Value Assignment by specific indices error in Tensorflow-CodePudding

I'm trying to build a custom loss function for my model, but whenever I try to convert Tensors into .numpy() arrays with run_eagerly = True, it gives "WARNING: Gradients do not exist for variables ...". So I debugged other custom loss functions implemented using TensorFlow. But in my case, I need to apply to mask and splitting index arrays and then use those arrays as indices to apply some sort of arithmetic functions using broadcasting. But I retrieved indices lists after masking, but I just have to access those indices and add specific functions. But I found no way in TensorFlow to implement that in a vectorized way.

error = y_true - y_false
print(y_true.shape, y_pred.shape)
print(error.shape)
print("Error values: ", error)

Output: (10, 1000), (10, 1000)

(10, 1000)

Error values: <tf.Tensor: shape=(10, 1000), dtype=float64, numpy= array([[-10, 0, 8, ..., 3, -1.5, -2.5], ..., [ 2.5, 8 , 6.5, ..., 5.5, 3.5, -0.5]])>

mask = tf.where(y_true > 5)
i = mask[0]
j = mask[1]
print(i[:5])
print(j[:5])

Results:

(<tf.Tensor: shape=(5,), dtype=int64, numpy=array([0, 0, 0, 0, 0], dtype=int64)>,

<tf.Tensor: shape=(5,), dtype=int64, numpy=array([19, 26, 28, 35, 39], dtype=int64)>)

In NumPy, I can access it using:

error[i, j] = error[i, j] * 5

What I want is to replace with new values of error in a specified position after executing the above code and get values like:

Error values: <tf.Tensor: shape=(10, 1000), dtype=float64, numpy= array([[-10, 0, 16*, ..., 3, -1.5, -2.5], ..., [ 2.5, 16* , 13*, ..., 11*, 3.5, -0.5]])>

But when I try to execute this as Tensors, it gives the following error:

TypeError                                 Traceback (most recent call last)
Input In [193], in <cell line: 1>()
----> 1 error[i, j]

File ~\AppData\Local\Programs\Python\Python310\lib\site-packages\tensorflow\python\util\traceback_utils.py:153, in filter_traceback.<locals>.error_handler(*args, **kwargs)
    151 except Exception as e:
    152   filtered_tb = _process_traceback_frames(e.__traceback__)
--> 153   raise e.with_traceback(filtered_tb) from None
    154 finally:
    155   del filtered_tb

File ~\AppData\Local\Programs\Python\Python310\lib\site-packages\tensorflow\python\ops\array_ops.py:899, in _check_index(idx)
    894 dtype = getattr(idx, "dtype", None)
    895 if (dtype is None or dtypes.as_dtype(dtype) not in _SUPPORTED_SLICE_DTYPES or
    896     idx.shape and len(idx.shape) == 1):
    897   # TODO(slebedev): IndexError seems more appropriate here, but it
    898   # will break `_slice_helper` contract.
--> 899   raise TypeError(_SLICE_TYPE_ERROR   ", got {!r}".format(idx))

TypeError: Only integers, slices (`:`), ellipsis (`...`), tf.newaxis (`None`) and scalar tf.int32/tf.int64 tensors are valid indices, got <tf.Tensor: shape=(4797,), dtype=int64, numpy=array([ 0,  0,  0, ..., 26, 26, 26], dtype=int64)>

I tried using other functions provided by TensorFlow too, but those did not work.

CodePudding user response：

You can't use indices assignement in tensorflow, but this can be overcome using tf.gather_nd and tf.scatter_nd.
Here is an example with dummy input data

y_true = tf.random.uniform(shape=(10, 6), minval=0, maxval=10, dtype=tf.int32)
mask = tf.where(y_true > 5)
y_mask = tf.scatter_nd(mask, tf.gather_nd(y_true, mask), shape=tf.cast(tf.shape(y_true), tf.int64))
print(y_true)
print(y_mask)
print(y_mask*4 y_true)

outputs
y_true:

