Where should I apply dropout to a convolutional layer?

Question

Because the word "layer" often means different things when applied to a convolutional layer (some treat everything up through pooling as a single layer, others treat convolution, nonlinearity, and pooling as separate "layers"; see fig 9.7) it's not clear to me where to apply dropout in a convolutional layer.

Does dropout happen between nonlinearity and pooling?

---

E.g., in TensorFlow would it be something like:

kernel_logits = tf.nn.conv2d(input_tensor, ...) + biases
activations = tf.nn.relu(kernel_logits)
kept_activations = tf.nn.dropout(activations, keep_prob)
output = pool_fn(kept_activations, ...)

Answer 1

You could probably try applying dropout at different places, but in terms of preventing overfitting not sure you're going to see much of a problem before pooling. What I've seen for CNN is that tensorflow.nn.dropout gets applied AFTER non-linearity and pooling:

    # Create a convolution + maxpool layer for each filter size
    pooled_outputs = []
    for i, filter_size in enumerate(filters):
        with tf.name_scope("conv-maxpool-%s" % filter_size):
            # Convolution Layer
            filter_shape = [filter_size, embedding_size, 1, num_filters]
            W = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1), name="W")
            b = tf.Variable(tf.constant(0.1, shape=[num_filters]), name="b")
            conv = tf.nn.conv2d(
                self.embedded_chars_expanded,
                W,
                strides=[1, 1, 1, 1],
                padding="VALID",
                name="conv")
            # Apply nonlinearity
            h = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu")
            # Maxpooling over the outputs
            pooled = tf.nn.max_pool(
                h,
                ksize=[1, sequence_length - filter_size + 1, 1, 1],
                strides=[1, 1, 1, 1],
                padding='VALID',
                name="pool")
            pooled_outputs.append(pooled)



    # Combine all the pooled features
    num_filters_total = num_filters * len(filters)
    self.h_pool = tf.concat(3, pooled_outputs)
    self.h_pool_flat = tf.reshape(self.h_pool, [-1, num_filters_total])

    # Add dropout
    with tf.name_scope("dropout"):
        self.h_drop = tf.nn.dropout(self.h_pool_flat, self.dropout_keep_prob)