Placement of dropout layers

[aalapshah12297]

From what I understand, droupout can only benefit the layers after it, since they must predict the same output with a subset of the input. In your implementation, dropout appears after the final conv layer, and is only succeeded by an upsampling and softmax layer (which do not have parameters to learn). It seems more logical to place the dropout layer(s) between the Classification layers i.e at line 110 and 114.

Fast-SCNN/tf_2_0_fast_scnn.py

Lines 105 to 121 in fcb381d

	"""## Step 4: Classifier"""
	classifier = tf.keras.layers.SeparableConv2D(128, (3, 3), padding='same', strides = (1, 1), name = 'DSConv1_classifier')(ff_final)
	classifier = tf.keras.layers.BatchNormalization()(classifier)
	classifier = tf.keras.activations.relu(classifier)
	classifier = tf.keras.layers.SeparableConv2D(128, (3, 3), padding='same', strides = (1, 1), name = 'DSConv2_classifier')(classifier)
	classifier = tf.keras.layers.BatchNormalization()(classifier)
	classifier = tf.keras.activations.relu(classifier)
	classifier = conv_block(classifier, 'conv', 19, (1, 1), strides=(1, 1), padding='same', relu=True)
	classifier = tf.keras.layers.Dropout(0.3)(classifier)
	classifier = tf.keras.layers.UpSampling2D((8, 8))(classifier)
	classifier = tf.keras.activations.softmax(classifier)