RaggedTensor that represents variable length data introduced in TensorFlow 2.1 or later, but if you try to write with ordinary Tensor glue, there are various things I'm addicted to it. This time, we will use window processing for signal processing. A frame with a certain time width is shifted little by little to extract a waveform that falls within the frame range.

- Ubuntu 18.04
- Python 3.6.9
- TensorFlow 2.2.0 (CPU)

Think of ` x`

as a batch and cut out a short interval waveform for each row of data. The length of the data varies.
Here, the frame width is set to 2, and the cutout position is shifted by 1.
`[3, 1, 4, 1]`

is like `[[3, 1], [1, 4], [4, 1]]`

.

For the usual ` Tensor`

, there is a handy function called tf.signal.frame, but unfortunately `It cannot be used for`

RaggedTensor``.

```
x = tf.ragged.constant([[3, 1, 4, 1], [], [5, 9, 2], [6], []])
print(tf.signal.frame(x, 2, 1)) # NG
# ValueError: TypeError: object of type 'RaggedTensor' has no len()
print(tf.signal.frame(x.to_tensor(), 2, 1)) #It works, but a lot of extra 0s come out
# tf.Tensor(
# [[[3 1]
# [1 4]
# [4 1]]
#
# [[0 0]
# [0 0]
# [0 0]]
#
# [[5 9]
# [9 2]
# [2 0]]
#
# [[6 0]
# [0 0]
# [0 0]]
#
# [[0 0]
# [0 0]
# [0 0]]], shape=(5, 3, 2), dtype=int32)
```

Think based on ` x.values`

, which represents the values that are flattened, and the length and offset of each row from ` RaggedTensor`

.

```
print(x.values) #Tensor with values
# tf.Tensor([3 1 4 1 5 9 2 6], shape=(8,), dtype=int32)
print(x.row_starts()) #Start index (offset) of each row in values
# tf.Tensor([0 4 4 7 8], shape=(5,), dtype=int64)
print(x.row_lengths()) #Length of each line
# tf.Tensor([4 0 3 1 0], shape=(5,), dtype=int64)
```

For each row of ` x`

, consider from which index of ` x.values`

the values should be taken (*).

--Line 0 is `[0, 1], [1, 2], [2, 3]`

--The first line is empty
--The second line is `[4, 5], [5, 6]`

--The third line is empty
--The fourth line is empty

First of all, if you make a ` RaggedTensor`

that has the first index of (*)

```
s = x.row_starts()
e = s + x.row_lengths() - 1
r = tf.ragged.range(s, e)
print(r)
# <tf.RaggedTensor [[0, 1, 2], [], [4, 5], [], []]>
```

In addition, you can combine the one-advanced indexes to see where in ` x.values`

you should get the values for the expected result after windowing. The results correspond to the previous bullet points (*).

```
ind = tf.stack([r, r+1], axis=2)
print(ind)
# <tf.RaggedTensor [[[0, 1], [1, 2], [2, 3]], [], [[4, 5], [5, 6]], [], []]>
```

After that, you can use `tf.gather ()`

to get the values from ` x.values`

based on the index entered in ```ind``.

```
ret = tf.gather(x.values, ind)
print(ret)
# <tf.RaggedTensor [[[3, 1], [1, 4], [4, 1]], [], [[5, 9], [9, 2]], [], []]>
```

The way to make `ʻe`` and`

ʻind` is slightly different, but the general idea is the same. Broadcast is used to create `ʻind`

. For that, we add a dimension of length 1 at the end as `r [:,:, tf.newaxis]`

.

```
len_frame = 3
s = x.row_starts()
e = s + x.row_lengths() + 1 - len_frame
r = tf.ragged.range(s, e)
ind = r[:, :, tf.newaxis] + tf.range(0, len_frame, dtype="int64")
ret = tf.gather(x.values, ind)
print(ret)
# <tf.RaggedTensor [[[3, 1, 4], [1, 4, 1]], [], [[5, 9, 2]], [], []]>
```

Of course, it can be used even when ` len_frame = 2`

.

It is OK if you change the step size of ` r`

.

```
len_frame = 2
shift_frame = 2
s = x.row_starts()
e = s + x.row_lengths() + 1 - len_frame
r = tf.ragged.range(s, e, shift_frame)
ind = r[:, :, tf.newaxis] + tf.range(0, len_frame, dtype="int64")
ret = tf.gather(x.values, ind)
print(ret)
# <tf.RaggedTensor [[[3, 1], [4, 1]], [], [[5, 9]], [], []]>
```

You can also use ` shift_frame = 1`

.

For example, in stereo audio, the L and R values are stored in pairs.

```
x = tf.ragged.constant([[[3, 2], [1, 7], [4, 1], [1, 8]], [], [[5, 2], [9, 8], [2, 1]], [[6, 8]], []])
```

In fact, it works in exactly the same way as before.

```
len_frame = 2
shift_frame = 1
s = x.row_starts()
e = s + x.row_lengths() + 1 - len_frame
r = tf.ragged.range(s, e, shift_frame)
ind = r[:, :, tf.newaxis] + tf.range(0, len_frame, dtype="int64")
ret = tf.gather(x.values, ind)
print(ret)
# <tf.RaggedTensor [[[[3, 2], [1, 7]], [[1, 7], [4, 1]], [[4, 1], [1, 8]]], [], [[[5, 2], [9, 8]], [[9, 8], [2, 1]]], [], []]>
```

The number of dimensions has increased so much that I can't tell if it fits just by looking at it, but it should be okay ...

Recommended Posts