Broken hyperopts in RNN block (upd: affects all blocks, has workarounnd)

[holinov]

Bug Description

I'm having troubles shitching autokeras from 1.0.12 to 1.0.15 - hyperparmas for blocks are broken.

Bug Reproduction

Code for reproducing the bug:

    input_node = ak.Input()

    prf_regressor = input_node
    prf_regressor = ak.RNNBlock(
        return_sequences=False, bidirectional=False)(prf_regressor)
    prf_regressor = ak.RegressionHead()(prf_regressor)

and just add to any automodel. on version 1.0.12 everything worked perfectly. and on 1.0.15 i'm constantly receiving this error:

Traceback (most recent call last):
  File "predictor.py", line 274, in <module>
    run_train(config)
  File "predictor.py", line 196, in run_train
    model = ak.AutoModel(
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/auto_model.py", line 142, in __init__
    self.tuner = tuner(
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/tuners/greedy.py", line 230, in __init__
    super().__init__(oracle=oracle, hypermodel=hypermodel, **kwargs)
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/engine/tuner.py", line 54, in __init__
    self.hypermodel.hypermodel.save(os.path.join(self.project_dir, "graph"))
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/graph.py", line 331, in save
    io_utils.save_json(filepath, self.get_config())
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/graph.py", line 188, in get_config
    blocks = [blocks_module.serialize(block) for block in self.blocks]
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/graph.py", line 188, in <listcomp>
    blocks = [blocks_module.serialize(block) for block in self.blocks]
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/blocks/__init__.py", line 46, in serialize
    return tf.keras.utils.serialize_keras_object(obj)
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/tensorflow/python/keras/utils/generic_utils.py", line 503, in serialize_keras_object
    config = instance.get_config()
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/autokeras/blocks/basic.py", line 192, in get_config
    "bidirectional": hyperparameters.serialize(self.bidirectional),
  File "/home/administrator/anaconda3/envs/libra2/lib/python3.8/site-packages/keras_tuner/engine/hyperparameters.py", line 1077, in serialize
    return {"class_name": obj.__class__.__name__, "config": obj.get_config()}
AttributeError: 'bool' object has no attribute 'get_config'

Setup Details

Include the details about the versions of:

• OS type and version: Ubuntu 20.04
• Python: 3.8
• autokeras: 1.0.12, 1.0.15
• keras-tuner: 1.0.3
• scikit-learn: 0.24.2
• numpy: 1.19.5
• pandas: 1.3.1
• tensorflow: 2.5.0

[holinov]

The only working work-around i'v found

    hp = keras_tuner.HyperParameters()
    is_bi_1 = hp.Fixed('rnn_block_1/bidirectional', False)
    input_node = ak.Input()

    prf_regressor = input_node
    prf_regressor = ak.RNNBlock(
        return_sequences=False
        #, bidirectional=False <----- !!!! DO NOT PASS PARAMETER CREATED EARLIER. YOU NEED ONLY TO CREATE IT. IF YOU PASS IT - IT BREAKES
        )(prf_regressor)
    prf_regressor = ak.RegressionHead()(prf_regressor)
    model = ak.AutoModel(
        inputs=[input_node],
        outputs=[prf_regressor],
        hyperparameters=hp
    )
    ```