,
-                 training_iterator_callable: Callable[[Dataset], Iterator],
+                 model: Union[Type[AbstractSupervisedModel], AbstractSupervisedModel],
+                 feature_transformer: FeatureTransformer = None,
+                 trainer: AbstractTrainer = None,
+                 training_iterator_callable: Callable[[Dataset], Iterator] = None,
                  prediction_iterator_callable: Callable[[Dataset], Iterator] = None,
-                 feature_transform_fun:
-                 Callable[[NamedTuple], np.ndarray] = None,
                  label_transform_fun:
                  Callable[[NamedTuple], np.ndarray] = None
                  ):
@@ -29,8 +32,11 @@ def __init__(self,
 
         Parameters
         ----------
-        model_class : class
-            Class of the Model to be fitted.
+        model : class or model instance
+            Class of the Model to be fitted or a pre-trained model.
+            If pre-trained model is passed and `fit` is called a new model instance will
+            be created. For fine-tuning of the passed model instance call
+            `partial_fit`.
             Must be a subclass of Podium's `AbstractSupervisedModel`
 
         trainer : AbstractTrainer
@@ -46,18 +52,24 @@ def __init__(self,
             a single tensor before being returned. If passed None, a SingleBatchIterator
             will be used as a default.
 
-        feature_transform_fun : Callable[[NamedTuple], np.ndarray]
-            Callable that transforms the input part of the batch returned by the iterator
-            into features that can be fed into the model.
+        feature_transformer : FeatureTransformer
+            FeatureTransformer that transforms the input part of the batch returned by the
+            iterator into features that can be fed into the model. Will also be fitted
+            during Experiment fitting.
 
         label_transform_fun : Callable[[NamedTuple], np.ndarray]
             Callable that transforms the target part of the batch returned by the iterator
             into the same format the model prediction is. For a hypothetical perfect model
             the prediction result of the model for some examples must be identical to the
             result of this callable for those same examples.
         """
-        self.model_class = model_class
-        self.model = None
+        if isclass(model):
+            self.model_class = model
+            self.model = None
+        else:
+            self.model_class = model.__class__
+            self.model = model
+
         self.trainer = trainer
         self.training_iterator_callable = training_iterator_callable
 
@@ -72,9 +84,9 @@ def default_prediction_iterator_callable(dataset):
         else:
             self.prediction_iterator_callable = prediction_iterator_callable
 
-        self.feature_transform_fun = feature_transform_fun \
-            if feature_transform_fun is not None \
-            else default_feature_transform
+        self.feature_transformer = feature_transformer \
+            if feature_transformer is not None \
+            else FeatureTransformer(default_feature_transform)
 
         self.label_transform_fun = label_transform_fun \
             if label_transform_fun is not None \
@@ -107,7 +119,10 @@ def set_default_trainer_args(self, **kwargs):
     def fit(self,
             dataset: Dataset,
             model_kwargs: Dict = None,
-            trainer_kwargs: Dict = None
+            trainer_kwargs: Dict = None,
+            feature_transformer: FeatureTransformer = None,
+            trainer: AbstractTrainer = None,
+            training_iterator_callable: Callable[[Dataset], Iterator] = None,
             ):
         """Fits the model to the provided Dataset. During fitting, the provided Iterator
         and Trainer are used.
@@ -126,7 +141,24 @@ def fit(self,
             Dict containing trainer arguments. Arguments passed to the trainer are the
             default arguments defined with `set_default_trainer_args` updated/overridden
             by 'trainer_kwargs'.
+
+        feature_transformer : FeatureTransformer, Optional
+            FeatureTransformer that transforms the input part of the batch returned by the
+            iterator into features that can be fed into the model. Will also be fitted
+            during Experiment fitting.
+            If None, the default FeatureTransformer provided in the constructor will be
+            used. Otherwise, this will overwrite the default feature transformer.
+
+        trainer : AbstractTrainer, Optional
+            Trainer used to fit the model. If None, the trainer provided in the
+            constructor will be used.
+
+        training_iterator_callable: Callable[[Dataset], Iterator]
+            Callable used to instantiate new instances of the Iterator used in fitting the
+            model. If None, the training_iterator_callable provided in the
+            constructor will be used.
         """
+
         model_kwargs = {} if model_kwargs is None else model_kwargs
         trainer_kwargs = {} if trainer_kwargs is None else trainer_kwargs
 
@@ -136,15 +168,91 @@ def fit(self,
         trainer_args = self.default_trainer_args.copy()
         trainer_args.update(trainer_kwargs)
 
+        trainer = trainer if trainer is not None else self.trainer
+        if trainer is None:
+            errmsg = "No trainer provided. Trainer must be provided either in the " \
+                     "constructor or as an argument to the fit method."
+            _LOGGER.error(errmsg)
+            raise RuntimeError(errmsg)
+
+        if feature_transformer is not None:
+            self.feature_transformer = feature_transformer
+
+        # Fit the feature transformer if it needs fitting
+        if self.feature_transformer.requires_fitting():
+            x_batch, y_batch = next(SingleBatchIterator(dataset).__iter__())
+            y = self.label_transform_fun(y_batch)
+            self.feature_transformer.fit(x_batch, y)
+
         # Create new model instance
         self.model = self.model_class(**model_args)
-        train_iterator = self.training_iterator_callable(dataset)
 
