A ResNet stage operates on a block. One bottleneck block is three convolutions in a row, all with "same" padding so the grid never changes, with a skip connection carrying the input around them and adding it back at the end:
| layer | kernel | filters |
|---|---|---|
| 1 | 64 | |
| 2 | 64 | |
| 3 | 256 |
How many weights does this block hold?
Count convolution weights only — ignore biases, and recall that a filter always spans the full depth of the input it is applied to.
Give your answer in thousands, rounded to 1 decimal place (so 12,300 weights would be entered as 12.3).