Adapt to tf-nightly (#1634)

* fixed tests in keras_layers_test

* fixed image classifier test

* fix bert tokenizer

* multi branch arch adapt preprocessing layers

* depending on tf-nightly

* coverage

Co-authored-by: Haifeng Jin <[email protected]>

.github/workflows/actions.yml

@@ -11,10 +11,10 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v2
-    - name: Set up Python 3.6
+    - name: Set up Python 3.7
       uses: actions/setup-python@v1
       with:
-        python-version: 3.6
+        python-version: 3.7
     - name: Get pip cache dir
       id: pip-cache
       run: |
@@ -50,10 +50,10 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v2
-      - name: Set up Python 3.6
+      - name: Set up Python 3.7
         uses: actions/setup-python@v1
         with:
-          python-version: 3.6
+          python-version: 3.7
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip setuptools
@@ -73,7 +73,7 @@ jobs:
     - name: Set up Python
       uses: actions/setup-python@v1
       with:
-        python-version: 3.6
+        python-version: 3.7
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip setuptools wheel twine

autokeras/engine/tuner.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import collections
 import copy
 import os
 
@@ -112,26 +113,32 @@ def adapt(model, dataset):
         # TODO: Use Keras Tuner for preprocessing layers adapt.
         x = dataset.map(lambda x, y: x)
 
-        def get_output_layer(tensor):
+        def get_output_layers(tensor):
+            output_layers = []
             tensor = nest.flatten(tensor)[0]
             for layer in model.layers:
                 if isinstance(layer, tf.keras.layers.InputLayer):
                     continue
                 input_node = nest.flatten(layer.input)[0]
                 if input_node is tensor:
-                    if not isinstance(layer, preprocessing.PreprocessingLayer):
-                        break
-                    return layer
-            return None
+                    if isinstance(layer, preprocessing.PreprocessingLayer):
+                        output_layers.append(layer)
+            return output_layers
+
+        dq = collections.deque()
 
         for index, input_node in enumerate(nest.flatten(model.input)):
-            temp_x = x.map(lambda *args: nest.flatten(args)[index])
-            layer = get_output_layer(input_node)
-            while layer is not None:
-                if isinstance(layer, preprocessing.PreprocessingLayer):
-                    layer.adapt(temp_x)
-                temp_x = temp_x.map(layer)
-                layer = get_output_layer(layer.output)
+            in_x = x.map(lambda *args: nest.flatten(args)[index])
+            for layer in get_output_layers(input_node):
+                dq.append((layer, in_x))
+
+        while len(dq):
+            layer, in_x = dq.popleft()
+            layer.adapt(in_x)
+            out_x = in_x.map(layer)
+            for next_layer in get_output_layers(layer.output):
+                dq.append((next_layer, out_x))
+
         return model
 
     def search(

autokeras/keras_layers.py

@@ -40,6 +40,9 @@ def get_config(self):
     def call(self, inputs):
         return data_utils.cast_to_float32(inputs)
 
+    def adapt(self, data):
+        return
+
 
 @tf.keras.utils.register_keras_serializable()
 class ExpandLastDim(preprocessing.PreprocessingLayer):
@@ -49,6 +52,9 @@ def get_config(self):
     def call(self, inputs):
         return tf.expand_dims(inputs, axis=-1)
 
+    def adapt(self, data):
+        return
+
 
 @tf.keras.utils.register_keras_serializable()
 class MultiCategoryEncoding(preprocessing.PreprocessingLayer):
@@ -75,9 +81,7 @@ def __init__(self, encoding: List[str], **kwargs):
                 # Set a temporary vocabulary to prevent the error of no
                 # vocabulary when calling the layer to build the model.  The
                 # vocabulary would be reset by adapting the layer later.
-                self.encoding_layers.append(
-                    preprocessing.StringLookup(vocabulary=["NONE"])
-                )
+                self.encoding_layers.append(preprocessing.StringLookup())
             elif encoding == ONE_HOT:
                 self.encoding_layers.append(None)
 
@@ -190,6 +194,9 @@ def bert_encode(self, input_tensor):
 
         return input_word_ids
 
+    def adapt(self, data):
+        return  # pragma: no cover
+
 
 # TODO: Remove after KerasNLP is ready.
 @tf.keras.utils.register_keras_serializable()
@@ -685,6 +692,9 @@ def call(self, inputs):
 
         return mask  # pragma: no cover
 
+    def get_config(self):
+        return super().get_config()
+
 
 @tf.keras.utils.register_keras_serializable()
 class Transformer(tf.keras.layers.Layer):

setup.cfg

@@ -14,7 +14,7 @@ filterwarnings =
     ignore::PendingDeprecationWarning
     ignore::FutureWarning
     ignore::numpy.VisibleDeprecationWarning
-    ignore::tensorflow.python.keras.utils.generic_utils.CustomMaskWarning
+    ignore::keras.utils.generic_utils.CustomMaskWarning
 
 addopts=-v
         --durations=10

setup.py

@@ -19,7 +19,7 @@
     install_requires=[
         "packaging",
         "keras-tuner>=1.0.2",
-        "tensorflow<=2.5.0,>=2.3.0",
+        "tf-nightly==2.8.0.dev20211016",
         "scikit-learn",
         "pandas",
     ],
@@ -41,7 +41,6 @@
         "Intended Audience :: Education",
         "Intended Audience :: Science/Research",
         "License :: OSI Approved :: Apache Software License",
-        "Programming Language :: Python :: 3.6",
         "Programming Language :: Python :: 3.7",
         "Programming Language :: Python :: 3.8",
         "Topic :: Scientific/Engineering :: Mathematics",

tests/unit_tests/keras_layers_test.py

@@ -62,9 +62,11 @@ def test_init_multi_one_hot_encode():
 
 
 def test_call_multi_with_single_column_return_right_shape():
+    x_train = np.array([["a"], ["b"], ["a"]])
     layer = layer_module.MultiCategoryEncoding(encoding=[layer_module.INT])
+    layer.adapt(tf.data.Dataset.from_tensor_slices(x_train).batch(32))
 
-    assert layer(np.array([["a"], ["b"], ["a"]])).shape == (3, 1)
+    assert layer(x_train).shape == (3, 1)
 
 
 def get_text_data():