How to pick selected rows from a tensor, regardless of dimension? - python

I have a tensor (tensorflow.Tensor) A, and I would like to form a new tensor containing certain rows from A, that is, A[i,:,:,...,:] for selected values of i.
Problem is I don't know before-hand how many axes A has. So how can I write this operation?

This is exactly what tf.gather() is for. See the example code below:
x = tf.reshape(tf.constant([1, 2, 3, 4, 5, 6, 7, 8]), [2, 2, 2])
# This is using tf.gather() on a 3D tensor.
print(tf.gather(x, [1]))
The result is:
<tf.Tensor: shape=(1, 2, 2), dtype=int32, numpy=
array([[[5, 6],
[7, 8]]], dtype=int32)>
x = tf.reshape(tf.constant([1, 2, 3, 4, 5, 6, 7, 8]), [2, 4])
# This is using tf.gather() on a 2D tensor.
print(tf.gather(x, [1]))
The result is:
tf.Tensor([[5 6 7 8]], shape=(1, 4), dtype=int32)

Related

How to do numpy like conditional assignment in Tensorflow

The following is how it works in Numpy
import numpy as np
vals_for_fives = [12, 18, 22, 33]
arr = np.array([5, 2, 3, 5, 5, 5])
arr[arr == 5] = vals_for_fives # It is guaranteed that length of vals_for_fives is equal to the number of fives in arr
# now the value of arr is [12, 2, 3, 18, 22, 33]
For broadcastable or constant assignment we can use where() and assign() in Tensorflow. How can we achieve the above scenario in TF?
tf.experimental.numpy.where is a thing in tensorflow v2.5.
But for now you could do this:
First find the positions of the 5's:
arr = np.array([5, 2, 3, 5, 5, 5])
where = tf.where(arr==5)
where = tf.cast(where, tf.int32)
print(where)
# <tf.Tensor: id=91, shape=(4, 1), dtype=int32, numpy=
array([[0],
[3],
[4],
[5]])>
Then use scatter_nd to "replace" elements by index:
tf.scatter_nd(where, tf.constant([12,18,22,23]), tf.constant([5]))
# <tf.Tensor: id=94, shape=(5,), dtype=int32, numpy=array([12, 0, 0, 18
, 22])>
Do a similar thing for the entries that were not 5 to find the missing tensor:
tf.scatter_nd(tf.constant([[1], [2]]), tf.constant([2,3]), tf.constant([5]))
# <tf.Tensor: id=98, shape=(5,), dtype=int32, numpy=array([0, 2, 3, 0, 0])>
Then sum the two tensors to get:
<tf.Tensor: id=113, shape=(5,), dtype=int32, numpy=array([12, 2, 3, 1, 8, 22])>

Tensorflow difference between *pool2D and *pooling2D

I am learning this TensorFlow-2.x-Tutorials where it use layers.MaxPooling2D. The autocompletion also hint layers.MaxPool2D, so I search for the difference between them.
Refer to this api_docs, I find their entire name tf.compat.v1.layers.MaxPooling2D and tf.keras.layers.MaxPool2D, which have almost same arguments, can I just consider layers.MaxPooling2D = layers.MaxPool2D, but the former is to tf1.x, the latter is to tf2.x?
What's more, I also find tf.keras.layers.GlobalMaxPool1D(Global max pooling operation for 1D temporal data) and tf.keras.layers.GlobalAveragePooling1D(Global average pooling operation for temporal data), these two have exact the same arguments, why is the syntax of function name different?
I'm only going to answer your second question because someone found a duplicate for your first one.
MaxPooling2D takes the maximum value from a 2D array. Take for example this input:
import tensorflow as tf
x = tf.random.uniform(minval=0, maxval=10, dtype=tf.int32, shape=(3, 3, 3), seed=42)
<tf.Tensor: shape=(3, 3, 3), dtype=int32, numpy=
array([[[2, 4, 3],
[9, 1, 8],
[8, 3, 5]],
[[6, 6, 9],
[9, 6, 1],
[7, 5, 2]],
[[2, 0, 8],
[1, 6, 1],
[2, 3, 9]]])>
MaxPooling2D will take the average value of all of these three elements:
gmp = tf.keras.layers.GlobalMaxPooling2D()
gmp(x[..., None])
<tf.Tensor: shape=(3, 1), dtype=int32, numpy=
array([[9],
[9],
[9]])>
There's a 9 in every elements so the operation returns a 9 for all three. For GlobalAveragePooling2D, it's the exact same thing but with averaging.
gap = tf.keras.layers.GlobalAveragePooling2D()
gap(x[..., None])
<tf.Tensor: shape=(3, 1), dtype=int32, numpy=
array([[3],
[6],
[5]])>

