Skip to content
Advertisement

How to use tf.data.Dataset.apply() for reshaping the dataset

I am working with time series models in tensorflow. My dataset contains physics signals. I need to divide this signals into windows as give this sliced windows as input to my model.

Here is how I am reading the data and slicing it:

import tensorflow as tf
import numpy as np
def _ds_slicer(data):
    win_len = 768
    return {"mix":(tf.stack(tf.split(data["mix"],win_len))),
            "pure":(tf.stack(tf.split(data["pure"],win_len)))}
dataset = tf.data.Dataset.from_tensor_slices({
    "mix" : np.random.uniform(0,1,[1000,24576]),
    "pure" : np.random.uniform(0,1,[1000,24576])
})
dataset = dataset.map(_ds_slicer)
print dataset.output_shapes
# {'mix': TensorShape([Dimension(768), Dimension(32)]), 'pure': TensorShape([Dimension(768), Dimension(32)])}

I want to reshape this dataset to # {'mix': TensorShape([Dimension(32)]), 'pure': TensorShape([Dimension(32))}

Equivalent transformation in numpy would be something like following:

signal  = np.random.uniform(0,1,[1000,24576])
sliced_sig = np.stack(np.split(signal,768,axis=1),axis=1)
print sliced_sig.shape #(1000, 768, 32)
sliced_sig=sliced_sig.reshape(-1, sliced_sig.shape[-1])
print sliced_sig.shape #(768000, 32)

I thought of using tf.contrib.data.group_by_window as an input to dataset.apply() but couldn’t figure out exactly how to use it. Is there a way I can use any custom transformation to reshape the dataset?

Advertisement

Answer

I think you’re just looking for the transformation tf.contrib.data.unbatch. This does exactly what you want:

x = np.zeros((1000, 768, 32))

dataset = tf.data.Dataset.from_tensor_slices(x)
print(dataset.output_shapes)  # (768, 32)
dataset = dataset.apply(tf.contrib.data.unbatch())
print(dataset.output_shapes)  # (32,)

From the documentation:

If elements of the dataset are shaped [B, a0, a1, …], where B may vary from element to element, then for each element in the dataset, the unbatched dataset will contain B consecutive elements of shape [a0, a1, …].

Edit for TF 2.0

(Thanks @DavidParks)

From TF 2.0, you can use directly tf.data.Dataset.unbatch:

x = np.zeros((1000, 768, 32))

dataset = tf.data.Dataset.from_tensor_slices(x)
print(dataset.output_shapes)  # (768, 32)
dataset = dataset.unbatch()
print(dataset.output_shapes)  # (32,)
Advertisement