keras.layers.RandomGrayscale can't be put in as part of image augmentation layer

[AlundorZhu]

Hi all,
When I build the model with RandomGrayscale in my image augmentation layers it gives the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[35], line 8
      5 x = keras.layers.Dropout(0.2)(x)
      6 output = keras.layers.Dense(1, activation=None)(x)
----> 8 model = keras.Model(inputs=input, outputs=output)
     10 model.compile(
     11     optimizer='adam',
     12     loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
     13     metrics=['accuracy'],
     14 )

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/utils/tracking.py:26, in no_automatic_dependency_tracking.<locals>.wrapper(*args, **kwargs)
     23 @wraps(fn)
     24 def wrapper(*args, **kwargs):
     25     with DotNotTrackScope():
---> 26         return fn(*args, **kwargs)

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/models/functional.py:135, in Functional.__init__(self, inputs, outputs, name, **kwargs)
    132 if not all(is_input_keras_tensor(t) for t in flat_inputs):
    133     inputs, outputs = clone_graph_nodes(inputs, outputs)
--> 135 Function.__init__(self, inputs, outputs, name=name)
    137 if trainable is not None:
    138     self.trainable = trainable

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/ops/function.py:77, in Function.__init__(self, inputs, outputs, name)
     74 if backend() == "tensorflow":
     75     self._self_setattr_tracking = _self_setattr_tracking
---> 77 (nodes, nodes_by_depth, operations, operations_by_depth) = map_graph(
     78     self._inputs, self._outputs
     79 )
     80 self._nodes = nodes
     81 self._nodes_by_depth = nodes_by_depth

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/ops/function.py:232, in map_graph(inputs, outputs)
    216 """Validates a graph's topology and gather its operations and nodes.
    217 
    218 Args:
   (...)
    228         instances.
    229 """
    230 # "depth" is number of operations between output Node and the Node.
    231 # Nodes are ordered from inputs -> outputs.
--> 232 nodes_in_decreasing_depth, operation_indices = _build_map(inputs, outputs)
    233 network_nodes = {
    234     make_node_key(node.operation, node.operation._inbound_nodes.index(node))
    235     for node in nodes_in_decreasing_depth
    236 }
    238 nodes_depths = {}  # dict {node: depth value}

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/ops/function.py:363, in _build_map(inputs, outputs)
    361 operation_indices = {}  # operation -> in traversal order.
    362 for output in tree.flatten(outputs):
--> 363     _build_map_helper(
    364         inputs,
    365         output,
    366         finished_nodes,
    367         nodes_in_progress,
    368         nodes_in_decreasing_depth,
    369         operation_indices,
    370     )
    371 return nodes_in_decreasing_depth, operation_indices

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/ops/function.py:412, in _build_map_helper(inputs, tensor, finished_nodes, nodes_in_progress, nodes_in_decreasing_depth, operation_indices)
    410 if not node.is_input and tensor not in tree.flatten(inputs):
    411     for tensor in node.input_tensors:
--> 412         _build_map_helper(
    413             inputs,
    414             tensor,
    415             finished_nodes,
    416             nodes_in_progress,
    417             nodes_in_decreasing_depth,
    418             operation_indices,
    419         )
    421 finished_nodes.add(node)
    422 nodes_in_progress.remove(node)

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/ops/function.py:412, in _build_map_helper(inputs, tensor, finished_nodes, nodes_in_progress, nodes_in_decreasing_depth, operation_indices)
    410 if not node.is_input and tensor not in tree.flatten(inputs):
    411     for tensor in node.input_tensors:
--> 412         _build_map_helper(
    413             inputs,
    414             tensor,
    415             finished_nodes,
    416             nodes_in_progress,
    417             nodes_in_decreasing_depth,
    418             operation_indices,
    419         )
    421 finished_nodes.add(node)
    422 nodes_in_progress.remove(node)

    [... skipping similar frames: _build_map_helper at line 412 (1 times)]

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/ops/function.py:412, in _build_map_helper(inputs, tensor, finished_nodes, nodes_in_progress, nodes_in_decreasing_depth, operation_indices)
    410 if not node.is_input and tensor not in tree.flatten(inputs):
    411     for tensor in node.input_tensors:
--> 412         _build_map_helper(
    413             inputs,
    414             tensor,
    415             finished_nodes,
    416             nodes_in_progress,
    417             nodes_in_decreasing_depth,
    418             operation_indices,
    419         )
    421 finished_nodes.add(node)
    422 nodes_in_progress.remove(node)

