What it does

Hyperas bridges Keras and Hyperopt so you can optimize dropout rates, layer sizes, optimizers, and even conditional layer stacks without writing Hyperopt’s verbose search-space syntax. You wrap your data and model in functions, drop {{uniform(0, 1)}} or {{choice([64, 128])}} where you want variation, and call optim.minimize(). Under the hood it does Jinja-style template replacement, spits out a temporary Python file, and feeds that to Hyperopt’s TPE sampler.

The interesting bit

The trick is that your create_model() function stops being valid Python the moment you add those double curly braces. Hyperas leans into this: it parses your source as text, swaps the templates for Hyperopt variables, and runs the generated code. It’s a hack, but a deliberate one that saves you from manually building Hyperopt’s space dictionary.

Key highlights

• Supports conditional architecture changes (e.g., add a fourth layer only if a choice resolves to 'four')
• Can swap entire layer blocks, not just scalar hyperparameters
• Distributed search via MongoDB by copying the generated temp_model.py to worker machines
• Data function stays separate so Hyperopt doesn’t reload your dataset on every trial
• Passes positional arguments to your data loader through data_args

Caveats

• Requires networkx==1.11 to avoid a known generator subscripting bug (issue #125)
• Jupyter notebooks need explicit notebook_name='...' parameter to avoid file-not-found errors
• The README itself warns: “If it’s not convenient to use in your situation, simply don’t use it — and choose Hyperopt instead”
• Keyword arguments to data() are not supported; only positional args that repr() can display

Verdict

Worth a look if you’re already using Keras and want to dip a toe into hyperparameter search without committing to Hyperopt’s API. Skip it if you need fine-grained control over the search space or if template-based code generation makes you nervous.