train.py 文件源码

def vectorize_fold(fold, tags, meta_df, use_metafeats=True):
  with time_me('Loaded pdicts'):
    scoreses = [common.pdict_for_tag(tag, fold) for tag in tags]
  df = meta_df[meta_df['fold']==fold]
  assert len(df)
  y = df['label']
  n_predictors = len(scoreses)
  with time_me('Munged scores for {} predictors'.format(n_predictors), mode='print'):
    # TODO: could use the logit loading fn added to user_wrapper module
    scores = munge_scoreses(scoreses, df)
  if not use_metafeats:
    X = scores
  else:
    meta_cols = metavectorize.metafeature_columns
    meta = df[meta_cols].values
    # Special f_0 dummy meta feature for learning vanilla weight term per predictor
    metafeats = np.hstack([np.ones( (len(df), 1) ), meta])
    # Oh fuck this, I've spent too long trying to understand np.einsum...
    # (Worth noting that sklearn.preprocessing has a 'PolynomialFeatures' utility
    # that might have been useful here. But this is fine.)
    n_metafeats = metafeats.shape[1]
    logging.info('{} predictors x {} metafeatures -> {} coefs'.format(
      n_predictors, n_metafeats, n_predictors*n_metafeats))
    # X is 'metafeat major'. i.e. the first n_p values for each vector are the 
    # raw scores for each predictor, they're followed by each predictor's score
    # multiplied by the first metafeature and so on.
    X = np.tile(scores, n_metafeats) * np.repeat(metafeats, n_predictors, axis=1)
  return X, y