How can combine two tensors so they are in one dataset? - tensorflow

Question

I'm working with the Titanic data set from the TensorFlow API.
I don't know how to make the features tensors model friendly.
Here's the best I got, but it's for one tensor at a time. How do I make it so it can handle all tensors in the features item?
import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.keras.optimizers import Adam
data = tfds.load("titanic",split='train', as_supervised=True).map(lambda x,y: (x,y)).prefetch(1)
for i in data.batch(1309):
xx1 = i[0]['age']
xx2 = i[0]['fare']
yyy = tf.convert_to_tensor(tf.one_hot(i[1],2))
model = tf.keras.models.Sequential([tf.keras.layers.Dense(1),
tf.keras.layers.Dense(13, activation='relu'),
tf.keras.layers.Dense(2, activation='softmax')])
model.compile(
optimizer=Adam(learning_rate=0.01),
loss='categorical_crossentropy',
metrics=['accuracy']
)
model.fit(xx1,yyy,epochs=30)
How do I concat the age and fare tensors so that they're in one data set?
I tried concat and stack to no avail.

Answer 1

This should be accomplishable with using tf.stack. As the input is already using the dataset API, I've refactored some code to leverage the dataset features for mapping the input format to the goal format that you have described. For convenience, here is a colab notebook with the example: https://colab.research.google.com/drive/1dHNe9rYaJSgqbj_QtQ1aJL_7WgKnLKsU?usp=sharing
# Nothing novel here
import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.keras.optimizers import Adam
data = tfds.load("titanic",split='train', as_supervised=True).map(lambda x,y: (x,y)).prefetch(1)
Basic demo of intended data restructuring
Take 1 item from the dataset and convert it to a tensor that includes both of the goal datapoints using tf.stack
for item in data.take(1):
age = item[0]['age']
fare = item[0]['fare']
output = tf.stack([age, fare], axis=0)
print(output)
Output: tf.Tensor([30. 13.], shape=(2,), dtype=float32)
Within the output we can see a single tensor with two values embedded as expected.
Usage as a TensorFlow Dataset
Tensorflow datasets can be provided directly for training, we can easily create a function which will map from the input data format to the goal format described in the problem. The below function will accomplish this, using the sample code from above.
# Input data and associated label
def transform_data(item, label):
# Extract values
age = item['age']
fare = item['fare']
# Create output tensor
output = tf.stack([age, fare], axis=0)
return output, label
# Create a training dataset from the base dataset - for each batch map the input format to the goal format by passing the mapping function
train_dataset = data.map(transform_data).batch(1200)
# Model - I made some minor changes to get it to run cleaner
model = tf.keras.models.Sequential([
tf.keras.layers.Dense(2),
tf.keras.layers.Dense(13, activation='relu'),
# As we have only two labels, this is really a binary problem, so I've created a single output neuron activated by sigmoid
tf.keras.layers.Dense(1,activation='sigmoid')
])
# Compiled with binary_crossentropy to complement the binary classification
model.compile(optimizer=Adam(learning_rate=0.01),loss='binary_crossentropy', metrics=['accuracy'])
model.fit(train_dataset,epochs=30)
Output:
Epoch 1/30
2/2 [==============================] - 0s 16ms/step - loss: 11.7881 - accuracy: 0.4385
Epoch 2/30
2/2 [==============================] - 0s 7ms/step - loss: 10.2350 - accuracy: 0.4270
...