File ~/Work/Research/.venv-metal/lib/python3.12/site-packages/keras/src/ops/function.py:399, in _build_map_helper(inputs, tensor, finished_nodes, nodes_in_progress, nodes_in_decreasing_depth, operation_indices)
    397 # Prevent cycles.
    398 if node in nodes_in_progress:
--> 399     raise ValueError(
    400         f"Tensor {tensor} from operation '{operation.name}' is part of a "
    401         "cycle."
    402     )
    404 # Store the traversal order for operation sorting.
    405 if operation not in operation_indices:

ValueError: Tensor <KerasTensor shape=(None, 224, 224, 3), dtype=float32, sparse=False, name=keras_tensor_1089> from operation 'random_grayscale_4' is part of a cycle.

The code I use to build the model:

data_augmentation_layers = [
    layers.RandomFlip("horizontal"),
    layers.RandomRotation(0.02),
    layers.RandomShear(x_factor=0.1, y_factor=0.1),
    layers.RandomTranslation(0.1, 0.1),
    layers.RandomGrayscale(0.2)
]


def data_augmentation(images):
    for layer in data_augmentation_layers:
        images = layer(images)
    return images


IMG_SHAPE = (224, 224, 3)

base_model = keras.applications.MobileNetV3Large(
    include_top=False,
    weights="imagenet",
    input_shape=IMG_SHAPE,
    pooling="avg",
    classifier_activation=None
)

preprocess_input = keras.applications.mobilenet_v3.preprocess_input

input = keras.Input(IMG_SHAPE)
x = data_augmentation(input)
x = preprocess_input(x)
x = base_model(x)
x = keras.layers.Dropout(0.2)(x)
output = keras.layers.Dense(1, activation=None)(x)

model = keras.Model(inputs=input, outputs=output)

model.compile(
    optimizer='adam',
    loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=['accuracy'],
)

Best,

[pctablet505]

The below code works for me. After fixing some bug for passing single unbatched image in RandomGrayscale.call
#21277

class DataAugmentation(keras.Layer):
    def __init__(self, name=None):
        super().__init__(name='Data Augmentation')
        # Store layers as attributes to ensure they are tracked by Keras
        self.random_flip = layers.RandomFlip("horizontal")
        self.random_rotation = layers.RandomRotation(0.02)
        self.random_shear = layers.RandomShear(x_factor=0.1, y_factor=0.1)
        self.random_translation = layers.RandomTranslation(0.1, 0.1)
        self.random_grayscale = layers.RandomGrayscale(0.2)

    def call(self, images, training=None):
        x = images
        # Pass the training argument to each layer
        x = self.random_flip(x, training=training)
        x = self.random_rotation(x, training=training)
        x = self.random_shear(x, training=training)
        x = self.random_translation(x, training=training)
        x = self.random_grayscale(x, training=training)
        return x



IMG_SHAPE = (224, 224, 3)

base_model = keras.applications.MobileNetV3Large(
    include_top=False,
    weights=None,
    input_shape=IMG_SHAPE,
    pooling="avg",
    classifier_activation=None
)

preprocess_input = keras.applications.mobilenet_v3.preprocess_input

input = keras.Input(IMG_SHAPE)
x = DataAugmentation()(input)
x = preprocess_input(x)
x = base_model(x)
x = keras.layers.Dropout(0.2)(x)
output = keras.layers.Dense(1, activation=None)(x)

model = keras.Model(inputs=input, outputs=output)

model.compile(
    optimizer='adam',
    loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=['accuracy'],
)

print(model)
print(model.summary())

[AlundorZhu]

Thank you! it works!

Also, final bit of change to make model.fit work:

def __init__(self, name=None):
        # Change the layer name to remove the space
        super().__init__(name='DataAugmentation')

[pctablet505]

import os
os.environ["KERAS_BACKEND"] = "torch"
import keras
from keras import layers


data_augmentation_layers = [
    layers.RandomFlip("horizontal"),
    layers.RandomRotation(0.02),
    layers.RandomShear(x_factor=0.1, y_factor=0.1),
    layers.RandomTranslation(0.1, 0.1),
    layers.RandomGrayscale(0.2)
]


def data_augmentation(images):
    for layer in data_augmentation_layers:
        images = layer(images)
    return images


