morel
Docs/Reference — config

Reference — config

Every field in the Config dataclass tree, with its meaning and the kind of validation it receives at construction time.

Top-level

Field Type Default Meaning
seed int 42 Global seed for torch, numpy, Python random, PYTHONHASHSEED, cuDNN.
device str "auto" "cpu", "cuda", or "auto".

data

Field Type Default Meaning
raw str "data/raw" Path to raw dataset directory.
processed str "data/processed" Path to processed dataset directory.
category str "Beauty" Dataset category (e.g. "Beauty").
min int 5 K-core minimum edges per user/item.

encoder

Field Type Default Meaning
text str "sentence-transformers/all-MiniLM-L6-v2" Text backbone.
visual str "resnet50" Visual backbone.
td int 384 Text feature dimension.
visual_dim int 2048 Visual feature dimension.
batch int 64 Encoding batch size.

masking

Field Type Default Meaning
kind str "bernoulli" Masking strategy (bernoulli, block).
ratio float 0.4 Fraction of modalities to mask. Must be in [0, 1].
seed int 42 Per-mask RNG seed.

retrieve

Field Type Default Meaning
kind str "mage" Retrieval strategy (mage, anchor, acs, bfs, none).
anchors int 10 Number of anchor nodes per query.
iters int 10 Number of expansion iterations.

encode

Field Type Default Meaning
kind str "transformer" Encoder kind.
hidden int 128 Hidden dimension. Must be positive.
layers int 2 Number of encoder layers.
heads int 4 Attention heads.
dropout float 0.5 Dropout rate.
pe int 20 Laplacian positional-encoding width.

route

Field Type Default Meaning
kind str "top" Soft-routing strategy (top, dense, gumbel, fixed).
p int 4 Top-p or number of entries. Must be positive.
tau float 0.5 Softmax temperature.

codebook

Field Type Default Meaning
kind str "gumbel" Codebook kind (gumbel, vq, identity).
size int 100 Codebook size. Must be positive.

complete

Field Type Default Meaning
kind str "mlp" Completer kind.
hidden int 128 Hidden dimension.

recommend

Field Type Default Meaning
kind str "light" Downstream ranker (light, mf, pop).
embed int 64 Embedding width.
layers int 3 GCN layers.

completion / recommendation

Training-loop hyperparameters. The *val fields must be in [0, 1]; the other integer fields are unbounded; grad_clip is float and clamps gradient norm.

Field Default (completion) Default (recommendation) Meaning
epochs 100 100 Number of training epochs.
batch 512 1024 Mini-batch size.
lr 1e-3 1e-3 Learning rate.
weight_decay 1e-5 1e-5 Adam weight decay.
usage 1.0 n/a Usage-loss weight (completion only).
balance 1.0 n/a Balance-loss weight (completion only).
grad_clip 1.0 1.0 Gradient-norm clip.
val 0.1 0.1 Validation fraction. Must be in [0, 1].
patience 10 10 Early-stopping patience.
amp False False Mixed-precision toggle.
negatives n/a 1 Negatives per positive (recommendation only).

eval

Field Type Default Meaning
ks tuple[int, ...] (10, 20) K values for ranking metrics. Each must be positive.
robustness tuple[float, ...] (0.1..0.9) Mask ratios for the robustness sweep.
ablations tuple[str, ...] (noretry, nope, nobook) Registered ablation condition names.

serve

Field Type Default Meaning
host str "0.0.0.0" Bind host.
port int 8080 Bind port.
workers int 1 Number of uvicorn workers.
auth bool False Whether auth is enabled (read-only; configure via env vars).

| metrics_enabled | bool | True | Expose the Prometheus metrics endpoint. | | metrics_path | str | /metrics | Prometheus endpoint path. |

log

Field Type Default Meaning
level str "INFO" Log level.
structured bool True Emit JSON instead of plain text.
directory str "runs" JSONL files directory.

Validation

Config.__post_init__ calls validate() on every construction path: Config(), Config.defaults(), Config.from_dict, Config.load, Config.env, and inline Config(**kwargs) inside notebooks and tests. A field out of range raises Cfg immediately, not downstream.

Saving

config.save("configs/my.yaml")

save writes the YAML round-trip; Config.load of the saved file reproduces the same cfg_hash, which is what the manifest sidecar binds to.