tf.Tensor(
[[9 9 0 1 4 2]
 [9 8 6 3 9 8]
 [9 0 1 4 7 7]
 [4 8 6 3 4 1]
 [9 1 8 9 3 9]
 [2 8 3 4 9 2]
 [2 5 7 5 2 2]
 [6 7 6 7 9 4]
 [2 8 9 5 2 1]
 [7 4 1 9 7 9]], shape=(10, 6), dtype=int32)

y_mask:

tf.Tensor(
[[9 9 0 0 0 0]
 [9 8 6 0 9 8]
 [9 0 0 0 7 7]
 [0 8 6 0 0 0]
 [9 0 8 9 0 9]
 [0 8 0 0 9 0]
 [0 0 7 0 0 0]
 [6 7 6 7 9 0]
 [0 8 9 0 0 0]
 [7 0 0 9 7 9]], shape=(10, 6), dtype=int32)

y_mask*4 y_true:

tf.Tensor(
[[45 45  0  1  4  2]
 [45 40 30  3 45 40]
 [45  0  1  4 35 35]
 [ 4 40 30  3  4  1]
 [45  1 40 45  3 45]
 [ 2 40  3  4 45  2]
 [ 2  5 35  5  2  2]
 [30 35 30 35 45  4]
 [ 2 40 45  5  2  1]
 [35  4  1 45 35 45]], shape=(10, 6), dtype=int32)

Another solution is to use directly the mask and cast it to int32 or float32 (or other):

mask_float = tf.cast(y_true>5, tf.int32)
print(mask_float*y_true)

y_mask v2:

tf.Tensor(
[[9 9 0 0 0 0]
 [9 8 6 0 9 8]
 [9 0 0 0 7 7]
 [0 8 6 0 0 0]
 [9 0 8 9 0 9]
 [0 8 0 0 9 0]
 [0 0 7 0 0 0]
 [6 7 6 7 9 0]
 [0 8 9 0 0 0]
 [7 0 0 9 7 9]], shape=(10, 6), dtype=int32)

CodePudding user response：

If I understand your comment correctly, the following code should be appropriate. It multiplies by alpha (=2. here) all error values greater than 5

# generate some error tensor
error = tf.random.uniform(shape=(10, 3), minval=0, maxval=10, dtype=tf.float64)
print('input error')
print(error)
float_mask = tf.cast(error>5, dtype=tf.float64)
print('mask')
print(float_mask)
alpha = 2.
print('gain = %f' % alpha)
error = error   (alpha-1.)*float_mask*error
print('output error')
print(error)

which gives:

input error

tf.Tensor( [[9.47020833 6.21211945 2.56257082]  [8.2855179
6.23372048 9.39559957]  [5.2926297  2.62602144 4.44665184]  [6.49200992 7.09389259 1.04311547]  [9.39402112 2.68713794 7.71738653] [6.4853496  2.99997236 9.88983946]  [3.57130888 5.73827016 5.91022104] [2.58102132 4.01791191 3.19829238]  [9.28263857 4.73230455 6.24950981] [0.38713425 3.56589859 8.74955686]], shape=(10, 3), dtype=float64)

mask

tf.Tensor( [[1. 1. 0.]  [1. 1. 1.]  [1. 0. 0.]  [1. 1. 0.]  [1.
0. 1.]  [1. 0. 1.]  [0. 1. 1.]  [0. 0. 0.]  [1. 0. 1.]  [0. 0. 1.]], shape=(10, 3), dtype=float64)

gain

gain = 2.000000

output error

tf.Tensor( [[18.94041665 12.42423889  2.56257082]  [16.5710358  12.46744096
18.79119913]  [10.58525941  2.62602144  4.44665184]  [12.98401983 14.18778517  1.04311547]  [18.78804224  2.68713794 15.43477305]  [12.9706992   2.99997236 19.77967893]  [ 3.57130888 11.47654031
11.82044208]  [ 2.58102132  4.01791191  3.19829238]  [18.56527714  4.73230455 12.49901962]  [ 0.38713425  3.56589859 17.49911371]], shape=(10, 3), dtype=float64)