Module kerod.utils.ops

None

None

View Source

import tensorflow as tf

def indices_to_dense_vector(indices, size, indices_value=1., default_value=0, dtype=tf.float32):

    """Creates dense vector with indices set to specific value and rest to zeros.

    This function exists because it is unclear if it is safe to use

        tf.sparse_to_dense(indices, [size], 1, validate_indices=False)

    with indices which are not ordered.

    This function accepts a dynamic size (e.g. tf.shape(tensor)[0])

    Arguments:

    - *indices*: 1d Tensor with integer indices which are to be set to

            indices_values.

    - *size*: scalar with size (integer) of output Tensor.

    - *indices_value*: values of elements specified by indices in the output vector

    - *default_value*: values of other elements in the output vector.

    - *dtype*: data type.

    Returns:

    A dense 1D Tensor of shape [size] with indices set to indices_values and the

            rest set to default_value.

    """

    size = tf.cast(size, dtype=tf.int32)

    zeros = tf.ones([size], dtype=dtype) * default_value

    values = tf.ones_like(indices, dtype=dtype) * indices_value

    return tf.dynamic_stitch([tf.range(size), tf.cast(indices, dtype=tf.int32)], [zeros, values])

def item_assignment(tensor: tf.Tensor, indicator: tf.Tensor, val):

    """Set the indicated fields of tensor to val.

    ```python

    tensor = tf.constant([1, 2, 3, 4])

    # won't work in tensorflow

    tensor[tensor == 2] = 1

    tensor = item_assignment(tensor, tensor == 2, 1)

    ```

    Arguments:

    - *tensor*: A tensor without shape constraint.

    - *indicator*: boolean tensor with same shape as `tensor`.

    - *val*: scalar with value to set.

    """

    indicator = tf.cast(indicator, tensor.dtype)

    return tensor * (1 - indicator) + val * indicator

def get_full_indices(indices):

    """ This operation allows to extract full indices from indices.

    These full-indices have the proper format for gather_nd operations.

    Example:

    ```python

    indices = [[0, 1], [2, 3]]

    full_indices = [[[0, 0], [0, 1]], [[1, 2], [1, 3]]]

    ```

    Arguments:

    - *indices*: Indices without their assorciated batch format [batch_size, k].

    Returns:

    Full-indices tensor [batch_size, k, 2]

    """

    batch_size = tf.shape(indices)[0]

    num_elements = tf.shape(indices)[1]

    batch_ids = tf.range(0, batch_size)

    # Repeat the batch indice for every indices per batch

    # [0, 1, ..., n] => [[0, ..., 0], [1, ..., 1], ..., [n, ..., n]]

    batch_ids = tf.tile(batch_ids[:, None], [1, num_elements])

    # [[a1, ..., au], ...,[n1, ..., nu]] =>

    # [[[0, a1], ..., [0, au]], ..., [[n, n1], ..., [n, nu]]]

    full_indices = tf.concat(

        [batch_ids[..., None], indices[..., None]],

         axis=-1

    )

    return full_indices

Functions

get_full_indices

def get_full_indices(
    indices
)

This operation allows to extract full indices from indices.

These full-indices have the proper format for gather_nd operations.

Example:

indices = [[0, 1], [2, 3]]
full_indices = [[[0, 0], [0, 1]], [[1, 2], [1, 3]]]

Arguments:

• indices: Indices without their assorciated batch format [batch_size, k].

Returns:

Full-indices tensor [batch_size, k, 2]

View Source

def get_full_indices(indices):

    """ This operation allows to extract full indices from indices.

    These full-indices have the proper format for gather_nd operations.

    Example:

    ```python

    indices = [[0, 1], [2, 3]]

    full_indices = [[[0, 0], [0, 1]], [[1, 2], [1, 3]]]

    ```

    Arguments:

    - *indices*: Indices without their assorciated batch format [batch_size, k].

