How to mask row in Tensorflow without for loop - python

I want to create a custom Layer for a Tensorflow model but the logic I have uses a for loop, which Tensorflow doesn't like. How can I modify my code to remove the for loop but still achieve the same result?
class CustomMask(tf.keras.layers.Layer):
def call(self, inputs):
mask = tf.where(inputs[:, 0] < 0.5, 1, 0)
for i,m in enumerate(mask):
if m:
inputs = inputs[i, 1:].assign(tf.zeros(4, dtype=tf.float32))
else:
first = tf.where(inputs[:, 1] >= 0.5, 0, 1)
assign = tf.multiply(tf.cast(first, tf.float32), inputs[:, 2])
inputs = inputs[:, 2].assign(assign)
third = tf.where(inputs[:, 1] >= 0.5, 1, 0)
assign = tf.multiply(tf.cast(third, tf.float32), inputs[:, 1])
inputs = inputs[:, 1].assign(assign)
return inputs
Example input Tensor:
<tf.Variable 'Variable:0' shape=(3, 5) dtype=float32, numpy=
array([[0.8, 0.7, 0.2, 0.6, 0.9],
[0.8, 0.4, 0.8, 0.3, 0.7],
[0.3, 0.2, 0.4, 0.3, 0.8]], dtype=float32)>
Corresponding output:
<tf.Variable 'UnreadVariable' shape=(3, 5) dtype=float32, numpy=
array([[0.8, 0.7, 0. , 0.6, 0.9],
[0.8, 0. , 0.8, 0.3, 0.7],
[0.3, 0. , 0. , 0. , 0. ]], dtype=float32)>
EDIT:
The layer should take an array of shape (batch_size, 5) and if the first value of a row is less than 0.5, set the rest of the row values to 0, otherwise if the 2nd element is above 0.5, set the 3rd element to 0 and if the 3rd element is greater than 0.5, set the 2nd element to 0

Without using any foor loop, ask in comments if it doesn't solve your issue (tested in colab)
import tensorflow as tf
mask1 = tf.convert_to_tensor([0.0,1.0,1.0,1.0,1.0])
mask2 = tf.convert_to_tensor([0.0,0.0,1.0,0.0,0.0])
mask3 = tf.convert_to_tensor([0.0,1.0,0.0,0.0,0.0])
def masking(x):
mask = tf.ones(x.shape, tf.float32)
cond1 = tf.cast(x[0] < 0.5, tf.float32)
x = tf.multiply(x, tf.subtract(mask, tf.multiply(mask1, cond1)))
cond2 = tf.cast(x[1] > 0.5, tf.float32)
x = tf.multiply(x, tf.subtract(mask, tf.multiply(mask2, cond2)))
cond3 = tf.cast(x[2] > 0.5, tf.float32)
x = tf.multiply(x, tf.subtract(mask, tf.multiply(mask3, cond3)))
return x
inputs = tf.convert_to_tensor([[0.8, 0.7, 0.2, 0.6, 0.9],
[0.8, 0.4, 0.8, 0.3, 0.7],
[0.3, 0.2, 0.4, 0.3, 0.8]])
res = tf.vectorized_map(masking, inputs)
print (res)
tf.Tensor(
[[0.8 0.7 0. 0.6 0.9]
[0.8 0. 0.8 0.3 0.7]
[0.3 0. 0. 0. 0. ]], shape=(3, 5), dtype=float32)
I tested it with
%timeit tf.map_fn(masking, inputs)
%timeit tf.vectorized_map(masking, inputs)
and the tf.vectorized_map(masking, inputs) get faster when the batch size increase

