EricLouchez.Shorokoo.Modules 0.1.8-dev

Shorokoo.Modules

Baseline neural-network library for Shorokoo: ready-made layers, loss functions, optimizers, and initializers built from Shorokoo modules.

• Initializers (Shorokoo.Modules.Initializers) — Zeros, Ones, Uniform, Normal, XavierUniform, XavierNormal, KaimingUniform, KaimingNormal, TruncatedNormal, LeCunNormal. All shape-only [TrainableParamInitializer]s; the random ones are seeded (deterministic), and Xavier/Kaiming/LeCun compute fan-in/fan-out in-graph from the shape vector.
• Layers (Shorokoo.Modules.Layers) — Linear, Conv1d, Conv2d, Conv3d (hyperparameter-driven geometry via the dynamic Conv lowering), ConvTranspose2d (default geometry, kernel inferred from the weight), BatchNorm2d/BatchNorm1d (training/eval flag, running stats via StateUpdate), LayerNorm, RMSNorm, GroupNorm, InstanceNorm2d, Dropout (training flag), Embedding, MultiHeadAttention / TransformerEncoderLayer (+ the Attention.ScaledDotProductAttention helper), LeakyReLU/ELU (hyper alpha), PReLU (learnable slope), and the Pooling / GatedLinear.GLU helpers (MaxPool2d, AvgPool2d, GlobalAvgPool2d, GlobalMaxPool2d, Flatten). Plain activations are tensor one-liners — x.Relu(), x.Gelu(), x.Sigmoid(), x.Tanh(), x.Softmax(axis) — and need no modules.
• Losses (Shorokoo.Modules.Losses) — L2Loss (MSE), L1Loss, HuberLoss(delta) / SmoothL1Loss, CrossEntropyLoss (logits + int64 class indices), NLLLoss, BCELoss, BCEWithLogitsLoss, KLDivLoss (log-probs + probs). All map (predictions, targets) → scalar loss.
• Optimizers (Shorokoo.Modules.Optimizers) — SGDOptimizer, SGDMomentumOptimizer, AdamOptimizer (with bias correction), AdamWOptimizer, RMSpropOptimizer, AdagradOptimizer, with strongly typed hyperparameter sets and learning-rate schedules (Schedules.*). Optimizer state (moments, velocity, accumulators) is created inside each module via optimizer-owned [StateInitializer]s — OptimizerStateZeros (param-shaped) and OptimizerScalarZeros (a rank-0 scalar, e.g. Adam's timestep) — and threaded with StateUpdate; never declared in the Inline signature.

dotnet add package Shorokoo.Modules

using Shorokoo.Modules.Optimizers;
using Shorokoo.Modules.Losses;

var rig = TrainingRig.FromScratch(
    MyModel.ComputationGraph,
    CrossEntropyLoss.ComputationGraph,
    AdamOptimizer.ComputationGraph,
    sampleInputs,
    new AdamOptimizerHyperparameters { LearningRate = 1e-3f });

Documentation: https://github.com/Eric-Louchez/Shorokoo