Question

我正在尝试将this functionality从Python移植到.NET Core上的C＃，以查看是否可以从我知道的静态类型语言中使用Tensorflow。为此，我使用this framework，它包装了Tensorflow C ++ API。该API不包含函数conv2d_transpose，因此我正在寻找一种用C ++ API中实际存在的函数替换它的方法。

我尝试挖掘Tensorflow的源代码，以查看是否可以找到conv2d_transpose的实现，并找到this code并将其引向this code，在我看来，这类似于以下内容是实际的肉：

with ops.name_scope(name, "conv2d_transpose", [input, filter, output_shape]) as name:
  if data_format is None:
    data_format = "NHWC"
  channel_index = 1 if data_format.startswith("NC") else 3

  strides = _get_sequence(strides, 2, channel_index, "strides")
  dilations = _get_sequence(dilations, 2, channel_index, "dilations")

  return gen_nn_ops.conv2d_backprop_input(
      input_sizes=output_shape,
      filter=filters,
      out_backprop=input,
      strides=strides,
      padding=padding,
      data_format=data_format,
      dilations=dilations,
      name=name)

我翻译了以下Python代码：

def _conv_tranpose_layer(net, num_filters, filter_size, strides):
    weights_init = _conv_init_vars(net, num_filters, filter_size, transpose=True)

    batch_size, rows, cols, in_channels = [i.value for i in net.get_shape()]
    new_rows, new_cols = int(rows * strides), int(cols * strides)

    new_shape = [batch_size, new_rows, new_cols, num_filters]
    tf_shape = tf.stack(new_shape)
    strides_shape = [1,strides,strides,1]

    net = tf.nn.conv2d_transpose(net, weights_init, tf_shape, strides_shape, padding='SAME')
    net = _instance_norm(net)
    return tf.nn.relu(net)

到以下C＃代码：

private Tensor _conv_transpose_layer(Tensor net, int num_filters, int filter_size, int strides)
{
    RefVariable weights_init = _conv_init_vars(net, num_filters, filter_size, transpose: true);
    int batch_size = net.shape[0];
    int rows = net.shape[1];
    int cols = net.shape[2];
    int in_channels = net.shape[3];
    int new_rows = rows * strides;
    int new_cols = cols * strides;
    int[] new_shape = new int[] { batch_size, new_rows, new_cols, num_filters };
    Tensor tf_shape = tf.stack(new_shape);
    int[] strides_shape = new int[] { 1, strides, strides, 1 };
    //net = tf.nn.conv2d_transpose(net, weights_init, tf_shape, strides_shape, "SAME");
    net = Tensorflow.Operations.gen_nn_ops.conv2d_backprop_input(new Tensorflow.Operations.Conv2dParams()
    {
        InputSizes = tf_shape,
        Filter = weights_init,
        OutBackProp = net,
        Strides = strides_shape,
        DataFormat = "NHWC",
        Dilations = null,
        Name = null
    });
    net = _instance_norm(net);
    return tf.nn.relu(net);
}

我本可以使用this post的答案来解决我的问题，但是我不知道如何将这些答案集成到我正在构建的图形中。特别是因为它们包含循环或Session.run

的单独调用

我对ML知之甚少，无法真正验证此代码的结果，并且我没有足够的计算资源来尝试我能想到的一切。

请告诉我我的翻译是否有意义，以及在哪里可以找到有关这些操作为何/如何发生联系的更多信息。

谢谢。

如何在Tensorflow中用conv2d替换conv2d_transpose？

0 个答案: