替换“ tf.gather_nd”

时间:2019-06-04 23:58:27

标签: python tensorflow deep-learning

我正在做一个项目,但其tensorflow版本不支持tf.gather_nd。我问是否可以使用tf.gather,tf.slice或tf.strided_slice重写tf.gather_nd的功能?

tf.gather_nd用于将切片从张量收集到张量中,张量具有由索引指定的形状。详细信息可以在https://www.tensorflow.org/api_docs/python/tf/gather_nd

中找到

谢谢

1 个答案:

答案 0 :(得分:0)

此功能应做等效的工作:

import tensorflow as tf
import numpy as np

def my_gather_nd(params, indices):
    idx_shape = tf.shape(indices)
    params_shape = tf.shape(params)
    idx_dims = idx_shape[-1]
    gather_shape = params_shape[idx_dims:]
    params_flat = tf.reshape(params, tf.concat([[-1], gather_shape], axis=0))
    axis_step = tf.cumprod(params_shape[:idx_dims], exclusive=True, reverse=True)
    indices_flat = tf.reduce_sum(indices * axis_step, axis=-1)
    result_flat = tf.gather(params_flat, indices_flat)
    return tf.reshape(result_flat, tf.concat([idx_shape[:-1], gather_shape], axis=0))

# Test
np.random.seed(0)
with tf.Graph().as_default(), tf.Session() as sess:
    params = tf.constant(np.random.rand(10, 20, 30).astype(np.float32))
    indices = tf.constant(np.stack([np.random.randint(10, size=(5, 8)),
                                    np.random.randint(20, size=(5, 8))], axis=-1))
    result1, result2 = sess.run((tf.gather_nd(params, indices),
                                 my_gather_nd(params, indices)))
    print(np.allclose(result1, result2))
    # True