转置卷积（反卷积）算法

Question

我正在使用 tensorflow 构建卷积神经网络。给定形状的张量 (none, 16, 16, 4, 192)，我想执行转置卷积，得到形状 (none, 32, 32, 7, 192)。

[2,2,4,192,192] 的过滤器大小和 [2,2,1,1,1] 的步幅是否会产生我想要的输出形状？

Answer 1

是的，你几乎是对的。

一个小的修正是tf.nn.conv3d_transpose 期望NCDHW 或NDHWC 输入格式（你的似乎是NHWDC），过滤器形状预计是[depth, height, width, output_channels, in_channels]。这会影响filter 和stride 中的维度顺序：

# Original format: NHWDC.
original = tf.placeholder(dtype=tf.float32, shape=[None, 16, 16, 4, 192])
print original.shape

# Convert to NDHWC format.
input = tf.reshape(original, shape=[-1, 4, 16, 16, 192])
print input.shape

# input shape:  [batch, depth, height, width, in_channels].
# filter shape: [depth, height, width, output_channels, in_channels].
# output shape: [batch, depth, height, width, output_channels].
filter = tf.get_variable('filter', shape=[4, 2, 2, 192, 192], dtype=tf.float32)
conv = tf.nn.conv3d_transpose(input,
                              filter=filter,
                              output_shape=[-1, 7, 32, 32, 192],
                              strides=[1, 1, 2, 2, 1],
                              padding='SAME')
print conv.shape

final = tf.reshape(conv, shape=[-1, 32, 32, 7, 192])
print final.shape

哪些输出：

(?, 16, 16, 4, 192)
(?, 4, 16, 16, 192)
(?, 7, 32, 32, 192)
(?, 32, 32, 7, 192)