#flops

Question 1: Single choice

d2l's AlexNet takes a $1 \times 224 \times 224$ input and has five convolutional layers followed by fully connected layers of sizes $6400 \to 4096 \to 4096 \to 10$ . It has about 46.8 million parameters and needs about 0.94 billion multiply–adds per image. How are these shared between the convolutional and the fully connected layers?

The FC layers hold about 92% of the parameters; the conv layers do about 95% of the multiply–adds1
The conv layers hold about 92% of the parameters; the FC layers do about 95% of the multiply–adds2
The FC layers dominate both, with about 92% of the parameters and 95% of the multiply–adds3
The conv layers dominate both, with about 92% of the parameters and 95% of the multiply–adds4

Timed test