DropBlock2D 层
DropBlock2D 类
keras_cv.layers.DropBlock2D(rate, block_size, seed=None, **kwargs)
对输入特征应用 DropBlock 正则化。
DropBlock 是一种结构化 dropout，其中特征图中相邻区域的单元被一起丢弃。DropBlock 比 dropout 在卷积层上的效果更好，因为卷积层的激活单元在空间上是相关的。
建议在网络的后续层中，在 Conv -> BatchNorm -> Activation 块之后使用 DropBlock。例如，论文中提到在 ResNet 块的第三和第四组中使用 DropBlock。
参考
参数
- rate: 浮点数。丢弃单元的概率。必须介于 0 和 1 之间。为了获得最佳结果，值应介于 0.05-0.25 之间。
- block_size: 整数，或整数元组。要丢弃的块的大小。如果是整数，则将丢弃一个正方形块。如果是元组，则数字表示块的（高度，宽度）。必须大于 0，并且不应该大于输入特征图大小。论文作者对大小为
14x14xchannels的输入特征使用block_size=7。如果此值大于或等于输入特征图大小，您将遇到nan值。 - seed: 整数。用作随机种子。
- name: 字符串。层的名称。
示例
DropBlock2D 可以用在 keras.Model 中
# (...)
x = Conv2D(32, (1, 1))(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = DropBlock2D(0.1, block_size=7)(x)
# (...)
直接使用时，该层将在相邻区域中将一些输入归零，并对剩余的值进行归一化。
# Small feature map shape for demonstration purposes:
features = tf.random.stateless_uniform((1, 4, 4, 1), seed=[0, 1])
# Preview the feature map
print(features[..., 0])
# tf.Tensor(
# [[[0.08216608 0.40928006 0.39318466 0.3162533 ]
# [0.34717774 0.73199546 0.56369007 0.9769211 ]
# [0.55243933 0.13101244 0.2941643 0.5130266 ]
# [0.38977218 0.80855536 0.6040567 0.10502195]]], shape=(1, 4, 4),
# dtype=float32)
layer = DropBlock2D(0.1, block_size=2, seed=1234) # Small size for
demonstration
output = layer(features, training=True)
# Preview the feature map after dropblock:
print(output[..., 0])
# tf.Tensor(
# [[[0.10955477 0.54570675 0.5242462 0.42167106]
# [0.46290365 0.97599393 0. 0. ]
# [0.7365858 0.17468326 0. 0. ]
# [0.51969624 1.0780739 0.80540895 0.14002927]]],
# shape=(1, 4, 4),
# dtype=float32)
# We can observe two things:
# 1. A 2x2 block has been dropped
# 2. The inputs have been slightly scaled to account for missing values.
# The number of blocks dropped can vary, between the channels - sometimes no
# blocks will be dropped, and sometimes there will be multiple overlapping
# blocks. Let's present on a larger feature map:
features = tf.random.stateless_uniform((1, 4, 4, 36), seed=[0, 1])
layer = DropBlock2D(0.1, (2, 2), seed=123)
output = layer(features, training=True)
print(output[..., 0]) # no drop
# tf.Tensor(
# [[[0.09136613 0.98085546 0.15265216 0.19690938]
# [0.48835075 0.52433217 0.1661478 0.7067729 ]
# [0.07383626 0.9938906 0.14309917 0.06882786]
# [0.43242374 0.04158871 0.24213943 0.1903095 ]]],
# shape=(1, 4, 4),
# dtype=float32)
print(output[..., 9]) # drop single block
# tf.Tensor(
# [[[0.14568178 0.01571623 0.9082305 1.0545396 ]
# [0.24126057 0.86874676 0. 0. ]
# [0.44101703 0.29805306 0. 0. ]
# [0.56835717 0.04925899 0.6745584 0.20550345]]],
# shape=(1, 4, 4),
# dtype=float32)
print(output[..., 22]) # drop two blocks
# tf.Tensor(
# [[[0.69479376 0.49463132 1.0627024 0.58349967]
# [0. 0. 0.36143216 0.58699244]
# [0. 0. 0. 0. ]
# [0.0315055 1.0117861 0. 0. ]]],
# shape=(1, 4, 4),
# dtype=float32)
print(output[..., 29]) # drop two blocks with overlap
# tf.Tensor(
# [[[0.2137237 0.9120104 0.9963533 0.33937347]
# [0.21868704 0.44030213 0.5068906 0.20034194]
# [0. 0. 0. 0.5915383 ]
# [0. 0. 0. 0.9526224 ]]],
# shape=(1, 4, 4),
# dtype=float32)