[Python] How do I format my a tensorflow dataset for a multi output model?

I have an image dataset where each image has multiple categorical features that I want to predict. I am getting this error when trying to train

ValueError: y_true and y_pred have different structures. y_true: * y_pred: ['', '', '', '']
​

Heres is my code

def getLabel(path):
path = path.numpy().decode("utf-8")
key = os.path.basename(path)[:9]

if key not in allLabelDict:
print("Missing key:", key)
raise ValueError("Missing label key.")
return tf.convert_to_tensor(allLabelDict[key], dtype=tf.uint8)

def getImage(path):
img = tf.io.read_file(path)
img = tf.io.decode_jpeg(img, channels=3)
return tf.image.resize_with_crop_or_pad(img, 360, 360)

def processData(file_path):
label = tf.py_function(func=getLabel, inp=[file_path], Tout=tf.uint8)
label.set_shape([4])

img = tf.py_function(func=getImage, inp=[file_path], Tout=tf.uint8)
img.set_shape([360, 360, 3])

return img, label


allLabelDict is a dict with a key of each images file name and a value with an array of the sparse one hot encoded categories.


valSize = int(imageFileNames.cardinality().numpy() * 0.2)

trainData = imageFileNames \
.skip(valSize) \
.map(process_path, num_parallel_calls=tf.data.AUTOTUNE) \
.cache() \
.batch(100) \
.prefetch(buffer_size=tf.data.AUTOTUNE)
valData = imageFileNames \
.take(valSize) \
.map(process_path, num_parallel_calls=tf.data.AUTOTUNE) \
.cache() \
.batch(100) \
.prefetch(buffer_size=tf.data.AUTOTUNE)

inputLayer = tf.keras.layers.Input(shape=(360, 360, 3))

baseLayers = tf.keras.layers.Rescaling(1./255)(inputLayer)
baseLayers = tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same')(baseLayers)
baseLayers = tf.keras.layers.MaxPooling2D((2, 2))(baseLayers)
baseLayers = tf.keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same')(baseLayers)
baseLayers = tf.keras.layers.MaxPooling2D((2, 2))(baseLayers)
baseLayers = tf.keras.layers.Conv2D(128, (3, 3), activation='relu', padding='same')(baseLayers)
baseLayers = tf.keras.layers.Flatten()(baseLayers)
baseLayers = tf.keras.layers.Dense(128, activation='relu')(baseLayers)

labelOutput = tf.keras.layers.Dense(len(labelTuple))(baseLayers)
cellShapeOutput = tf.keras.layers.Dense(len(cellShapeTuple))(baseLayers)
nucleusShapeOutput = tf.keras.layers.Dense(len(nucleusShapeTuple))(baseLayers)
cytoplasmVacuoleOutput = tf.keras.layers.Dense(len(cytoplasmVacuoleTuple))(baseLayers)

model = tf.keras.Model(inputs=inputLayer, outputs=[labelOutput, cellShapeOutput, nucleusShapeOutput, cytoplasmVacuoleOutput])

model.compile(
optimizer=tf.keras.optimizers.Adam(),
loss={
"labelOutput": tf.keras.losses.SparseCategoricalCrossentropy(),
"cellShapeOutput": tf.keras.losses.SparseCategoricalCrossentropy(),
"nucleusShapeOutput": tf.keras.losses.SparseCategoricalCrossentropy(),
"cytoplasmVacuoleOutput": tf.keras.losses.SparseCategoricalCrossentropy()
},
metrics={
"labelOutput": ["sparse_categorical_accuracy"],
"cellShapeOutput": ["sparse_categorical_accuracy"],
"nucleusShapeOutput": ["sparse_categorical_accuracy"],
"cytoplasmVacuoleOutput": ["sparse_categorical_accuracy"]
}
)

model.summary()

Continue reading...