CNN model parameters

This reference targets Aner 0.0.5 or newer. Updating documentation or editor help alone does not upgrade an installed runtime; check Downloads for verified platform packages.

This is the public Aner source reference for immutable CNNModel values in aner.nn. Start from Neural networks, the CNN workflow, or the model directory. The API inventory records every registered overload. The mathematical Tensor nn.conv2d primitive has a different signature from the model builder described here.

Model creation and architecture variables

import aner.nn

let model = nn.cnn(channels: 1, height: 8, width: 8, seed: 42)
    .conv2d(outputs: 4, kernel: 3, stride: 1, padding: 1)
    .relu().global_pool().flatten()
    .dropout(probability: 0.2).dense(outputs: 2)
print(model.summary())
print(model.parameter_count())

Construction declares shapes and does not allocate initialized weights. prepare, fitting, inference or parameter inspection initializes the relevant snapshot. Configure topology before initialization: adding layers to an initialized/trained snapshot is rejected. Builders return new values; they do not mutate an earlier model.

For one spatial dimension, convolution computes floor((input + 2*padding - dilation*(kernel-1) - 1)/stride) + 1. Infer both height and width this way; checked geometry and tensor/work budgets can reject a large combination. A group count equal to input channels admits depthwise style grouping only when output divisibility also holds. The architecture registry caps combined parameter/buffer storage at 150M elements; per operation and graph budgets can reject a smaller declared network.

Placement, compute policy and resource variables

performance_summary(), device(), dtype(), loss_scale(), attempts(), updates() and skipped_updates() disclose actual policy and state. Selecting a policy does not guarantee a speedup or any particular GPU instruction. Device performance and CNN execution details explain numerical and collective boundaries.

Every source fit variable

model.fit(images, targets, epochs=1, batch=32, rate=0.001,
          seed=42, optimizer="adam")

The = notation above summarizes implemented shorter overloads. The explicit execution overload adds a ParallelContext. Positional calls use the complete prefix; named calls can supply execution while the checker fills earlier optional fit defaults.

The native CNN fitting implementation uses global norm clipping 1, weight decay zero and shuffling enabled for these registered source overloads. Its native FitOptions structure has additional fields, but clip_norm, weight_decay and shuffle are not public CNN source .fit arguments. The recurrent model has a different, larger fit signature.

A run admits at most 1,000,000 attempted batch updates, including attempts whose FP16 overflow skips the update. Parameters, moments and successful update count do not advance on a skipped update. Check skipped_updates() rather than treating every attempt as optimization progress. CNN fitting accepts raw tensors and does not automatically enforce Dataset partition roles or fit a standardizer; retain split IDs and preprocessing yourself.

Results, weights and continuation

logits(images) returns detached [N,classes], probabilities(images) applies softmax, predict(images) returns [N,1] class IDs, and accuracy(images, targets) measures the caller supplied labels. These calls disable training dropout and use retained BatchNorm buffers. last_loss() is the recorded training observation; it is not a separately held out validation metric. Evaluate independently retained validation data explicitly.

parameter_names() and parameter(name) inspect exact named weights. load_weights(path) imports compatible local Safetensors values into the declared topology and starts the appropriate fresh training state; it does not download a model. read_images(path, samples) reads a raw little endian Float32 or Float64 NCHW payload in the model's selected dtype. path is required; positive samples requires exactly samples*C*H*W finite values. It does not decode PNG/JPEG files or infer normalization.

save(path) exclusively creates a family specific checkpoint containing topology, parameters, BatchNorm buffers, optimizer moments, successful/attempted/shuffle counters and performance/scaler policy. nn.load_cnn(path) restores CPU portable storage; select placement explicitly afterward. Use identical data/order, seeds and numerical implementation for exact continuation. This format is independent of Dense classifier v7.

Run CNN basics, architecture inspection, and CPU replica training to inspect complete native workflows. Parameter counts describe stored values; convolution's im2col workspace, saved activations, gradients and optimizer moments can dominate memory. Use smaller batches/spatial maps and inspect declared budgets before increasing them.