spindle

Spindle neurons enable whales to be deep thinkers.

How to run

1. Download MNIST dataset and extract .gz files into ./data/ directory
2. Start and get inside the Docker container:

   cd infra-dev/
   docker-compose up -d
   docker-compose exec -it spindle bash

3. Training:

   zig build run -- ../data/

   Logs:

   Before training: loss=1.6351758242, acc=10.53%
   Epoch 0: elapsed_time=0:24, loss=0.0950836316, acc=89.35%
   Epoch 1: elapsed_time=0:47, loss=0.0602058768, acc=92.61%
   Epoch 2: elapsed_time=1:11, loss=0.0562265515, acc=93.25%
   Epoch 3: elapsed_time=1:35, loss=0.0533989370, acc=93.56%
   Epoch 4: elapsed_time=1:59, loss=0.0503539443, acc=94.18%
   Epoch 5: elapsed_time=2:23, loss=0.0506065302, acc=94.01%
   Epoch 6: elapsed_time=2:46, loss=0.0488770492, acc=94.37%
   Epoch 7: elapsed_time=3:10, loss=0.0465301722, acc=94.54%
   Epoch 8: elapsed_time=3:34, loss=0.0448890477, acc=94.69%
   Epoch 9: elapsed_time=3:58, loss=0.0450211503, acc=94.71%
   Epoch 10: elapsed_time=4:21, loss=0.0452360660, acc=94.56%
   Epoch 11: elapsed_time=4:45, loss=0.0468302444, acc=94.37%
   Epoch 12: elapsed_time=5:09, loss=0.0434075668, acc=94.85%
   Epoch 13: elapsed_time=5:33, loss=0.0441294163, acc=94.76%
   Epoch 14: elapsed_time=5:56, loss=0.0428987928, acc=94.98%
   

References

• Using neural nets to recognize handwritten digits
• DNNs from Scratch in Zig