Hyperopt: defining search space

Question

I'm having trouble with the logic of the search space definition.

I want to search over these:

• The type of model to use (features_and_hours, features_only, hours_only, no_features_no_hours)
• The number of hidden units (output_units)
• regularization for the kernel matrix (type = l1, l2, or l1l2)
• regularization values for the kernel matrix (anywhere from 0.0 to 0.5)
• regularization for the activities (type = l1, l2, or l1l2)
• regularization values for the activities (anywhere from 0.0 to 0.5)
• number of epochs (num_epochs, either 1, 5, or 10)
• optimizer to be used (either adadelta, adam, or rmsprop)
• whether and how to apply attention (before, after, or none)

Here's the way I've set this up, following this example (second post on page, by jacobzweig)

def para_space():

        space_paras = {'model_type': hp.choice('model_type', ['features_and_hours', 'features_only', 'hours_only', 'no_features_no_hours']),
                        'output_units': hp.uniform('output_units', 1, 10),
                        'kernel_reg': hp.choice('kernel_reg', [{'reg_type':'l1', 'reg_vals': hp.uniform('reg_vals', 0.0, 0.5)},
                                                                {'reg_type':'l2','reg_vals': hp.uniform('reg_vals', 0.0, 0.5)},
                                                                {'reg_type':'l1l2', 'reg_vals': hp.uniform('reg_vals', 0.0, 0.5)}]),
                        'activity_reg': hp.choice('activity_reg', [{'reg_type':'l1', 'reg_vals': hp.uniform('reg_vals', 0.0, 0.5)},
                                                                {'reg_type':'l2','reg_vals': hp.uniform('reg_vals', 0.0, 0.5)},
                                                                {'reg_type':'l1l2', 'reg_vals': hp.uniform('reg_vals', 0.0, 0.5)}]),
                        'num_epochs': hp.choice('num_epochs', [1, 5, 10]),
                        'optimizer': hp.choice('optimizer', ['adadelta', 'adam', 'rmsprop']),
                        'attention': hp.choice('attention', ['before', 'after', 'none'])}
        return space_paras

The error I'm getting is:

Using TensorFlow backend.
Traceback (most recent call last):
  File "Jan22Model1.py", line 374, in <module>
    best = fmin(lstm_model_1, params, algo=tpe.suggest, max_evals=5, trials=trials)
  File "/user/pkgs/anaconda2/lib/python2.7/site-packages/hyperopt/fmin.py", line 307, in fmin
    return_argmin=return_argmin,
  File "/user/pkgs/anaconda2/lib/python2.7/site-packages/hyperopt/base.py", line 635, in fmin
    return_argmin=return_argmin)
  File "/user/pkgs/anaconda2/lib/python2.7/site-packages/hyperopt/fmin.py", line 314, in fmin
    pass_expr_memo_ctrl=pass_expr_memo_ctrl)
  File "/user/pkgs/anaconda2/lib/python2.7/site-packages/hyperopt/base.py", line 776, in __init__
    raise DuplicateLabel(label)
hyperopt.exceptions.DuplicateLabel: reg_vals

But in the example, there seem to be duplicate labels that don't throw an error. What am I doing wrong?

Answer 1

I ended up restructuring the space, which solved the problem:

def para_space():
    space_paras = {'model_type': hp.choice('model_type', ['features_and_hours', 'features_only', 'hours_only', 'no_features_no_hours']),
                    'output_units': hp.uniform('output_units', 1, 10),
                    'kernel_reg': hp.choice('kernel_reg', ['l1', 'l2', 'l1_l2']),
                    'kernel_reg_value': hp.uniform('kernel_reg_value', 0.0, 0.5),
                    'activity_reg': hp.choice('activity_reg', ['l1', 'l2', 'l1_l2']),
                    'activity_reg_value': hp.uniform('activity_reg_value', 0.0, 0.5),
                     'optimizer': hp.choice('optimizer', ['adadelta', 'adam', 'rmsprop']),
                     'attention': hp.choice('attention', ['before', 'after', 'none'])} 
   return space_paras