-        self.trainer.train(self.model,
-                           train_iterator,
-                           self.feature_transform_fun,
-                           self.label_transform_fun,
-                           **trainer_args)
+        training_iterator_callable = training_iterator_callable \
+            if training_iterator_callable is not None \
+            else self.training_iterator_callable
+
+        train_iterator = training_iterator_callable(dataset)
+        # Train the model
+        trainer.train(self.model,
+                      train_iterator,
+                      self.feature_transformer,
+                      self.label_transform_fun,
+                      **trainer_args)
+
+    def partial_fit(self,
+                    dataset: Dataset,
+                    trainer_kwargs: Dict = None,
+                    trainer: AbstractTrainer = None,
+                    training_iterator_callable: Callable[[Dataset], Iterator] = None):
+        """Fits the model to the data without resetting the model.
+
+        Parameters
+        ----------
+         dataset : Dataset
+            Dataset to fit the model to.
+
+        trainer_kwargs : dict
+            Dict containing trainer arguments. Arguments passed to the trainer are the
+            default arguments defined with `set_default_trainer_args` updated/overridden
+            by 'trainer_kwargs'.
+
+        trainer : AbstractTrainer, Optional
+            Trainer used to fit the model. If None, the trainer provided in the
+            constructor will be used.
+
+        training_iterator_callable: Callable[[Dataset], Iterator]
+            Callable used to instantiate new instances of the Iterator used in fitting the
+            model. If None, the training_iterator_callable provided in the
+            constructor will be used.
+
+        Returns
+        -------
+
+        """
+        self._check_if_model_exists()
+
+        trainer = trainer if trainer is not None else self.trainer
+        if trainer is None:
+            errmsg = "No trainer provided. Trainer must be provided either in the " \
+                     "constructor or as an argument to the partial_fit method."
+            _LOGGER.error(errmsg)
+            raise RuntimeError(errmsg)
+
+        trainer_kwargs = {} if trainer_kwargs is None else trainer_kwargs
+        trainer_args = self.default_trainer_args.copy()
+        trainer_args.update(trainer_kwargs)
+
+        training_iterator_callable = training_iterator_callable \
+            if training_iterator_callable is not None \
+            else self.training_iterator_callable
+
+        train_iterator = training_iterator_callable(dataset)
+
+        trainer.train(self.model,
+                      train_iterator,
+                      self.feature_transformer,
+                      self.label_transform_fun,
+                      **trainer_args)
 
     def predict(self,
                 dataset: Dataset,
@@ -167,13 +275,21 @@ def predict(self,
         """
         # TODO: new method of providing examples must be defined.
         # examples is taken in dataset form as proof-of-concept.
+        self._check_if_model_exists()
 
         y = []
 
         for x_batch, _ in self.prediction_iterator_callable(dataset):
-            x_batch_tensor = self.feature_transform_fun(x_batch)
+            x_batch_tensor = self.feature_transformer.transform(x_batch)
             batch_prediction = self.model.predict(x_batch_tensor, **kwargs)
             prediction_tensor = batch_prediction[AbstractSupervisedModel.PREDICTION_KEY]
             y.append(prediction_tensor)
 
         return np.concatenate(y)
+
+    def _check_if_model_exists(self):
+        if self.model is None:
+            errmsg = "Model instance not available. Please provide a model instance in " \
+                     "the constructor or call `fit` before calling `partial_fit.`"
+            _LOGGER.error(errmsg)
+            raise RuntimeError(errmsg)

takepod/models/impl/simple_trainers.py

@@ -16,13 +16,13 @@ class SimpleTrainer(AbstractTrainer):
     def train(self,
               model,
               iterator,
-              feature_transform_fun,
+              feature_transformer,
               label_transform_fun,
               **kwargs):
         self._check_kwargs(**kwargs)
         for _ in range(kwargs[SimpleTrainer.MAX_EPOCH_KEY]):
             for x_batch, y_batch in iterator:
-                x = feature_transform_fun(x_batch)
+                x = feature_transformer.transform(x_batch)
                 y = label_transform_fun(y_batch)
                 model.fit(X=x, y=y)
 

takepod/models/trainer.py

@@ -4,7 +4,7 @@
 
 import numpy as np
 
-from takepod.models import AbstractSupervisedModel
+from takepod.models import AbstractSupervisedModel, FeatureTransformer
 from takepod.datasets import Iterator
 
 
@@ -15,8 +15,7 @@ class AbstractTrainer(ABC):
     def train(self,
               model: AbstractSupervisedModel,
               iterator: Iterator,
-              feature_transform_fun:
-                  Callable[[NamedTuple], np.ndarray],
+              feature_transformer: FeatureTransformer,
               label_transform_fun:
                   Callable[[NamedTuple], np.ndarray],
               **kwargs):
@@ -27,7 +26,7 @@ def train(self,
             The model that needs to be trained.
         iterator : Iterator
             Iterator instance that provides data from a dataset
-        feature_transform_fun: Callable[[NamedTuple], np.ndarray]
+        feature_transformer: Callable[[NamedTuple], np.ndarray]
             Callable that transforms the input part of the batch returned by the iterator
             into features that can be fed into the model.
         label_transform_fun: Callable[[NamedTuple], np.ndarray]