🐛🔥 remove old code, document and fix bug in scikit-learn interface

- scikit-learn interface for CF did not specify validation split correctly

vaep/models/collab.py

@@ -5,8 +5,8 @@
 import pandas as pd
 # import explicit objects for functional annotations
 from fastai.collab import *
-from fastai.collab import (Categorify, CollabDataLoaders, IndexSplitter,
-                           TabularCollab, TransformBlock)
+from fastai.collab import (Categorify, IndexSplitter, TabularCollab,
+                           TransformBlock)
 from fastai.tabular.all import *
 
 import vaep.io.dataloaders
@@ -49,37 +49,30 @@ def __init__(self,
                  item_column: str = 'peptide',
                  target_column: str = 'intensity',
                  model_kwargs: dict = None,
-                 batch_size: int = 64):
+                 batch_size: int = 1_024):
         if datasplits.val_y is not None:
-            self.X, self.frac = combine_data(datasplits.train_X,
-                                             datasplits.val_y)
+            self.X, _ = combine_data(datasplits.train_X,
+                                     datasplits.val_y)
         else:
-            self.X, self.frac = datasplits.train_X.reset_index(), 0.0
+            self.X, _ = datasplits.train_X.reset_index(), 0.0
         self.batch_size = batch_size
-        self.dls = CollabDataLoaders.from_df(self.X, valid_pct=self.frac,
-                                             seed=42,
-                                             user_name=sample_column,
-                                             item_name=item_column,
-                                             rating_name=target_column,
-                                             bs=self.batch_size)
         user_name = sample_column
         item_name = item_column
         rating_name = target_column
         cat_names = [user_name, item_name]
-        ratings = self.X
         splits = None
         if datasplits.val_y is not None:
             idx_splitter = IndexSplitter(
-                list(range(len(datasplits.train_X), len(datasplits.train_X) + len(datasplits.val_y))))
+                list(range(len(datasplits.train_X), len(self.X))))
             splits = idx_splitter(self.X)
-        to = TabularCollab(
-            ratings,
-            [Categorify],
-            cat_names,
+        self.to = TabularCollab(
+            self.X,
+            procs=[Categorify],
+            cat_names=cat_names,
             y_names=[rating_name],
             y_block=TransformBlock(),
             splits=splits)
-        self.dls = to.dataloaders(path='.', bs=self.batch_size)
+        self.dls = self.to.dataloaders(path='.', bs=self.batch_size)
         self.params = {}
         if model_kwargs is None:
             model_kwargs = {}

vaep/sklearn/cf_transformer.py

@@ -8,12 +8,13 @@
 from fastai import learner
 from fastai.callback.tracker import EarlyStoppingCallback
 from fastai.collab import *
-from fastai.collab import CollabDataLoaders, EmbeddingDotBias, TabularCollab
+from fastai.collab import EmbeddingDotBias, TabularCollab
 from fastai.data.block import TransformBlock
 from fastai.data.transforms import IndexSplitter
 from fastai.learner import Learner
 from fastai.losses import MSELossFlat
 from fastai.tabular.all import *
+from fastai.tabular.all import TransformBlock
 from fastai.tabular.core import Categorify
 from fastai.torch_core import default_device
 from sklearn.base import BaseEstimator, TransformerMixin
@@ -89,21 +90,16 @@ def fit(self, X: pd.Series, y: pd.Series = None,
         if not cuda:
             default_device(use=False)  # set to cpu
         if y is not None:
-            X, frac = collab.combine_data(X, y)
+            # Concatenate train and validation observations into on dataframe
+            first_N_train = len(X)
+            X, _ = collab.combine_data(X, y)
         else:
-            X, frac = X.reset_index(), 0.0
-
-        self.dls = CollabDataLoaders.from_df(
-            X,
-            valid_pct=frac,
-            seed=42,
-            user_name=self.sample_column,
-            item_name=self.item_column,
-            rating_name=self.target_column,
-            bs=self.batch_size)
+            X, _ = X.reset_index(), 0.0
+
         splits = None
         if y is not None:
-            idx_splitter = IndexSplitter(list(range(len(X), len(X) + len(y))))
+            # specify positional indices of validation data
+            idx_splitter = IndexSplitter(list(range(first_N_train, len(X))))
             splits = idx_splitter(X)
 
         self.cat_names = [self.sample_column, self.item_column]