Adapt to KerasTuner 1.1.0rc0. (#1640)

Co-authored-by: Haifeng Jin <haifeng-jin@users.noreply.github.com>

Author: haifeng-jin

autokeras/auto_model.py

@@ -193,8 +193,12 @@ def _build_graph(self):
             graph = graph_module.Graph(inputs=self.inputs, outputs=self.outputs)
         # Using input/output API.
         elif all([isinstance(output, head_module.Head) for output in self.outputs]):
+            # Clear session to reset get_uid(). The names of the blocks will
+            # start to count from 1 for new blocks in a new AutoModel afterwards.
+            tf.keras.backend.clear_session()
             graph = self._assemble()
             self.outputs = graph.outputs
+            tf.keras.backend.clear_session()
 
         return graph
 
@@ -375,6 +379,7 @@ def _build_hyper_pipeline(self, dataset):
             inputs=[node.get_hyper_preprocessors() for node in self.inputs],
             outputs=[head.get_hyper_preprocessors() for head in self._heads],
         )
+        self.tuner.hypermodel.hyper_pipeline = self.tuner.hyper_pipeline
 
     def _convert_to_dataset(self, x, y, validation_data, batch_size):
         """Convert the data to tf.data.Dataset."""

autokeras/engine/tuner.py

@@ -18,7 +18,6 @@
 
 import keras_tuner
 import tensorflow as tf
-from keras_tuner.engine import hypermodel as hm_module
 from tensorflow.keras import callbacks as tf_callbacks
 from tensorflow.keras.layers.experimental import preprocessing
 from tensorflow.python.util import nest
@@ -43,7 +42,7 @@ class AutoTuner(keras_tuner.engine.tuner.Tuner):
 
     # Arguments
         oracle: keras_tuner Oracle.
-        hypermodel: keras_tuner KerasHyperModel.
+        hypermodel: keras_tuner HyperModel.
         **kwargs: The args supported by KerasTuner.
     """
 
@@ -52,15 +51,15 @@ def __init__(self, oracle, hypermodel, **kwargs):
         self._finished = False
         super().__init__(oracle, hypermodel, **kwargs)
         # Save or load the HyperModel.
-        self.hypermodel.hypermodel.save(os.path.join(self.project_dir, "graph"))
+        self.hypermodel.save(os.path.join(self.project_dir, "graph"))
         self.hyper_pipeline = None
 
     def _populate_initial_space(self):
         # Override the function to prevent building the model during initialization.
         return
 
     def get_best_model(self):
-        with hm_module.maybe_distribute(self.distribution_strategy):
+        with keras_tuner.engine.tuner.maybe_distribute(self.distribution_strategy):
             model = tf.keras.models.load_model(self.best_model_path)
         return model
 
@@ -80,27 +79,27 @@ def _prepare_model_build(self, hp, **kwargs):
         pipeline = self.hyper_pipeline.build(hp, dataset)
         pipeline.fit(dataset)
         dataset = pipeline.transform(dataset)
-        self.hypermodel.hypermodel.set_io_shapes(data_utils.dataset_shape(dataset))
+        self.hypermodel.set_io_shapes(data_utils.dataset_shape(dataset))
 
         if "validation_data" in kwargs:
             validation_data = pipeline.transform(kwargs["validation_data"])
         else:
             validation_data = None
         return pipeline, dataset, validation_data
 
-    def _build_and_fit_model(self, trial, fit_args, fit_kwargs):
+    def _build_and_fit_model(self, trial, *args, **kwargs):
+        model = self.hypermodel.build(trial.hyperparameters)
         (
             pipeline,
-            fit_kwargs["x"],
-            fit_kwargs["validation_data"],
-        ) = self._prepare_model_build(trial.hyperparameters, **fit_kwargs)
+            kwargs["x"],
+            kwargs["validation_data"],
+        ) = self._prepare_model_build(trial.hyperparameters, **kwargs)
         pipeline.save(self._pipeline_path(trial.trial_id))
 
-        model = self.hypermodel.build(trial.hyperparameters)
-        self.adapt(model, fit_kwargs["x"])
+        self.adapt(model, kwargs["x"])
 
         _, history = utils.fit_with_adaptive_batch_size(
-            model, self.hypermodel.hypermodel.batch_size, **fit_kwargs
+            model, self.hypermodel.batch_size, **kwargs
         )
         return history
 
@@ -165,7 +164,7 @@ def search(
         if callbacks is None:
             callbacks = []
 
-        self.hypermodel.hypermodel.set_fit_args(validation_split, epochs=epochs)
+        self.hypermodel.set_fit_args(validation_split, epochs=epochs)
 
         # Insert early-stopping for adaptive number of epochs.
         epochs_provided = True
@@ -216,9 +215,7 @@ def search(
                 )
                 copied_fit_kwargs.pop("validation_data")
 
-            self.hypermodel.hypermodel.set_fit_args(
-                0, epochs=copied_fit_kwargs["epochs"]
-            )
+            self.hypermodel.set_fit_args(0, epochs=copied_fit_kwargs["epochs"])
             pipeline, model, history = self.final_fit(**copied_fit_kwargs)
         else:
             # TODO: Add return history functionality in Keras Tuner
@@ -270,7 +267,7 @@ def final_fit(self, **kwargs):
         model = self._build_best_model()
         self.adapt(model, kwargs["x"])
         model, history = utils.fit_with_adaptive_batch_size(
-            model, self.hypermodel.hypermodel.batch_size, **kwargs
+            model, self.hypermodel.batch_size, **kwargs
         )
         return pipeline, model, history
 

autokeras/tuners/greedy.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import copy
 from typing import Any
 from typing import Dict
 from typing import List
@@ -94,7 +95,7 @@ class GreedyOracle(keras_tuner.Oracle):
 
     def __init__(self, initial_hps=None, seed=None, **kwargs):
         super().__init__(seed=seed, **kwargs)
-        self.initial_hps = initial_hps or []
+        self.initial_hps = copy.deepcopy(initial_hps) or []
         self._tried_initial_hps = [False] * len(self.initial_hps)
 
     def get_state(self):

setup.py

@@ -18,7 +18,7 @@
     keywords=["AutoML", "Keras"],
     install_requires=[
         "packaging",
-        "keras-tuner>=1.0.2",
+        "keras-tuner==1.1.0rc0",
         "tf-nightly==2.8.0.dev20211016",
         "scikit-learn",
         "pandas",

tests/unit_tests/blocks/basic_test.py

@@ -229,7 +229,7 @@ def test_conv_get_config_has_all_attributes():
 
 
 def test_rnn_build_return_tensor():
-    block = blocks.RNNBlock()
+    block = blocks.RNNBlock(bidirectional=False)
 
     outputs = block.build(
         keras_tuner.HyperParameters(),

tests/unit_tests/tuners/task_specific_test.py

@@ -65,8 +65,8 @@ def test_txt_clf_init_hp2_equals_hp_of_a_model(tmp_path):
     clf.inputs[0].batch_size = 6
     clf.inputs[0].num_samples = 1000
     clf.outputs[0].in_blocks[0].shape = (10,)
-    clf.tuner.hypermodel.hypermodel.epochs = 1000
-    clf.tuner.hypermodel.hypermodel.num_samples = 20000
+    clf.tuner.hypermodel.epochs = 1000
+    clf.tuner.hypermodel.num_samples = 20000
     init_hp = task_specific.TEXT_CLASSIFIER[2]
     hp = keras_tuner.HyperParameters()
     hp.values = copy.copy(init_hp)