ReceptiveFieldReport

ReceptiveFieldReport

Demonstrates TNNet.ReceptiveFieldReport: a purely analytical per-layer receptive-field (RF) walk of a TNNet. No probe batch, no forward pass, no backward pass — it reads each layer's kernel/stride/padding geometry and propagates the textbook RF recurrence.

What it does

1. Builds a plain VGG-style 3x3 stride-1 stack (32x32 input) and prints its per-layer RF table. Each 3x3 conv grows the RF by 2 and the jump stays at 1.
2. Builds a stride-2 downsampling stack (32x32 input) and prints the same table. Every stride-2 layer (and the 2x2 max-pool) doubles the jump, so the RF grows much faster and covers the whole input far sooner.

No training and no data: the report walks NN.Layers and uses each layer's recorded kernel size, stride and padding.

The recurrence

Tracked independently for X and Y (so rectangular kernels/strides work):

r_k     = r_{k-1} + (kernel_k - 1) * jump_{k-1}     // RF size on the input
jump_k  = jump_{k-1} * stride_k                      // effective stride
start_k = start_{k-1} + ((kernel_k - 1)/2 - pad_k) * jump_{k-1}

• Convolution / Deconvolution read FeatureSizeX/Y, Stride, Padding. A deconvolution that grows the map is treated as a fractional stride.
• Pool family (TNNetMaxPool, TNNetAvgPool, TNNetMinPool, ...) read PoolSize, Stride, Padding.
• Global channel reductions (TNNetMaxChannel, TNNetAvgChannel, TNNetGlobalSumPool) span the whole current feature map → RF jumps to the full input.
• Upsample / DeMaxPool / DeAvgPool divide the jump (fractional stride).
• Pad layers shift the start offset.
• Everything else (FullConnect, activations, norms, ...) is treated as kernel=1, stride=1, padding=0 (pass-through).

Output

For each network the report prints one row per layer plus a summary:

Idx   Class                          OutShape       k/s/p(x,y)        RF(x,y)  jump(x,y)     cover%
--------------------------------------------------------------------------------------------------------
0     TNNetInput                     (32,32,3)      1/1/0,1/1/0          1x1        1x1       0.1%
1     TNNetConvolutionReLU           (32,32,16)     3/1/1,3/1/1          3x3        1x1       0.9%
...
--------------------------------------------------------------------------------------------------------
Final receptive field: ... on a 32 x 32 input.
Final jump (effective stride): ...
First layer whose RF covers the whole input: ... — the rest is global mixing.
Flags:
  ...

cover% is the fraction of the input plane a single output unit of that layer can see; it is flagged >100% once the RF exceeds the input (the "already global" point). The flag list also calls out any layer whose RF stopped growing (the pointwise / 1x1 tail).

Running

cd examples/ReceptiveFieldReport
lazbuild ReceptiveFieldReport.lpi
../../bin/<arch>/bin/ReceptiveFieldReport

Or directly with fpc:

cd examples/ReceptiveFieldReport
fpc -B -Fu../../neural -Mobjfpc -Sh -O2 ReceptiveFieldReport.lpr
./ReceptiveFieldReport

Runs in under a second on CPU.