CODEWITHSUNDEEP
pythontensorflow
Zabir Al Nazi Nabil
Zabir Al Nazi Nabil

Reputation: 11198

How to get the tensor shape dynamically for bucket_by_sequence_length with dynamic padding?

I am trying to train a text to speech model in TensorFlow. To generate batches, I'm using bucket_by_sequence_length function.

But, after I get each batch, I need to know the shape of each tensor. The problem is that the tensors come from different sized arrays.

I have tried get_shape().as_list() and 'tf.shape' but I don't get the shape in runtime. For the temporal dimension, I always get None. But, as I am planning to add a custom layer which needs to know the temporal dimension too, I am getting errors for dimension mismatch.

_, (texts, mels, mags, fnames) = tf.contrib.training.bucket_by_sequence_length(
                                            input_length=text_length,
                                            tensors=[text, mel, mag, fname],
                                            batch_size=hp.B,
                                            bucket_boundaries=[i for i in range(minlen + 1, maxlen - 1, 20)],
                                            num_threads=1,
                                            capacity=hp.B,
                                            dynamic_pad=True)

Each time I get batch size, for the tensors I get the following shape with get_shape().as_list() (10, None, 156) Here, mel, mag are two tensors with unknown temporal dimension (depends on the audio file length).

Here, 10 is the batch size, and 156 is the feature size. I also need the exact temporal dimension where I get None.

tf.shape returns Tensor("Text2Mel/AudioEnc/strided_slice:0", shape=(), dtype=int32) I don't get anything useful from this too.

How can I infer the exact shape at runtime?

Actual Source which I have modified: https://github.com/Kyubyong/dc_tts/blob/master/data_load.py

Upvotes: 0

Views: 399

Answers (1)

pratsbhatt
pratsbhatt

Reputation: 1538

I can see that you are using an older version of TensorFlow. The API has since moved to tf.data.experimental. If you use the tf.data.Dataset API then you can use the following code to get the size of the tensor.

def _element_length_fn(x, y=None):
    return array_ops.shape(x)[0]

Upvotes: 1

Related Questions