// DOCS NAVIGATION
std.ml — Machine Learning_
SOURCE: content/docs/reference/05-stdlib.md — 05 Stdlib / 7. std.ml — Machine Learning
PT-BR:
std.ml fornece a camada de alto nível da Phase 6 para treinamento em CPU usando handles de std.tensor.
EN-US:
std.ml provides the Phase 6 high-level CPU training layer on top of std.tensor handles.
import std.tensor as tensor
import std.ml as ml
| Função / Function | Descrição / Description |
|---|---|
module_new() | Creates a module handle |
module_add_parameter, module_parameter_count, module_parameter | Parameter registration and discovery |
module_set_training, module_is_training | Training/eval mode |
linear(input, weight, bias) | Differentiable dense layer |
conv2d(input, kernel, bias, batch, in_ch, h, w, out_ch, kh, kw) | Differentiable valid 2D convolution over flattened NCHW tensors |
dropout(input, p, training) | Deterministic baseline dropout/inference helper |
max_pool2d(input, batch, channels, h, w, pool_h, pool_w) | Max pooling over flattened NCHW tensors |
mse_loss, bce_loss, cross_entropy_loss, nll_loss | Scalar tensor losses compatible with tensor.backward |
sgd_step, sgd_momentum_step, adam_step, adamw_step | Optimizers that update tensor parameters in place |
exp_lr(base, gamma, step) | Exponential learning-rate scheduling |
dataset_from_tensors, dataset_len | Tensor-backed datasets |
dataset_from_csv, dataset_from_jsonl, dataset_from_npy, dataset_from_directory | File-backed numerical datasets |
dataset_map_features, dataset_filter_label_min | Materialized dataset transforms |
dataset_train_split, dataset_test_split | Deterministic train/test dataset splits |
dataloader_new, dataloader_batch_count, dataloader_batch_features, dataloader_batch_labels | Deterministic minibatching |
dataframe_from_csv, dataframe_rows, dataframe_cols, dataframe_column | Numeric dataframe handles and column extraction |
experiment_start, experiment_finish | Tracked experiment lifecycle |
experiment_set_config, experiment_log_metric, experiment_log_artifact | Experiment config, metrics, and artifacts |
experiment_set_lockfile, experiment_set_model_output | Reproducibility lockfile and model output records |
experiment_manifest_path, experiment_repro_command, experiment_compare_manifests | Manifest path, reproduction command, and manifest comparison |
distributed_session_start, distributed_worker_step, distributed_global_step | Single-machine simulated distributed training coordination |
distributed_checkpoint_save, distributed_resume, distributed_summary, distributed_worker_step_count | Checkpoint/resume and worker progress inspection |
onnx_export, onnx_import_summary, onnx_validate, onnx_roundtrip | Binary ONNX subset export/import/round-trip for supported AI model blocks |
embedding_lookup, positional_encoding, layer_norm, gelu, swiglu, attention | Transformer tensor primitives |
kv_cache_new, kv_cache_append, kv_cache_keys, kv_cache_values, kv_cache_len, logits_sample | LLM KV-cache and logits sampling helpers |
tokenizer_wordpiece, tokenizer_encode, tokenizer_decode, text_embed | Deterministic tokenization and text embedding utilities |
vector_index_new, vector_index_insert, vector_index_query, vector_index_persist, vector_index_load, vector_index_set_metadata, vector_index_metrics | Deterministic HNSW vector index APIs backed by the R-3003 Artifact Container v1; legacy JSON is rejected |
rag_chunk_text, rag_build_prompt, rag_evaluate_answer | RAG chunking, prompt assembly, and evaluation |
Exemplos completos estão em:
tests/validation/72_ml_phase6_mlp_training.spectratests/validation/73_ml_phase6_cnn_training.spectra
Estado Phase 6: MLP e CNN pequenos treinam end-to-end nos testes de runtime, com exemplos Spectra compilando e executando pela API pública. Readers CSV/imagem/JSONL, serialização de modelos e prefetch paralelo são trabalho futuro.