Related

Extract arrays based on positions indicated in another array

I have the data below as an example:
import numpy as np
data=[np.array([[0.9,0.6,0.5,0.4,0.7],[0.8,0.0,0.0,0.8,0.2],
[0.9,0.0,0.4,0.4,0.3],[0.9,0.6,0.3,0.2,0.5],[0.8,0.0,0.3,0.1,0.5]]),
np.array([[0.9,0.0,0.2,0.4,0.3],[0.0,0.2,0.4,0.0,0.0],
[0.0,0.0,0.0,0.2,0.0],[0.5,0.0,0.3,0.6,0.8],[0.5,0.6,0.9,0.0,0.0]])]
and I want to extract the relevant data based on these positions below:
positions_non_zero=[np.array([2,3,4]),np.array([1,4])]
the desired output should be this:
[array([[0.9, 0. , 0.4, 0.4, 0.3],
[0.9, 0.6, 0.3, 0.2, 0.5],
[0.8, 0. , 0.3, 0.1, 0.5]]),
array([[0. , 0.2, 0.4, 0. , 0. ],
[0.5, 0.6, 0.9, 0. , 0. ]])]
The reason is this:
The problem with my code is that only the np.array([1,4]) is taken under consideration.
My code:
df_class11=[]
for n in data:
def data_target(df_class_target):
for z in df_class_target:
x_classA=[n[i] for i in z]
x_classA=np.vstack(x_classA)
return x_classA
df_class11.append(data_target(positions_non_zero))
df_class11

Is this a Python/Numpy bug or a subtle gotcha?

