What the Second Block Inherits

Two Inception blocks are stacked, one feeding the other. Every convolution uses "same" padding, so every path in a block returns the same grid size, and the paths are joined by concatenating along the depth axis.

Block A receives a 192-deep input:

path	layers	filters in the final conv
$1 \times 1$	one $1 \times 1$ conv	64
$3 \times 3$	$1 \times 1$ conv with 96 filters, then $3 \times 3$ conv	128
$5 \times 5$	$1 \times 1$ conv with 16 filters, then $5 \times 5$ conv	32
pool	$3 \times 3$ max-pool (stride 1), then $1 \times 1$ conv	32

Block B takes block A's output as its input, and is built the same way:

path	layers	filters in the final conv
$1 \times 1$	one $1 \times 1$ conv	128
$3 \times 3$	$1 \times 1$ conv with 128 filters, then $3 \times 3$ conv	192
$5 \times 5$	$1 \times 1$ conv with 32 filters, then $5 \times 5$ conv	96
pool	$3 \times 3$ max-pool (stride 1), then $1 \times 1$ conv	64

How many weights does block B hold?

Count convolution weights only — ignore biases, and remember pooling has none.

Give your answer in thousands, rounded to 1 decimal place (so 12,300 weights would be entered as 12.3).

What the Second Block Inherits

Hints