Error on converting ragged tensors to dense

Keras
,
1

I’m training SegNet architecture model and my dataset contains images of different resolutions. For this reason, I need to use ragged tensors when training. To do this, I wrote the simplest layer:

class RaggedToDenseTensor(tf.keras.Layer):

def __init__(self, **kwargs):
    super(RaggedToDenseTensor, self).__init__(**kwargs)

def call(self, inputs):
    if type(inputs) is tf.RaggedTensor:
        inputs = inputs.to_tensor()
    return inputs

This layer accepts a tensor from Input:

x = layers.Input(shape=(None, None, 3)) # 3-channel RGB image
x = RaggedToDenseTensor()(x)
…

There are no error on eager mode. But if I do not use this mode, then this error occurs:

…
target = tf.convert_to_tensor(target)
TypeError: Failed to convert elements of tf.RaggedTensor(values=tf.RaggedTensor(values=Tensor(“data_4:0”, shape=(None,), dtype=float32), row_splits=Tensor(“data_6:0”, shape=(None,), dtype=int64)), row_splits=Tensor(“data_5:0”, shape=(None,), dtype=int64)) to Tensor. Consider casting elements to a supported type.

Maybe I don’t even need to use ragged tensors because I use mini-batch training mode and when the mini-batch ragged tensor to dense tensor conversion is done, all the training examples are padded with zeros to have the same shape. Maybe do it before calling Model.fit method and save time this way?

Or in such cases do I need to write my own training method to be able to use batch mode?

3

Hi @Dmitry_Redgrave,

Sorry for the delay in response.
Yes, I would suggest to convert ragged tensors to dense tensors before model.fit if you are using mini-batch training and fixed-size images within each batch in graph mode.In graph mode, generally TF creates a static graph and expects a static type of data during execution, this creates discrepancy in your case.So I recommend to create a preprocessing layers(conversion,padding…etc) before training it especially in graph mode while using ragged tensors.

Hope this helps.Thank You.