Consider the following two implemtations of the same piece of code. I would have thought they are identical but they are not.
Is this a Python/Numpy bug or a subtle gotcha? If the latter, what rule would make it obvious why it does not work as expected?
I was working with multiple arrays of data and having to process each array item by item, with each array manipulated by a table depending on it's metadata.
In the real world example 'n' is multiple factors and offsets but the following code still demonstrates the issue that I was getting the wrong result in all but one case.
import numpy as np
# Change the following line to True to show different behaviour
NEEDS_BUGS = False # Changeme
# Create some data
data = np.linspace(0, 1, 10)
print(data)
# Create an array of vector functions each of which does a different operation on a set of data
vfuncd = dict()
# Two implementations
if NEEDS_BUGS:
# Lets do this in a loop because we like loops - However WARNING this does not work!!
for n in range(10):
vfuncd[n] = np.vectorize(lambda x: x * n)
else:
# Unwrap the loop - NOTE: Spoiler - this works
vfuncd[0] = np.vectorize(lambda x: x * 0)
vfuncd[1] = np.vectorize(lambda x: x * 1)
vfuncd[2] = np.vectorize(lambda x: x * 2)
vfuncd[3] = np.vectorize(lambda x: x * 3)
vfuncd[4] = np.vectorize(lambda x: x * 4)
vfuncd[5] = np.vectorize(lambda x: x * 5)
vfuncd[6] = np.vectorize(lambda x: x * 6)
vfuncd[7] = np.vectorize(lambda x: x * 7)
vfuncd[8] = np.vectorize(lambda x: x * 8)
vfuncd[9] = np.vectorize(lambda x: x * 9)
# Prove we have multiple different vectorised functions
for k, vfunc in vfuncd.items():
print(k, vfunc)
# Do the work
res = {k: vfuncd[k](data) for k in vfuncd.keys()}
# Show the result
for k, r in res.items():
print(k, r)
I don't know what exactly you're trying to achieve and if it's a bad idea or not (in terms of np.vectorize), but the issue you're facing is because of the way python makes closures. Quoting from an answer to the linked question:
Scoping in Python is lexical. A closure will always
remember the name and scope of the variable, not the object it's
pointing to. Since all the functions in your example are created in
the same scope and use the same variable name, they always refer to
the same variable.
in other words when you make that closure over n, you're not actually closing off the state of n, just the name. So when n changes, the value in your closure also changes. This is quite unexpected to me, but others find it natural.
Here is one fix using partial:
from functools import partial
.
.
.
def func(x, n):
return x * n
for n in range(10):
vfuncd[n] = np.vectorize(partial(func, n=n))
Or another using a factory method
def func_factory(n):
return lambda x: x * n
for n in range(10):
vfuncd[n] = np.vectorize(func_factory(n))
It seems that the python variable n is bound to the vectorized expression:
for n in range(10):
vfuncd[n] = np.vectorize(lambda x: x * n)
This fixes it as it creates a new object with which to bind:
for n in range(10):
vfuncd[n] = np.vectorize(lambda x: x * np.scalar(n))
In fact this has implications in terms of performance as I assume the value of the python variable would have to be fetched repeatedly.
In [13]: data = np.linspace(0,1,11)
Since the data array can be multiplied with a simple:
In [14]: data*3
Out[14]: array([0. , 0.3, 0.6, 0.9, 1.2, 1.5, 1.8, 2.1, 2.4, 2.7, 3. ])
we don't need the complication of np.vectorize to see the closure issue. A simple lambda is enough.
In [15]: vfuncd = {}
...: for n in range(3):
...: vfuncd[n] = lambda x:x*n
...:
In [16]: vfuncd
Out[16]:
{0: <function __main__.<lambda>(x)>,
1: <function __main__.<lambda>(x)>,
2: <function __main__.<lambda>(x)>}
In [17]: {k:v(data) for k,v in vfuncd.items()}
Out[17]:
{0: array([0. , 0.2, 0.4, 0.6, 0.8, 1. , 1.2, 1.4, 1.6, 1.8, 2. ]),
1: array([0. , 0.2, 0.4, 0.6, 0.8, 1. , 1.2, 1.4, 1.6, 1.8, 2. ]),
2: array([0. , 0.2, 0.4, 0.6, 0.8, 1. , 1.2, 1.4, 1.6, 1.8, 2. ])}
We won't get the closure problem if we use a proper numpy "vectorization":
In [18]: data * np.arange(3)[:,None]
Out[18]:
array([[0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[0. , 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. ],
[0. , 0.2, 0.4, 0.6, 0.8, 1. , 1.2, 1.4, 1.6, 1.8, 2. ]])
Or a simple iteration is we need a dictionary:
In [20]: {k:data*k for k in range(3)}
Out[20]:
{0: array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]),
1: array([0. , 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. ]),
2: array([0. , 0.2, 0.4, 0.6, 0.8, 1. , 1.2, 1.4, 1.6, 1.8, 2. ])}
np.vectorize has a speed disclaimer. But it is justified where the function only takes scalar inputs, and we want the flexibility of numpy broadcasting - i.e. for 2 or more arguments.
Creating multiple vectorize is clearly an 'anti-pattern'. I'd rather see one vectorize with the appropriate arguments:
In [25]: f = np.vectorize(lambda x,n: x*n)
In [26]: {n: f(data,n) for n in range(3)}
Out[26]:
{0: array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]),
1: array([0. , 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. ]),
2: array([0. , 0.2, 0.4, 0.6, 0.8, 1. , 1.2, 1.4, 1.6, 1.8, 2. ])}
That f can also produce the array Out[18] (but is slower):
In [27]: f(data, np.arange(3)[:,None])
Out[27]:
array([[0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[0. , 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. ],
[0. , 0.2, 0.4, 0.6, 0.8, 1. , 1.2, 1.4, 1.6, 1.8, 2. ]])

