qkeras.qtools.quantized_operators.fused_bn_factory

quantized batch normliaztion quantizer implementation.

Classes

FusedBNFactory()

determine which quantizer implementation to use.
class qkeras.qtools.quantized_operators.fused_bn_factory.FusedBNFactory[source]

Bases: object

determine which quantizer implementation to use.

Create an fused bn instance. The type and bit width of the output_quantizer is deteremined from both the previous layer and batchnorm weight types:

z = bn(y) = bn_inv * x - fused_bias is the output of the previous layer and the following bn layer, with:

bn_inv = gamma * rsqrt(variance^2+epsilon) is computed from the

bn layer weights with inverse_quantizer datatype

x is the previous layer’s output fused_bias = bn_inv * bias + beta - bn_inv*mean where bias is

the bias term from the previous layer, beta and mean are the bn layer weights.

make_quantizer(prev_output_quantizer, beta_quantizer, mean_quantizer, inverse_quantizer, prev_bias_quantizer, use_beta, use_bias, qkeras_inverse_quantizer)[source]

Makes a fused_bn quantizer.

Parameters:

  • prev_output_quantizer (IQuantizer) – IQuantizer type. Previous layer output quantizer

  • beta_quantizer (IQuantizer) – IQuantizer type. bn layer beta quantizer

  • mean_quantizer (IQuantizer) – IQuantizer type. layer mean quantizer

  • inverse_quantizer (IQuantizer) – IQuantizer type. bn layer inverse quantizer

  • prev_bias_quantizer (IQuantizer) – IQuantizer type. conv layer bias quantizer

  • use_beta (bool) – Bool. whether enabling beta in batch_normalization layer

  • use_bias (bool) – Bool. Whether bias is used in conv layer.

  • qkeras_inverse_quantizer (BaseQuantizer) – qkeras quantizer type. bn layer inverse quantizer with qkeras quantizer type

Returns:

None