Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/mrjake222/fogml-riscv
https://github.com/mrjake222/fogml-riscv
Last synced: 9 days ago
JSON representation
- Host: GitHub
- URL: https://github.com/mrjake222/fogml-riscv
- Owner: MrJake222
- Created: 2024-10-17T12:38:44.000Z (28 days ago)
- Default Branch: master
- Last Pushed: 2024-10-27T12:13:49.000Z (18 days ago)
- Last Synced: 2024-10-27T13:47:10.561Z (18 days ago)
- Language: C
- Size: 161 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# FogML-RiscV
Based on [FogML-Arduino](https://github.com/tszydlo/FogML-Arduino).
### Compilation
This project uses CMake, so: `mkdir build`, `cd build`, `cmake ..`, `make`
## Usage
For now, you'll need dummy data from [FogML-Linux](https://github.com/MrJake222/FogML-Linux/).
- Generate data: `gendata/sim.py > data-sim-x24` (FogML-Linux),
- Redirect & send data: `./send.py < data-sim-x24`,
- Reset the core (press the button),
- Observe the model learn for 16, then ok for 8, and fail for 8 iterations.
### Send protocol
It sends raw float data in little-endian binary format. For each sample (3 axis x 4-byte float)
12 bytes is needed. The code echoes read data in human-readable format for verification.
## Results
In my [Risc-V survey](https://github.com/MrJake222/riscv-ice40) core VexRiscV (smprod_my, RV32I) @ 16MHz, -O2
learning took from 24s (1st pass) to 31.4s (24th pass) and 0.6s for inference pass:
```
fogml vector size: 12
```
```
processing...
learning...
DSP took 0.06 sec
RES took 0.00 sec
LOF took 23.98 sec (99.8%)
finished learning
[P 01] lof took: 24.04 sec
```
```
processing...
classifying...
DSP took 0.06 sec
LOF took 0.50 sec (89.3%)
LOF Score = 0.98, ok
[P 17] lof took: 0.56 sec
```
**These results are with FFT turned off.**
Most learning time is spent in `tinyml_lof_learn()` -- 99.8%.
Out of it 98% is calculating k-nearest neighbors O(n^2) with euclidean distance.
Modifications (learning time 1st pass, applied separately):
- 24.0s -O3,
- 24.2s -Os,
- 23.4s no sqrt on distance,
- 14.3s taxicab distance (learning 0.22s),
- 8.5s with dummy +1 in loop,
- 0.4s no loop.
Last 4 modification applied to `tinyml_lof_normal_distance_vec()` in `anomaly_rt/fogml_lof.c`.
### With FFT
This is -Os since defaults changes (minor speed decrease).
The DSP time naturally increased (added FFT block), but also the LOF time,
as the feature vector got longer.
Code & results on `fft` branch.
```
fogml vector size: 18
```
```
processing...
learning...
DSP took 0.38 sec
RES took 0.00 sec
LOF took 35.48 sec (98.9%)
finished learning
[P 01] lof took: 35.87 sec
```
```
processing...
classifying...
DSP took 0.24 sec
LOF took 0.71 sec (74.7%)
LOF Score = 0.98, ok
[P 17] lof took: 0.95 sec
```
Percent of LOF time in learning has barely reduced (~1%p). Inference
shifted more (~14.5%p)
### Random forests
Random forest takes 0.07s to complete (including 0.06s for DSP).
Only -Os tested (-O2 random forest won't fit):
```
RF DSP took 0.06 sec
RF cls took 0.00 sec
RF class = 2
[P 01] rf took: 0.07 sec
```
## Other files
- `size.ods` -- size analysis of this code
- `tiny.pdf` -- tutorial how to make small binaries