BalancedTernaryEngine.
Copy Markdown
A neural network layer with ternary weights.
Supports:
- Forward pass with add/sub/skip (no multiplication)
- Per-channel scale factors and bias
- Conversion from dense (FP32) weights
- Packed weight storage (10 trits per 16-bit word)
- Serialization to/from binary format
Summary
Functions
Forward pass through the ternary layer.
Forward pass with Nx tensors (for batched inference).
Deserialize layer from binary.
Create a TernaryLayer from a dense (FP32) weight matrix.
Serialize layer to binary (.tbin record payload).
Types
@type t() :: %BalancedTernaryEngine.Layer{ bias: [float()], delta: float(), name: String.t(), original_shape: term(), packing: term(), scales: [float()], shape: {non_neg_integer(), non_neg_integer()}, sparsity: float(), weight_packed: binary() }
Functions
Forward pass through the ternary layer.
Computes: output[i] = scale[i] sum_j(W_ternary[i][j] x[j]) + bias[i]
Uses add/sub/skip — no multiplication for the weight-activation product.
Examples
iex> w = Nx.tensor([[0.8, -0.3, 0.1], [0.5, -0.2, 0.7]])
iex> layer = BalancedTernaryEngine.Layer.from_dense(w, "test", delta: 0.4)
iex> BalancedTernaryEngine.Layer.forward(layer, [1.0, 2.0, 3.0])
[0.5666666626930237, 1.619999885559082]@spec forward_nx(t(), Nx.Tensor.t()) :: Nx.Tensor.t()
Forward pass with Nx tensors (for batched inference).
Deserialize layer from binary.
Mirrors to_binary/1 field-for-field. The pattern match is total:
a truncated or malformed payload fails to match and raises, which the
Exporter wraps into an error tuple.
@spec from_dense(Nx.Tensor.t(), String.t(), keyword()) :: t()
Create a TernaryLayer from a dense (FP32) weight matrix.
Quantization is delegated to BalancedTernaryEngine.Quantizer.quantize/2;
pass mode: (:delta, :twn, :absmean) and delta: options
through. Automatically packs weights and records per-channel scales.
Examples
iex> w = Nx.tensor([[0.8, -0.3, 0.1, -0.9], [0.5, -0.2, 0.7, -0.4]])
iex> layer = BalancedTernaryEngine.Layer.from_dense(w, "fc1", delta: 0.4)
iex> layer.shape
{2, 4}
iex> layer.name
"fc1"Serialize layer to binary (.tbin record payload).
Layout (all integers big-endian):
name_len::16 | name | out::32 | in::32
delta::float-32 | sparsity::float-32
num_scales::32 | scales (num_scales × float-16)
num_biases::32 | biases (num_biases × float-16)
packed_size::32 | packed trits (see Packer)Scales/biases use half precision: they are calibration constants, and f16 halves their footprint at negligible accuracy cost.