ripe swallow
Hello, I have two models, one outputs a ragged tensor:
[[
[1, 2, 1],
],
[
[2, 0, 0],
[1, 1, 0]
]]
while the second model outputs a tensor:
[
[5, 4],
[6, 7]
]
I need to "expand" the second models ouput to concatenate it with the first model. Basically I want to add a dimension and copy the result as many times as there are in the ragged tensor, like:
[[
[5, 4],
],
[
[6, 7],
[6, 7]
]]
Does anyone know how to do this? This example would have a batch size of 2.
steep heart
how should the output look like?
ripe swallow
Like the last code block. But I think I got something that work for my model architecture with
class RaggedExpansionLayer(keras.layers.Layer):
def __init__(self, **kwargs):
super(RaggedExpansionLayer, self).__init__(**kwargs)
def call(self, inputs):
ragged_tensor, dense_tensor = inputs
row_lengths = ragged_tensor.row_lengths()
batch_indices = tf.range(tf.shape(dense_tensor)[0])
repeated_indices = tf.repeat(batch_indices, row_lengths)
gathered = tf.gather(dense_tensor, repeated_indices)
ragged = tf.RaggedTensor.from_row_lengths(gathered, row_lengths)
return tf.concat([ragged_tensor, ragged], axis=-1)
It doesn't solve my issue:
AttributeError: 'SymbolicTensor' object has no attribute 'row_lengths'
Call arguments received by layer "ragged_expansion_layer" (type RaggedExpansionLayer):
• inputs=['tf.Tensor(shape=(None, None, 100), dtype=float32)', 'tf.Tensor(shape=(None, 4000), dtype=float32)']
I mean how does the concatenated output of your example tensors look like?
the code constructs a graph rather than eagerly evaluate things, that's why it's a symbolic tensor and length can't find computed
ripe swallow
The output should look like:
[[
[1, 2, 1, 5, 4],
],
[
[2, 0, 0, 6, 7],
[1, 1, 0, 6, 7]
]]
I am currently working on it. I will come back to you if cant figure it out at all and then provide a minimal working example. Dont want to waste your time 🙂