times.txt

Benchmark Results
-----------------

Train/Test Set:

- mnist 64 samples per class
- 28x28 - cropped to 24x24, downsized to 8x8 and thresholded to 1 bit per color

Network:

- Convolutional layer with kernels size of 3x3 + bais
  Conversion: 1x8x8 -> Nx6x6 where N is number of kernels

- Max pooling 2x2 + ReLU : Nx6x6 -> Nx3x3
- Fully Connected with bias: Nx3x3->M classes 
- Euclidean loss layer


Two Training variants:

2 dig.:
  - two digits only 0,1
  - 4 kernels
  - 2 epochs
  - parameters: Conv 4*3*3 + 4 offset, FC 2x36 + 2 offset, total: 114 prams

10 dig.:
  - all 10 digits
  - 12 kernels
  - 5 epochs
  - parameters: Conv 12*3*3 + 12 offset, FC 10x108 + 10 offset, total: 1210 prams

Implementations:

- BASIC - standard ZX Spectrum basic  (train.bas)
- ToBoSFP - same code compiled with ToBoSFP (train.bas)
- C/float - floating point implementation using z88dk (train.c)
- C/fixed - fixed point (4:12) computation with custom multiplication
            written in assembly, using z88dk (train_f16.c)


Results:
--------
                BASIC   ToBoSFP C/float C/fixed
                -       ToBoSFP z88dk   z88dk+asm
2 dig. Train:    5h20m  12m     3.7m    1.5m   
2 dig. Test:     2h19    6m     1.2m    0.5m
Total:          12h59m  30m     8.6m    3.5m

                BASIC   ToBoSFP C/float C/fixed
                -       ToBoSFP z88dk   z88dk+asm
10 dig Train:   N/A      3h15m   2h16m    41.4m
10 dig Test:    N/A      1h21m     41m    13.6m
Total:                  17h36m  12h01m  3h40m