#feature-vector

Question 1: Select all that apply

A pretrained VGG-19 receives $224 \times 224$ RGB images. Its convolutional part ends with a max-pooling layer that outputs a $512 \times 7 \times 7$ map, which is flattened and fed to fc6 (4096 units), fc7 (4096 units) and a 1000-way output layer. The convolution conv4_3 sits in the fourth block, where the feature maps are $28 \times 28$ and have 512 channels. You want fixed feature vectors for a new classifier. Which statements are correct? Select all that apply.

The fc7 activations give a 4096-dimensional feature vector1
The 1000 class scores of the output layer can serve as a 1000-dimensional feature vector2
Flattening the last pooled map gives a 25,088-dimensional feature vector3
The conv4_3 output is a 512-dimensional feature vector as it stands4
The fc6 output is a 25,088-dimensional feature vector5

Timed test