IMG_SHAPE = (224, 224, 3)

input = keras.Input(IMG_SHAPE)
x = data_augmentation(input)

output = x

model = keras.Model(inputs=input, outputs=output)

model.compile(
    optimizer='adam',
    loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=['accuracy'],
)

print(model)
print(model.summary())

this is giving error

/usr/local/bin/python3 /Users/hellorahul/PyCharmMiscProject/script.py 
Traceback (most recent call last):
  File "/Users/hellorahul/PyCharmMiscProject/script.py", line 54, in <module>
    model = keras.Model(inputs=input, outputs=output)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/hellorahul/Projects/keras/keras/src/utils/tracking.py", line 26, in wrapper
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/Users/hellorahul/Projects/keras/keras/src/models/functional.py", line 136, in __init__
    Function.__init__(self, inputs, outputs, name=name)
  File "/Users/hellorahul/Projects/keras/keras/src/ops/function.py", line 77, in __init__
    (nodes, nodes_by_depth, operations, operations_by_depth) = map_graph(
                                                               ^^^^^^^^^^
  File "/Users/hellorahul/Projects/keras/keras/src/ops/function.py", line 238, in map_graph
    nodes_in_decreasing_depth, operation_indices = _build_map(inputs, outputs)
                                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/hellorahul/Projects/keras/keras/src/ops/function.py", line 369, in _build_map
    _build_map_helper(
  File "/Users/hellorahul/Projects/keras/keras/src/ops/function.py", line 418, in _build_map_helper
    _build_map_helper(
  File "/Users/hellorahul/Projects/keras/keras/src/ops/function.py", line 405, in _build_map_helper
    raise ValueError(
ValueError: Tensor <KerasTensor shape=(None, 224, 224, 3), dtype=float32, sparse=False, ragged=False, name=keras_tensor_4> from operation 'random_grayscale' is part of a cycle.

Process finished with exit code 1

whereas if we use the Subclassing based technique, it works

import os
# os.environ["KERAS_BACKEND"] = "torch"
import keras
from keras import layers



#%%

class DataAugmentation(keras.Layer):
    def __init__(self, name=None):
        super().__init__(name='Data Augmentation')
        # Store layers as attributes to ensure they are tracked by Keras
        self.random_flip = layers.RandomFlip("horizontal")
        self.random_rotation = layers.RandomRotation(0.02)
        self.random_shear = layers.RandomShear(x_factor=0.1, y_factor=0.1)
        self.random_translation = layers.RandomTranslation(0.1, 0.1)
        self.random_grayscale = layers.RandomGrayscale(0.2)

    def call(self, images, training=None):
        x = images
        # Pass the training argument to each layer
        x = self.random_flip(x, training=training)
        x = self.random_rotation(x, training=training)
        x = self.random_shear(x, training=training)
        x = self.random_translation(x, training=training)
        x = self.random_grayscale(x, training=training)
        return x



# data_augmentation_layers = [
#     layers.RandomFlip("horizontal"),
#     layers.RandomRotation(0.02),
#     layers.RandomShear(x_factor=0.1, y_factor=0.1),
#     layers.RandomTranslation(0.1, 0.1),
#     layers.RandomGrayscale(0.2)
# ]
#
#
# def data_augmentation(images):
#     for layer in data_augmentation_layers:
#         images = layer(images)
#     return images


IMG_SHAPE = (224, 224, 3)

input = keras.Input(IMG_SHAPE)
x = DataAugmentation()(input)

output = x

model = keras.Model(inputs=input, outputs=output)

model.compile(
    optimizer='adam',
    loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=['accuracy'],
)

print(model)
print(model.summary())

Output:

/usr/local/bin/python3 /Users/hellorahul/PyCharmMiscProject/script.py 
<Functional name=functional, built=True>
Model: "functional"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ input_layer (InputLayer)        │ (None, 224, 224, 3)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ Data Augmentation               │ (None, 224, 224, 3)    │             0 │
│ (DataAugmentation)              │                        │               │
└─────────────────────────────────┴────────────────────────┴───────────────┘
 Total params: 0 (0.00 B)
 Trainable params: 0 (0.00 B)
 Non-trainable params: 0 (0.00 B)

[pctablet505]

Why is it that other keras image preprocessing layers not causing cycle, whereas random_grayscale causing cycle?
I tried multiple things like mimicking the codeflow from random_flip, random_contrast, etc in random_grayscale, but it still detects a cycle.