machine-learning - 我如何将 bool 张量输入到 tf.cond() 而不仅仅是一个 bool 值？

Question

这是我想用 tensorflow 实现 f(x)

输入x = (x1,x2,x3,x4,x5,x6,x7,x8,x9)

定义f(x) = f1(x1,x2,x3,x4,x5) + f2(x5,x6,x7,x8,x9)

哪里

f1(x1,x2,x3,x4,x5) = {1 if (x1,x2,x3,x4,x5)=(0,0,0,0,0),

                   g1(x1,x2,x3,x4,x5) otherwise}

f2(x5,x6,x7,x8,x9) = {1 if (x5,x6,x7,x8,x9)=(0,0,0,0,0),

                   g2(x5,x6,x7,x8,x9) otherwise}

这是我的 tensorflow 代码

import tensorflow as tf
import numpy as np
ph = tf.placeholder(dtype=tf.float32, shape=[None, 9])
x1 = tf.slice(ph, [0, 0], [-1, 5])
x2 = tf.slice(ph, [0, 4], [-1, 5])
fixed1 = tf.placeholder(dtype=tf.float32, shape=[1, 5])
fixed2 = tf.placeholder(dtype=tf.float32, shape=[1, 5])

# MLP 1
w1 = tf.Variable(tf.ones([5, 1]))
g1 = tf.matmul(x1, w1)

# MLP 2
w2 = tf.Variable(-tf.ones([5, 1]))
g2 = tf.matmul(x2, w2)

check1 = tf.reduce_all(tf.equal(x1, fixed1), axis=1, keep_dims=True)

check2 = tf.reduce_all(tf.equal(x2, fixed2), axis=1, keep_dims=True)

#### with Problem
f1 = tf.cond(check1,
                   lambda: tf.constant([2], dtype=tf.float32), lambda: g1)
f2 = tf.cond(check2,
                   lambda: tf.constant([1], dtype=tf.float32), lambda: g2)
####
f = tf.add(f1, f2)

x = np.array([[0, 0, 0, 0, 0, 0, 0, 0, 0],
              [0, 0, 0, 0, 0, 0, 0, 0, 1],
              [1, 0, 0, 0, 0, 0, 0, 0, 0],
              [2, 0, 0, 0, 0, 0, 0, 0, 0],
              [9, 0, 0, 0, 0, 0, 0, 0, 0]])

fixed = np.array([[0, 0, 0, 0, 0]])

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    print('(1)\n', sess.run(check1, feed_dict={ph: x, fixed1: fixed, fixed2: fixed}))
    print('(2)\n', sess.run(check2, feed_dict={ph: x, fixed1: fixed, fixed2: fixed}))
    print('(3)\n', sess.run(f, feed_dict={ph: x, fixed1: fixed, fixed2: fixed}))
    print('(4)\n', sess.run(f1, feed_dict={ph: x, fixed1: fixed, fixed2: fixed}))
    print('(5)\n', sess.run(f2, feed_dict={ph: x, fixed1: fixed, fixed2: fixed}))

<小时/>

在这种情况下，

check1 是 [[ True], [ True], [False], [False], [False]]，形状为 (5, 1)

check2 是 [[ True], [False], [ True], [ True], [ True]]，形状为 (5, 1)

我期望 f 的结果是 [[3], [1], [2], [3], [10]]

但似乎 tf.cond() 无法将输入处理为形状为 (5, 1) 的 bool 张量

请您建议如何使用 tensorflow 实现 f(x)。

这是我收到的错误消息

Traceback (most recent call last): File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\common_shapes.py", line 670, in _call_cpp_shape_fn_impl status) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\contextlib.py", line 66, in exit next(self.gen) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\errors_impl.py", line 469, in raise_exception_on_not_ok_status pywrap_tensorflow.TF_GetCode(status)) tensorflow.python.framework.errors_impl.InvalidArgumentError: Shape must be rank 0 but is rank 2 for 'cond/Switch' (op: 'Switch') with input shapes: [?,1], [?,1].

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "C:/Users/hong/Dropbox/MLILAB/Research/GM-MLP/code/tensorflow_cond.py", line 23, in lambda: tf.constant([2], dtype=tf.float32), lambda: g1) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\ops\control_flow_ops.py", line 1765, in cond p_2, p_1 = switch(pred, pred) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\ops\control_flow_ops.py", line 318, in switch return gen_control_flow_ops._switch(data, pred, name=name) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\ops\gen_control_flow_ops.py", line 368, in _switch result = _op_def_lib.apply_op("Switch", data=data, pred=pred, name=name) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\op_def_library.py", line 759, in apply_op op_def=op_def) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 2242, in create_op set_shapes_for_outputs(ret) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 1617, in set_shapes_for_outputs shapes = shape_func(op) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 1568, in call_with_requiring return call_cpp_shape_fn(op, require_shape_fn=True) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\common_shapes.py", line 610, in call_cpp_shape_fn debug_python_shape_fn, require_shape_fn) File "C:\Users\hong\AppData\Local\Continuum\Anaconda3\lib\site-packages\tensorflow\python\framework\common_shapes.py", line 675, in _call_cpp_shape_fn_impl raise ValueError(err.message) ValueError: Shape must be rank 0 but is rank 2 for 'cond/Switch' (op: 'Switch') with input shapes: [?,1], [?,1].

Process finished with exit code 1

Answer 1

我认为你需要 tf.where，而不是 tf.cond。

查看此问题的答案:How to use tf.cond for batch processing