# `Text.Language.Classifier.Fasttext.Model`
[🔗](https://github.com/kipcole9/text/blob/v0.5.0/lib/language/classifier/fasttext/model.ex#L1)

A fully-loaded fastText model.

Holds the parsed args and dictionary plus the input and output matrices as
`Nx` tensors. The input matrix is the largest piece of memory in the
struct: for `lid.176.bin` it is a `{nwords + bucket, dim}` tensor of
float32 values, approximately 128 MB.

Models are produced by `Text.Language.Classifier.Fasttext.ModelLoader.load/2`
from a fastText `.bin` file.

### Fields

* `args` is the parsed `Text.Language.Classifier.Fasttext.Args` struct.

* `dictionary` is the parsed
  `Text.Language.Classifier.Fasttext.Dictionary` struct.

* `input_matrix` is an `Nx` tensor of shape `{args.bucket + dictionary.nwords, args.dim}`
  holding the row-major input embeddings (word rows first, then subword
  n-gram rows).

* `output_matrix` is an `Nx` tensor of shape
  `{dictionary.nlabels, args.dim}` holding the per-label output vectors.

* `labels` is the list of language label strings (with the `__label__`
  prefix stripped) in row order matching `output_matrix`. For `lid.176`
  this is a 176-element list such as `["en", "zh-Hans", ...]`.

# `loss_state`

```elixir
@type loss_state() :: Text.Language.Classifier.Fasttext.HuffmanTree.t() | nil
```

Loss-specific decoding state, built once at load time and reused for every
prediction.

* For `:hs` (hierarchical softmax) the state is a
  `Text.Language.Classifier.Fasttext.HuffmanTree` constructed from the
  label counts.

* For `:softmax` no extra state is needed; the field is `nil`.

* `:ns` and `:ova` are not yet supported by the inference path.

# `t`

```elixir
@type t() :: %Text.Language.Classifier.Fasttext.Model{
  args: Text.Language.Classifier.Fasttext.Args.t(),
  dictionary: Text.Language.Classifier.Fasttext.Dictionary.t(),
  input_matrix: Nx.Tensor.t(),
  labels: [String.t()],
  loss_state: loss_state(),
  output_matrix: Nx.Tensor.t()
}
```

---

*Consult [api-reference.md](api-reference.md) for complete listing*