    Returns:

    Full-indices tensor [batch_size, k, 2]

    """

    batch_size = tf.shape(indices)[0]

    num_elements = tf.shape(indices)[1]

    batch_ids = tf.range(0, batch_size)

    # Repeat the batch indice for every indices per batch

    # [0, 1, ..., n] => [[0, ..., 0], [1, ..., 1], ..., [n, ..., n]]

    batch_ids = tf.tile(batch_ids[:, None], [1, num_elements])

    # [[a1, ..., au], ...,[n1, ..., nu]] =>

    # [[[0, a1], ..., [0, au]], ..., [[n, n1], ..., [n, nu]]]

    full_indices = tf.concat(

        [batch_ids[..., None], indices[..., None]],

         axis=-1

    )

    return full_indices

indices_to_dense_vector

def indices_to_dense_vector(
    indices,
    size,
    indices_value=1.0,
    default_value=0,
    dtype=tf.float32
)

Creates dense vector with indices set to specific value and rest to zeros.

This function exists because it is unclear if it is safe to use tf.sparse_to_dense(indices, [size], 1, validate_indices=False) with indices which are not ordered. This function accepts a dynamic size (e.g. tf.shape(tensor)[0])

Arguments:

• indices: 1d Tensor with integer indices which are to be set to indices_values.
• size: scalar with size (integer) of output Tensor.
• indices_value: values of elements specified by indices in the output vector
• default_value: values of other elements in the output vector.
• dtype: data type.

Returns:

A dense 1D Tensor of shape [size] with indices set to indices_values and the rest set to default_value.

View Source

def indices_to_dense_vector(indices, size, indices_value=1., default_value=0, dtype=tf.float32):

    """Creates dense vector with indices set to specific value and rest to zeros.

    This function exists because it is unclear if it is safe to use

        tf.sparse_to_dense(indices, [size], 1, validate_indices=False)

    with indices which are not ordered.

    This function accepts a dynamic size (e.g. tf.shape(tensor)[0])

    Arguments:

    - *indices*: 1d Tensor with integer indices which are to be set to

            indices_values.

    - *size*: scalar with size (integer) of output Tensor.

    - *indices_value*: values of elements specified by indices in the output vector

    - *default_value*: values of other elements in the output vector.

    - *dtype*: data type.

    Returns:

    A dense 1D Tensor of shape [size] with indices set to indices_values and the

            rest set to default_value.

    """

    size = tf.cast(size, dtype=tf.int32)

    zeros = tf.ones([size], dtype=dtype) * default_value

    values = tf.ones_like(indices, dtype=dtype) * indices_value

    return tf.dynamic_stitch([tf.range(size), tf.cast(indices, dtype=tf.int32)], [zeros, values])

item_assignment

def item_assignment(
    tensor: tensorflow.python.framework.ops.Tensor,
    indicator: tensorflow.python.framework.ops.Tensor,
    val
)

Set the indicated fields of tensor to val.

tensor = tf.constant([1, 2, 3, 4])
# won't work in tensorflow
tensor[tensor == 2] = 1

tensor = item_assignment(tensor, tensor == 2, 1)

Arguments:

• tensor: A tensor without shape constraint.
• indicator: boolean tensor with same shape as tensor.
• val: scalar with value to set.

View Source

def item_assignment(tensor: tf.Tensor, indicator: tf.Tensor, val):

    """Set the indicated fields of tensor to val.

    ```python

    tensor = tf.constant([1, 2, 3, 4])

    # won't work in tensorflow

    tensor[tensor == 2] = 1

    tensor = item_assignment(tensor, tensor == 2, 1)

    ```

    Arguments:

    - *tensor*: A tensor without shape constraint.

    - *indicator*: boolean tensor with same shape as `tensor`.

    - *val*: scalar with value to set.

    """

    indicator = tf.cast(indicator, tensor.dtype)

    return tensor * (1 - indicator) + val * indicator