noise_shape in dropout layers

[YUHAN666]

Hi, Thanks for your fantastic work. May I ask why do you set the noise_shape in the tf.nn.dropout to (None, 1, 1, 1)? https://github.com/qubvel/efficientnet/blob/6ceca713c1186168f41a7a628e4d6e2fdc4aa7ef/efficientnet/model.py#L240-L242 Does the noise_shape=(None, 1, 1, 1) means that some data in the batch is randomly abandoned at drop_rate during the training phase?