NMF n_components are not properly getting reflected in output when using Grid Search CV

When I executed the example code from this link and analyzed the grid search output (grid.cv_results_[‘params’]), n_components are not properly getting reflected in output.

Posting a small snippet of output of grid.cv_results_[‘params’]:

{'classify__C': 1000,
  'reduce_dim': NMF(alpha=0.0, beta_loss='frobenius', init=None, l1_ratio=0.0, max_iter=200,
    **n_components=None,** random_state=None, shuffle=False, solver='cd',
    tol=0.0001, verbose=0),
  'reduce_dim__n_components': 2},
 {'classify__C': 1000,
  'reduce_dim': NMF(alpha=0.0, beta_loss='frobenius', init=None, l1_ratio=0.0, max_iter=200,
    **n_components=None**, random_state=None, shuffle=False, solver='cd',
    tol=0.0001, verbose=0),
  'reduce_dim__n_components': 4},
 {'classify__C': 1000,
  'reduce_dim': NMF(alpha=0.0, beta_loss='frobenius', init=None, l1_ratio=0.0, max_iter=200,
    **n_components=None**, random_state=None, shuffle=False, solver='cd',
    tol=0.0001, verbose=0),
  'reduce_dim__n_components': 8},

where reduce_dim__n_components are updating for NMF but not the actual n_components in NMF

Thanks, Pat

[amueller]

Yes, but I’m concerned about the side-effects. Let’s say someone did

est = RandomForestClassifier(n_estimators=100)
grid1 = {'clf': est, 'clf__max_depth': [1, 2, 3]}
grid2 = {'clf': est, 'clf__max_leaf_nodes': [2, 3, 6]}

and then ran a grid-search for each. If you run grid1 first, the estimator in grid2 would have max_depth=3. That’s very surprising to me.

[jnothman]

another thing we should consider is whether the bug, or its fix, will affect mutable parameters other than estimators