Return the biggest value less than one from a numpy vector

I have a numpy vector in python and I want to find the index of the max value of the vector with the condition that it is less than one. I have as an example the following:
temp_res = [0.9, 0.8, 0.7, 0.99, 1.2, 1.5, 0.1, 0.5, 0.1, 0.01, 0.12, 0.56, 0.89, 0.23, 0.56, 0.78]
temp_res = np.asarray(temp_res)
indices = np.where((temp_res == temp_res.max()) & (temp_res < 1))
However, what I tried always return an empty matrix since those two conditions cannot be met. HU want to return as final result the index = 3 which correspond to 0.99 the biggest value that it is less than 1. How can I do so?
You need to perform the max() function after filtering your array:
temp_res = np.asarray(temp_res)
temp_res[temp_res < 1].max()
Out[60]: 0.99
If you want to find all the indexes, here is a more genera approach:
mask = temp_res < 1
indices = np.where(mask)
maximum = temp_res[mask].max()
max_indices = np.where(temp_res == maximum)
Example:
...: temp_res = [0.9, 0.8, 0.7, 1, 0.99, 0.99, 1.2, 1.5, 0.1, 0.5, 0.1, 0.01, 0.12, 0.56, 0.89, 0.23, 0.56, 0.78]
...: temp_res = np.asarray(temp_res)
...: mask = temp_res < 1
...: indices = np.where(mask)
...: maximum = temp_res[mask].max()
...: max_indices = np.where(temp_res == maximum)
...:
In [72]: max_indices
Out[72]: (array([4, 5]),)
You can use:
np.where(temp_res == temp_res[temp_res < 1].max())[0]
Example:
In [49]: temp_res
Out[49]:
array([0.9 , 0.8 , 0.7 , 0.99, 1.2 , 1.5 , 0.1 , 0.5 , 0.1 , 0.01, 0.12,
0.56, 0.89, 0.23, 0.56, 0.78])
In [50]: np.where(temp_res == temp_res[temp_res < 1].max())[0]
...:
Out[50]: array([3])

TensorFlow - dense vector to one-hot

Suppose I have the following tensor:
T = [[0.1, 0.3, 0.7],
[0.2, 0.5, 0.3],
[0.1, 0.1, 0.8]]
I want to transform this into a one-hot tensor, such that the indexes with the maximum value over dimension 0 get set to 1 and all the other ones get set to zero, like this:
T_onehot = [[0, 0, 1],
[0, 1, 0],
[0, 0, 1]]
I know there's tf.argmax to get the indices of the largest elements in the tensor, but is there any method which allows me to do what I want to do in one step?
I don't know if there's a way to do this in one step, but there's a one_hot function in tensorflow:
import tensorflow as tf
T = tf.constant([[0.1, 0.3, 0.7], [0.2, 0.5, 0.3], [0.1, 0.1, 0.8]])
T_onehot = tf.one_hot(tf.argmax(T, 1), T.shape[1])
tf.InteractiveSession()
print(T_onehot.eval())
# [[ 0. 0. 1.]
# [ 0. 1. 0.]
# [ 0. 0. 1.]]

What is the equivalent of "zip()" in Python's numpy?

I am trying to do the following but with numpy arrays:
x = [(0.1, 1.), (0.1, 2.), (0.1, 3.), (0.1, 4.), (0.1, 5.)]
normal_result = zip(*x)
This should give a result of:
normal_result = [(0.1, 0.1, 0.1, 0.1, 0.1), (1., 2., 3., 4., 5.)]
But if the input vector is a numpy array:
y = np.array(x)
numpy_result = zip(*y)
print type(numpy_result)
It (expectedly) returns a:
<type 'list'>
The issue is that I will need to transform the result back into a numpy array after this.
What I would like to know is what is if there is an efficient numpy function that will avoid these back-and-forth transformations?
You can just transpose it...
>>> a = np.array([(0.1, 1.), (0.1, 2.), (0.1, 3.), (0.1, 4.), (0.1, 5.)])
>>> a
array([[ 0.1, 1. ],
[ 0.1, 2. ],
[ 0.1, 3. ],
[ 0.1, 4. ],
[ 0.1, 5. ]])
>>> a.T
array([[ 0.1, 0.1, 0.1, 0.1, 0.1],
[ 1. , 2. , 3. , 4. , 5. ]])
Try using dstack:
>>> from numpy import *
>>> a = array([[1,2],[3,4]]) # shapes of a and b can only differ in the 3rd dimension (if present)
>>> b = array([[5,6],[7,8]])
>>> dstack((a,b)) # stack arrays along a third axis (depth wise)
array([[[1, 5],
[2, 6]],
[[3, 7],
[4, 8]]])
so in your case it would be:
x = [(0.1, 1.), (0.1, 2.), (0.1, 3.), (0.1, 4.), (0.1, 5.)]
y = np.array(x)
np.dstack(y)
>>> array([[[ 0.1, 0.1, 0.1, 0.1, 0.1],
[ 1. , 2. , 3. , 4. , 5. ]]])

Categories

Resources