Weights for model sequential have not yet been created error

I’m getting a weights error when run a simple tuner.

• ENV : tf version 22.0.0-beta1 keras 2.2.4-tf

• RUNNING on Kaggle GPU powered

class FBPDHyperModel(HyperModel):
    def __init__(self,features):
        self.features_layers = features # DenseFeatures Layer         

    def build(self, hp):
        model = tf.keras.Sequential()
        model.add(self.features_layers)
        model.add(layers.Dense(units=hp.Range('units',32, 512, 32),activation='relu'))
        model.add(layers.Dense(3, activation='softmax'))
        model.compile(optimizer=tf.keras.optimizers.Adam(),loss='sparse_categorical_crossentropy',metrics=['accuracy'])
        return model  

tuner = RandomSearch(
    FBPDHyperModel(features=features_layer),
    objective='val_accuracy',
    max_trials=5,
    directory='tuner_dir')

On running, I get this error

---

ValueError Traceback (most recent call last) <ipython-input-49-720dad76188f> in <module> 3 objective=‘val_accuracy’, 4 max_trials=5, ----> 5 directory=‘tuner_dir’)

/opt/conda/lib/python3.6/site-packages/kerastuner/tuners/randomsearch.py in init(self, hypermodel, objective, max_trials, seed, **kwargs) 120 objective, 121 max_trials, –> 122 **kwargs)

/opt/conda/lib/python3.6/site-packages/kerastuner/engine/tuner.py in init(self, oracle, hypermodel, objective, max_trials, executions_per_trial, max_model_size, optimizer, loss, metrics, hyperparameters, tune_new_entries, allow_new_entries, distribution_strategy, directory, project_name) 200 # Populate initial search space 201 if not self.hyperparameters.space and self.tune_new_entries: –> 202 self._build_model(self.hyperparameters) 203 204 def search(self, *fit_args, **fit_kwargs):

/opt/conda/lib/python3.6/site-packages/kerastuner/engine/tuner.py in _build_model(self, hp) 599 600 # Check model size. –> 601 size = utils.compute_model_size(model) 602 if self.max_model_size and size > self.max_model_size: 603 oversized_streak += 1

/opt/conda/lib/python3.6/site-packages/kerastuner/utils.py in compute_model_size(model) 21 def compute_model_size(model): 22 “comput the size of a given model” —> 23 params = [K.count_params(p) for p in set(model.trainable_weights)] 24 return int(np.sum(params)) 25

/opt/conda/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py in trainable_weights(self) 578 @property 579 def trainable_weights(self): –> 580 self._assert_weights_created() 581 return trackable_layer_utils.gather_trainable_weights( 582 trainable=self.trainable,

/opt/conda/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py in _assert_weights_created(self) 1662 'Weights are created when the Model is first called on ’ 1663 ‘inputs or build() is called with an input_shape.’ % -> 1664 self.name) 1665 1666 @property

ValueError: Weights for model sequential have not yet been created. Weights are created when the Model is first called on inputs or build() is called with an input_shape.

[omalleyt12]

I was able to repro. A workaround for now is to pass input_shape=(<my_shape>) to the first layer of the Sequential model

[franchesoni]

If the number of layers do not depend upon the hyperparameters used I still get an error.

First, following the previous exchange: If I don’t implement get_config, I find in keras/engine/network.py, line 919:

  def get_config(self):
    if not self._is_graph_network:
      raise NotImplementedError

where ```self```` is my custom model.

If I implement get_config, now the traceback becomes:

Traceback (most recent call last):
  File "/home/franchesoni/mine/les/our_research/redes/0.7/tcn_1_search.py", line 116, in <module>
    tuner.search(train_gen, validation_data=test_gen)
  File "/home/franchesoni/aminiconda3/envs/tf/lib/python3.6/site-packages/kerastuner/engine/tuner.py", line 223, in search
    self.run_trial(trial, hp, fit_args, fit_kwargs)
  File "/home/franchesoni/aminiconda3/envs/tf/lib/python3.6/site-packages/kerastuner/tuners/hyperband.py", line 326, in run_trial
    super(Hyperband, self).run_trial(trial, hp, fit_args, fit_kwargs)
  File "/home/franchesoni/aminiconda3/envs/tf/lib/python3.6/site-packages/kerastuner/engine/tuner.py", line 253, in run_trial
    self.on_execution_begin(trial, execution, model)
  File "/home/franchesoni/aminiconda3/envs/tf/lib/python3.6/site-packages/kerastuner/tuners/hyperband.py", line 321, in on_execution_begin
    model.load_weights(best_checkpoint)
  File "/home/franchesoni/aminiconda3/envs/tf/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 162, in load_weights
    return super(Model, self).load_weights(filepath, by_name)
  File "/home/franchesoni/aminiconda3/envs/tf/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 1415, in load_weights
    saving.load_weights_from_hdf5_group(f, self.layers)
  File "/home/franchesoni/aminiconda3/envs/tf/lib/python3.6/site-packages/tensorflow/python/keras/saving/hdf5_format.py", line 671, in load_weights_from_hdf5_group
    ' layers.')
ValueError: You are trying to load a weight file containing 4 layers into a model with 0 layers.

I haven’t dig in yet, layers in my model are build at __init__