Easy:

• in test_common, check that the ValueError raise has a useful error message. (see sparse test for an example)
• put as many of the "specific" tests in test_clustering, test_transformers, ... into test_non_meta_estimators.

Not so easy:

• calling fit forgets the previous model if any
• check how classifiers handle only one class being present
• test how models handle non-float input (does uint8 cause overflows?)

Things done

We should factorize common tests in a new file test_common.py (or maybe test_input.py?). Things to check:

• can pickle the object
• raise an exception when data contains nans
• raise an exception for invalid input (e.g., np.matrix or sp.csr_matrix if dense only implementation)
• raise an exception if n_features is not the same in fit and predict or transform
• __repr__ and clone work
• check that we can pickle and unpickle estimators.
• check that all classifiers have a classes_ attribute (needs some fixes)

Edit by @amueller!

Edit by @GaelVaroquaux on Aug 13th 2014 to reflect progress in the codebase.