Broadcasting with ragged tensor

Define x as:
>>> import tensorflow as tf
>>> x = tf.constant([1, 2, 3])
Why does this normal tensor multiplication work fine with broacasting:
>>> tf.constant([[1, 2, 3], [4, 5, 6]]) * tf.expand_dims(x, axis=0)
<tf.Tensor: shape=(2, 3), dtype=int32, numpy=
array([[ 1, 4, 9],
[ 4, 10, 18]], dtype=int32)>
while this one with a ragged tensor does not?
>>> tf.ragged.constant([[1, 2, 3], [4, 5, 6]]) * tf.expand_dims(x, axis=0)
*** tensorflow.python.framework.errors_impl.InvalidArgumentError: Expected 'tf.Tensor(False, shape=(), dtype=bool)' to be true. Summarized data: b'Unable to broadcast: dimension size mismatch in dimension'
1
b'lengths='
3
b'dim_size='
3, 3
How can I get a 1-D tensor to broadcast over a 2-D ragged tensor? (I am using TensorFlow 2.1.)
The problem will be resolved if you add ragged_rank=0 to the Ragged Tensor, as shown below:
tf.ragged.constant([[1, 2, 3], [4, 5, 6]], ragged_rank=0) * tf.expand_dims(x, axis=0)
Complete working code is:
%tensorflow_version 2.x
import tensorflow as tf
x = tf.constant([1, 2, 3])
print(tf.ragged.constant([[1, 2, 3], [4, 5, 6]], ragged_rank=0) * tf.expand_dims(x, axis=0))
Output of the above code is:
tf.Tensor(
[[ 1 4 9]
[ 4 10 18]], shape=(2, 3), dtype=int32)
One more correction.
As per the definition of Broadcasting, Broadcasting is the process of **making** tensors with different shapes have compatible shapes for elementwise operations, there is no need to specify tf.expand_dims explicitly, Tensorflow will take care of it.
So, below code works and demonstrates the property of Broadcasting well:
%tensorflow_version 2.x
import tensorflow as tf
x = tf.constant([1, 2, 3])
print(tf.ragged.constant([[1, 2, 3], [4, 5, 6]], ragged_rank=0) * x)
Output of the above code is:
tf.Tensor(
[[ 1 4 9]
[ 4 10 18]], shape=(2, 3), dtype=int32)
For more information, please refer this link.
Hope this helps. Happy Learning!

Unable to expand numpy array

I have an array containing information about images. It contains information about 21495 images in an array named 'shuffled'.
np.shape(shuffled) = (21495, 1)
np.shape(shuffled[0]) = (1,)
np.shape(shuffled[0][0]) = (128, 128, 3) # (These are the image dimensions, with 3 channels of RGB)
How do I convert this array to an array of shape (21495, 128, 128, 3) to feed to my model?
There are 2 ways that I can think of:
One is using the vstack() fucntion of numpy, but it gets quite slow overtime when the size of array starts to increase.
Another way (which I use) is to take an empty list and keep appending the images array to that list using .append(), then finally convert that list to a numpy array.
Try
np.stack(shuffled[:,0])
stack, a form of concatenate, joins a list (or array) of arrays on a new initial dimension. We need to get get rid of the size 1 dimension first.
In [23]: arr = np.empty((4,1),object)
In [24]: for i in range(4): arr[i,0] = np.arange(i,i+6).reshape(2,3)
In [25]: arr
Out[25]:
array([[array([[0, 1, 2],
[3, 4, 5]])],
[array([[1, 2, 3],
[4, 5, 6]])],
[array([[2, 3, 4],
[5, 6, 7]])],
[array([[3, 4, 5],
[6, 7, 8]])]], dtype=object)
In [26]: arr.shape
Out[26]: (4, 1)
In [27]: arr[0,0].shape
Out[27]: (2, 3)
In [28]: np.stack(arr[:,0])
Out[28]:
array([[[0, 1, 2],
[3, 4, 5]],
[[1, 2, 3],
[4, 5, 6]],
[[2, 3, 4],
[5, 6, 7]],
[[3, 4, 5],
[6, 7, 8]]])
In [29]: _.shape
Out[29]: (4, 2, 3)
But beware, if the subarrays differ in shape, say one or two is b/w rather than 3 channel, this won't work.

How to use numpy as_strided (from np.stride_tricks) correctly?

