Tensorflow：tf.argmax 和切片

Question

我想设计这个损失函数：

sum((y[argmax(y_)] - y_[argmax(y_)])²)

我找不到方法 y[argmax(y_)]。我尝试了 y[k]、y[:,k] 和 y[None,k] none 这些工作。这是我的代码：

    Na = 3
    x = tf.placeholder(tf.float32, [None, 2])
    W = tf.Variable(tf.zeros([2, Na]))
    b = tf.Variable(tf.zeros([Na]))
    y = tf.nn.relu(tf.matmul(x, W) + b)
    y_ = tf.placeholder(tf.float32, [None, 3])
    k = tf.argmax(y_, 1)
    diff = y[k] - y_[k]
    loss = tf.reduce_sum(tf.square(diff))

错误：

  File "/home/ncarrara/phd/code/cython/robotnavigation/ftq/cftq19.py", line 156, in <module>
    diff = y[k] - y_[k]
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/ops/array_ops.py", line 499, in _SliceHelper
    name=name)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/ops/array_ops.py", line 663, in strided_slice
    shrink_axis_mask=shrink_axis_mask)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/ops/gen_array_ops.py", line 3515, in strided_slice
    shrink_axis_mask=shrink_axis_mask, name=name)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 767, in apply_op
    op_def=op_def)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2508, in create_op
    set_shapes_for_outputs(ret)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1873, in set_shapes_for_outputs
    shapes = shape_func(op)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1823, in call_with_requiring
    return call_cpp_shape_fn(op, require_shape_fn=True)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/framework/common_shapes.py", line 610, in call_cpp_shape_fn
    debug_python_shape_fn, require_shape_fn)
  File "/home/ncarrara/miniconda3/lib/python2.7/site-packages/tensorflow/python/framework/common_shapes.py", line 676, in _call_cpp_shape_fn_impl
    raise ValueError(err.message)
ValueError: Shape must be rank 1 but is rank 2 for 'strided_slice' (op: 'StridedSlice') with input shapes: [?,3], [1,?], [1,?], [1].

Answer 1

可以使用 tf.gather_nd:

import tensorflow as tf

Na = 3
x = tf.placeholder(tf.float32, [None, 2])
W = tf.Variable(tf.zeros([2, Na]))
b = tf.Variable(tf.zeros([Na]))
y = tf.nn.relu(tf.matmul(x, W) + b)
y_ = tf.placeholder(tf.float32, [None, 3])
k = tf.argmax(y_, 1)
# Make index tensor with row and column indices
num_examples = tf.cast(tf.shape(x)[0], dtype=k.dtype)
idx = tf.stack([tf.range(num_examples), k], axis=-1)
diff = tf.gather_nd(y, idx) - tf.gather_nd(y_, idx)
loss = tf.reduce_sum(tf.square(diff))

解释：

在这种情况下，tf.gather_nd 的想法是制作一个矩阵（二维张量），其中每一行包含要在输出中包含的行和列的索引。例如，如果我有一个矩阵 a 包含：

| 1 2 3 |
| 4 5 6 |
| 7 8 9 |

和一个矩阵 i 包含：

| 1 2 |
| 0 1 |
| 2 2 |
| 1 0 |

那么 tf.gather_nd(a, i) 的结果将是向量（一维张量）包含：

| 6 |
| 2 |
| 9 |
| 4 |

在这种情况下，列索引由 k 中的 tf.argmax 给出；它会告诉您，对于每一行，哪一列是具有最高值的。现在您只需要将行索引与其中的每一个放在一起。 k 中的第一个元素是第 0 行中最大值列的索引，下一个元素是第 1 行的索引，依此类推。 num_examples 只是 x 中的行数，然后 tf.range(num_examples) 为您提供一个从 0 到 x 中的行数减去 1 的向量（即所有序列行索引）。现在你只需要将它与 k 放在一起，这就是 tf.stack 所做的，结果 idx 是 tf.gather_nd.

的参数

Tensorflow：tf.argmax 和切片

Tensorflow : tf.argmax and slicing

python

tensorflow

slice

argmax