Basic Tensorflow Question (Input and Output Array)

If you want to feed multidimensional array you need to make some adjustments. For example: Your xs = [[1,3,4,5], [9,2,3,4]] has a shape of [2,4]. More precisely, it will have that shape if it'll be transformed to numpy array, like this:

np.array([np.array(x) for x in xs])

So, to train a model on data of shape [2,4], first you need to set input shape

model = tf.keras.Sequential([layers.Dense(units=1, input_shape=[2,4])])
model.add(layers.Dense(100))
model.add(layers.Dense(4))
model.compile(optimizer='sgd', loss='mean_squared_error')

The output layer has 4 units, same as last dimension. Then create some random data

xs = np.array([np.random.rand(2,4) for i in range(100)])
ys = np.array([xs[i]**2 for i in range(100)])

Here we create an array of arrays, each of shape [2,4] and same shaped array of labels, which are squares just for example.

model.fit(xs, ys, epochs=10, batch_size=1)

Then pass this data into the model, setting the batch size to 1, which means that a single array of shape [2,4] will be processed at a time.

And finally try to predict values with another random array. Note that here we need to add another dimension, which is batch size. It doesn't need to be 1, instead you can make any number of predictions, considering another dimension are [2,4]

p = np.random.rand(1,2,4)
print(model.predict(p))