A small image classifier is fed a batch of colour photographs. The batch arrives as a 4-D tensor whose axes are (batch, height, width, channels):
input shape=(32, 10, 10, 3)
The network then runs the pipeline below.
- The flatten step leaves the batch axis untouched and merges every remaining axis into a single axis of length d.
- Each dense layer computes Z=AW+1bT, where W has shape (inputs, outputs) and the bias vector b holds one bias per output unit.
| Stage | Output shape |
|---|
| Input batch | (32, 10, 10, 3) |
| Flatten | (32, d) |
| Dense 1 (with bias) | (32, 20) |
| Dense 2 (with bias) | (32, 20) |
| Dense 3 (with bias) | (32, 4) |
How many learnable numbers does this network hold in total — every weight and every bias, across all three dense layers?
Your answer is an exact whole number.