Implement Convolution, Dropout and Pooling layers - Githubissues

fff-rs / juice

The Hacker's Machine Learning Engine

1.1k stars 77 forks source link

Implement Convolution, Dropout and Pooling layers #180

Closed hweom closed 1 year ago

hweom commented 1 year ago

What does this PR accomplish?

Implements Convolution, Dropout and Pooling layers in the new architecture.

🦚 Feature
🧭 Architecture

Changes proposed by this PR:

Additional by-product changes:

Passes the backend to layer/net init functions as it's sometimes needed to create backend-depended helper structures.
Adjusts the layer testing helpers to better handle different input/output dimensionality.
Adjusts the layer testing helpers to additionally take backend ref to support layers that are currently only implemented for CUDA.
Automatically extend the CUDA descriptor size to a minimum of 3 instead of returning an error.

📜 Checklist

[x] Test coverage is excellent
[x] All unit tests pass
[x] The juice-examples run just fine
[x] Documentation is thorough, extensive and explicit