How to merge two tf.data.Dataset into one, alternating elements with known ratio

user_4105823 0
Viewed 2656

I have two tf.data.Dataset, let call them d1 and d2 and I want to construct another dataset that constains the elements of d1 and d2 alternating. It is easier to explain with an example. Let say:

d1 = [0,1,2,3,4,5,6,7,...] # it is not a list, just the content of the dataset

d2 = ["a", "b", "c", "d",... ]

and I have the couple specifying the number of consecutive elements from each dataset (for example (3,1)).

The result that I am looking for is:

result = [0, 1, 2, "a", 3, 4, 5, "b", 6, 7, 8, "c"...]

EDIT: d1 and d2 are objects of the class tf.data.Dataset. The example above shows just the content of the datasets but it is not code.

tensorflow dataset python
2 Answers

Assuming TF 2.0. The trick is based on batch followed by datasets interleave and unbatch.

import tensorflow as tf 

# input datasets
d1 = tf.data.Dataset.from_tensors([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]).unbatch()
d2 = tf.data.Dataset.from_tensors([100, 101, 102]).unbatch()
# replaced letters with numbers to make tensor types match

# define ratio
r1 = 3
r2 = 1

b1 = d1.batch(r1)
b2 = d2.batch(r2)

zipped = tf.data.Dataset.zip((b1, b2)).map(lambda x, y: tf.concat((x, y), axis=0))
result = zipped.unbatch()

Output:

In [9]: list(result)                                                                                                                  
Out[9]: 
[<tf.Tensor: id=224, shape=(), dtype=int32, numpy=0>,
 <tf.Tensor: id=225, shape=(), dtype=int32, numpy=1>,
 <tf.Tensor: id=226, shape=(), dtype=int32, numpy=2>,
 <tf.Tensor: id=227, shape=(), dtype=int32, numpy=100>,
 <tf.Tensor: id=228, shape=(), dtype=int32, numpy=3>,
 <tf.Tensor: id=229, shape=(), dtype=int32, numpy=4>,
 <tf.Tensor: id=230, shape=(), dtype=int32, numpy=5>,
 <tf.Tensor: id=231, shape=(), dtype=int32, numpy=101>,
 <tf.Tensor: id=232, shape=(), dtype=int32, numpy=6>,
 <tf.Tensor: id=233, shape=(), dtype=int32, numpy=7>,
 <tf.Tensor: id=234, shape=(), dtype=int32, numpy=8>,
 <tf.Tensor: id=235, shape=(), dtype=int32, numpy=102>]

Note: This solution might remove some elements at the end of d1 or d2 - their lengths must be adjusted to the ratio.

user_100750900 
print(d1)
print("---------------------------")
print(d2)
print("---------------------------")

def interweave(x, d1, y, d2):
    """
    x  = How many lines of d1 to add before adding a line from d2
    d1 = the d1 dataframe
    y  = How many lines of d2 to add before adding a line from d1 again
    d2 = the d2 dataframe
    """
    d3 = pd.DataFrame()
    countx = 0
    county = 0
    length = len(d1) if len(d1) > len(d2) else len(d2) 

    for count in range(0,length):
        for i in range(countx, countx + x):
            try: # This will prevent script halt from unequal or indivisible lengths 
                row = d1.iloc[(i)]
            except: 
                break
            d3 = d3.append(row)
            countx += 1

        for j in range(county, county + y):
            try: # This will prevent script halt from unequal or indivisible lengths 
                row = d2.iloc[j]
            except:
                break
            d3 = d3.append(row)
            county += 1

    d3 = d3.reset_index(drop = True)
    return d3

d3 = interweave(3, d1, 1, d2)
print(d3)

OUTPUT:

       Col1  Col2
0      0     0
1      1    10
2      2    20
3      3    30
4      4    40
5      5    50
6      6    60
7      7    70
8      8    80
9      9    90
10    10   100
---------------------------
     Col1 Col2
0    a    A
1    b    B
2    c    C
---------------------------
      Col1 Col2
0     0    0
1     1   10
2     2   20
3     a    A
4     3   30
5     4   40
6     5   50
7     b    B
8     6   60
9     7   70
10    8   80
11    c    C
12    9   90
13   10  100
user_14407640
Related