Nested search space refinement

[leelaylay]

1. Change the directory structure (move some common class and function to utils.py)
2. Add unit test for evolution_tuner, hyperopt_tuner and utils.
3. HyperoptTuner (Random, Anneal, TPE) and EvolutionTuner support new format of nested search space.

One simple json example follows as below.

{
    "layer0": {
        "_name": "Avg_pool",
        "pooling_size": 3
    },
    "layer1": {
        "_name": "Conv",
        "kernel_size": 2
    },
    "layer2": {
        "_name": "Empty"
    },
    "layer3": {
        "_name": "Conv",
        "kernel_size": 5
    }
}

related to issue #1028 .

[PurityFan]

Well done! BTW, why Grid Search can't support nested dict?

[PurityFan]

Can you add the doc for nested dict in SearchSpaceSpec.html ?

[leelaylay]

Can you add the doc for nested dict in SearchSpaceSpec.html ?

1. OK, I will add the doc for nested dict layer in SearchSpaceSpec.doc.
2. Grid Search and some other tuners didn't seem to support this feature originally. (Maybe we can support it later?)

[xuehui1991]

Can you add the doc for nested dict in SearchSpaceSpec.html ?

1. OK, I will add the doc for nested dict layer in SearchSpaceSpec.doc.
2. Grid Search and some other tuners didn't seem to support this feature originally. (Maybe we can support it later?)

For the original version, only those tuners based on hyperopt support nested search space.

[QuanluZhang]

code refactor looks good, thanks