02. Project Q&A

ADL Project Walkthrough V2

Points of Clarification

  • At 7:36 the following code is presented:
    tf.save_model.loader.load(sess, [vgg_tag], vgg_tag) ,
    this line should read: tf.save_model.loader.load(sess, [vgg_tag], vgg_path)

  • At 3:43 in the video, it is stated that the pretrained VGG-16 is frozen. Actually, the model is not frozen, and all the weights are trainable

  • It is stated that the model’s output tensor needs to be reshaped into 2D since that is what TensorFlow’s softmax function requires. Actually, TensorFlow’s softmax function (and tf.nn.softmax_cross_entropy_with_logits) accepts tensors of any shape and will apply the softmax function on the last axis of the tensor. Consequently, there is no need to reshape the output of the model.

  • It is stated that we need to add 1x1 convolutions on top of the VGG-16 network in order to preserve the spatial information of the data. Actually, the pretrained VGG-16 model is already fully convolutional, containing all the convolutions that replace the fully connected layers and retain spatial information. We must add 1x1 convolutions on top of the VGG to reduce the number of filters from 4096 to the number of classes for our specific model.