I'm trying to reshape a numpy array using numpy.strided_tricks. This is the guide I'm following: https://stackoverflow.com/a/2487551/4909087
My use case is very similar, with the difference being that I need strides of 3.
Given this array:
a = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9])
I'd like to get:
array([[1, 2, 3],
[2, 3, 4],
[3, 4, 5],
[4, 5, 6],
[5, 6, 7],
[6, 7, 8],
[7, 8, 9]])
Here's what I tried:
import numpy as np
as_strided = np.lib.stride_tricks.as_strided
a = np.arange(1, 10)
as_strided(a, (len(a) - 2, 3), (3, 3))
array([[ 1, 2199023255552, 131072],
[ 2199023255552, 131072, 216172782113783808],
[ 131072, 216172782113783808, 12884901888],
[216172782113783808, 12884901888, 768],
[ 12884901888, 768, 1125899906842624],
[ 768, 1125899906842624, 67108864],
[ 1125899906842624, 67108864, 4]])
I was pretty sure I'd followed the example to a T, but evidently not. Where am I going wrong?
The accepted answer (and discussion) is good, but for the benefit of readers who don't want to run their own test case, I'll try to illustrate what's going on:
In [374]: a = np.arange(1,10)
In [375]: as_strided = np.lib.stride_tricks.as_strided
In [376]: a.shape
Out[376]: (9,)
In [377]: a.strides
Out[377]: (4,)
For a contiguous 1d array, strides is the size of the element, here 4 bytes, an int32. To go from one element to the next it steps forward 4 bytes.
What the OP tried:
In [380]: as_strided(a, shape=(7,3), strides=(3,3))
Out[380]:
array([[ 1, 512, 196608],
[ 512, 196608, 67108864],
[ 196608, 67108864, 4],
[ 67108864, 4, 1280],
[ 4, 1280, 393216],
[ 1280, 393216, 117440512],
[ 393216, 117440512, 7]])
This is stepping by 3 bytes, crossing int32 boundaries, and giving mostly unintelligable numbers. If might make more sense if the dtype had been bytes or uint8.
Instead using a.strides*2 (tuple replication), or (4,4) we get the desired array:
In [381]: as_strided(a, shape=(7,3), strides=(4,4))
Out[381]:
array([[1, 2, 3],
[2, 3, 4],
[3, 4, 5],
[4, 5, 6],
[5, 6, 7],
[6, 7, 8],
[7, 8, 9]])
Columns and rows both step one element, resulting in a 1 step moving window. We could have also set shape=(3,7), 3 windows 7 elements long.
In [382]: _.strides
Out[382]: (4, 4)
Changing strides to (8,4) steps 2 elements for each window.
In [383]: as_strided(a, shape=(7,3), strides=(8,4))
Out[383]:
array([[ 1, 2, 3],
[ 3, 4, 5],
[ 5, 6, 7],
[ 7, 8, 9],
[ 9, 25, -1316948568],
[-1316948568, 184787224, -1420192452],
[-1420192452, 0, 0]])
But shape is off, showing us bytes off the end of the original databuffer. That could be dangerous (we don't know if those bytes belong to some other object or array). With this size of array we don't get a full set of 2 step windows.
Now step 3 elements for each row (3*4, 4):
In [384]: as_strided(a, shape=(3,3), strides=(12,4))
Out[384]:
array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
In [385]: a.reshape(3,3).strides
Out[385]: (12, 4)
This is the same shape and strides as a 3x3 reshape.
We can set negative stride values and 0 values. In fact, negative-step slicing along a dimension with a positive stride will give a negative stride, and broadcasting works by setting 0 strides:
In [399]: np.broadcast_to(a, (2,9))
Out[399]:
array([[1, 2, 3, 4, 5, 6, 7, 8, 9],
[1, 2, 3, 4, 5, 6, 7, 8, 9]])
In [400]: _.strides
Out[400]: (0, 4)
In [401]: a.reshape(3,3)[::-1,:]
Out[401]:
array([[7, 8, 9],
[4, 5, 6],
[1, 2, 3]])
In [402]: _.strides
Out[402]: (-12, 4)
However, negative strides require adjusting which element of the original array is the first element of the view, and as_strided has no parameter for that.
I have no idea why you think you need strides of 3. You need strides the distance in bytes between one element of a and the next, which you can get using a.strides:
as_strided(a, (len(a) - 2, 3), a.strides*2)
I was trying to do a similar operation and run into the same problem.
In your case, as stated in this comment, the problems were:
You were not taking into account the size of your element when stored in memory (int32 = 4, which can be checked using a.dtype.itemsize).
You didn't specify appropriately the number of strides you had to skip, which in your case were also 4, as you were skipping only one element.
I made myself a function based on this answer, in which I compute the segmentation of a given array, using a window of n-elements and specifying the number of elements to overlap (given by window - number_of_elements_to_skip).
I share it here in case someone else needs it, since it took me a while to figure out how stride_tricks work:
def window_signal(signal, window, overlap):
"""
Windowing function for data segmentation.
Parameters:
------------
signal: ndarray
The signal to segment.
window: int
Window length, in samples.
overlap: int
Number of samples to overlap
Returns:
--------
nd-array
A copy of the signal array with shape (rows, window),
where row = (N-window)//(window-overlap) + 1
"""
N = signal.reshape(-1).shape[0]
if (window == overlap):
rows = N//window
overlap = 0
else:
rows = (N-window)//(window-overlap) + 1
miss = (N-window)%(window-overlap)
if(miss != 0):
print('Windowing led to the loss of ', miss, ' samples.')
item_size = signal.dtype.itemsize
strides = (window - overlap) * item_size
return np.lib.stride_tricks.as_strided(signal, shape=(rows, window),
strides=(strides, item_size))
The solution for this case is, according to your code:
as_strided(a, (len(a) - 2, 3), (4, 4))
Alternatively, using the function window_signal:
window_signal(a, 3, 2)
Both return as output the following array:
array([[1, 2, 3],
[2, 3, 4],
[3, 4, 5],
[4, 5, 6],
[5, 6, 7],
[6, 7, 8],
[7, 8, 9]])

Categories