[FIXED] Keras call model.fit where x is two-tuple of np.ndarray

Issue

I have a regression tf.keras.Model that takes in:

• x: tuple[np.ndarray, np.ndarray], where the two items have different shapes
  • Shapes are (128, 1152) and (1, 256)
• y: float

I have my model and training codified like so:

class MyModel(tf.keras.Model):

    def __init__(self):
        ...  # Omitted for brevity

    def call(self, inputs: tuple[tf.Tensor, tf.Tensor], training=None, mask=None):
        # Unpacks the two-tuple
        weights_1, weights_2 = inputs
        ...  # Omitted for brevity


# NOTE: item 0's shape is (128, 1152), item 1's shape is (1, 256)
datapoint_x: tuple[np.ndarray, np.ndarray]
datapoint_y: float

model = MyModel()
model(inputs=datapoint_x)  # Works fine

However, when I go to fit the model, I get an Exception:

>>> model.fit(x=datapoint_x, y=np.array(datapoint_y))

Traceback (most recent call last):
  File "/path/to/python3.10/site-packages/IPython/core/interactiveshell.py", line 3433, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-5-a5dfb3dd4846>", line 1, in <module>
    model.fit(x=datapoint_x, y=np.array(datapoint_y))
  File "/path/to/python3.10/site-packages/keras/utils/traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/path/to/python3.10/site-packages/tensorflow/python/framework/tensor_shape.py", line 910, in __getitem__
    return self._dims[key]
IndexError: tuple index out of range

I investigated this, and self._dims is () and key is 0.

What is the proper way to call Model.fit on a dataset with two-tuple x's?

Solution

The answer is Model.fit is iterating over the x's and y's, so I had to add a batch dimension up-front to my x[0] and y.

This could be easily done using np.newaxis or np.expand_dims.

import numpy as np

# NOTE: item 0's shape is (128, 1152), item 1's shape is (1, 256)
datapoint_x: tuple[np.ndarray, np.ndarray]
datapoint_y: float

# NOTE: now item 0's shape is (1, 128, 1152), and item 1's shape remains (1, 256)
batch_x = (datapoint_x[0][np.newaxis, :], datapoint_x[1])
# NOTE: now y's shape is (1,)
batch_y = np.array([datapoint_y])

model.fit(x=batch_x, y=batch_y)

Answered By - Intrastellar Explorer