Using loops, tf.Variable, tf.tensorarray inside a tf.function function in tensorflow

Question

I have been trying to get this to work for a while but I don't know how. Ive tried a dozen combinations.

What I'm trying to do is fetch the input array, compare each element to see if its positive or negative. Apply a unique transformation depending on positivity or negativity. Take the mean and return the result. This is what I have but it doesn't work.

edit: my latest attempt

@tf.function
def transform(x, y):
    et = tf.math.subtract(x, y)
    et_variable = tf.Variable(et)

    max_loop = tf.shape(x)[0]

    loss = tf.TensorArray(tf.float32, size=et.shape[0], clear_after_read=False)
    loss = loss.unstack(et_variable)

    i = tf.constant(0)
    while tf.math.less(i, max_loop):
        if loss.gather([i])[0] > 0:
            val = 3 * loss.gather([i])[0]
        else:
            val = 2 * loss.gather([i])[0]
        loss = loss.scatter([i], value=[val])
        i = tf.Variable(i)
        i = i + 1
        i = tf.constant(i)
    return(K.mean(loss.stack()))

I get an error:

ValueError: condition of if statement expected to be `tf.bool` scalar, got Tensor("loss_funk/while/Greater:0", shape=(11,), dtype=bool); to check for None, use `is not None`

I've tried using tf.Variable and tf.tensorarray but neither work with tf.function.

FancyXun · Accepted Answer · 2021-08-31 02:29:10Z

2

Below is an example to meet your expectations. Tensor does not have gather_ Nd, it is tensorflow's API. My tensorflow version is 2.3

import numpy as np
import tensorflow as tf
from tensorflow.keras import backend as K

y_true = np.random.normal(size=(30, 50)).astype('float32')
x = np.random.normal(size=(30, )).astype('float32')
y = np.random.normal(size=(30, )).astype('float32')


@tf.function
def transform(x, y):
    et = tf.math.subtract(x, y)
    max_loop = tf.shape(y_true)[0]
    res = tf.TensorArray(tf.float32, size=et.shape[0], clear_after_read=False)
    i = tf.constant(0)
    while tf.math.less(i, max_loop):
        if tf.gather_nd(et, [i]) > 0:
            val = 0.03 * tf.gather_nd(et, [i])
        else:
            val = 0.05 * tf.gather_nd(et, [i])
        res = res.write(i, val)
        i = i + 1
    return K.mean(res.stack())


transform(x, y)

edited Aug 31, 2021 at 2:29

answered Aug 31, 2021 at 2:22

FancyXun

1,2981 gold badge10 silver badges17 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

intuition Over a year ago

Thanks but it didn't work. I get an error on the line: if tf.gather_nd(et, [i]) > 0: ValueError: condition of if statement expected to be tf.bool scalar, got Tensor("transform/while/Greater:0", shape=(11,), dtype=bool);

FancyXun Over a year ago

what's your tensorflow version? this work in tf 2.3

intuition Over a year ago

ok it works on a headless linux server but not my windows machine?

FancyXun Over a year ago

It seems to be independent of platform. I guess your x is 2-d tensor?

Collectives™ on Stack Overflow

Using loops, tf.Variable, tf.tensorarray inside a tf.function function in tensorflow

1 Answer 1

4 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Your Answer

Sign up or log in

Post as a guest

Related