Quantization-Guided Training¶
Category: callback.quantization
Version: 1.0.0
UUID: f964437880c5a1fb1cbf3fb51d67964f86dcfc12f294dde594dfdde076e8fef3
Execution context: leip_af
Choice Priority: 390.0
Value Parameters¶
| Name | Synonyms | Type | Values |
|---|---|---|---|
| Bit precision for biases | quantization.guided_training.bias_bit_precision |
scalar | int |
| Bit precision | quantization.guided_training.bit_precision |
scalar | int |
| The number of quantization bits specified on a per named-layer | quantization.guided_training.custom_layer_bits |
dict | int |
| Lambda of quantization loss | quantization.guided_training.qunt_loss_multiplier,quantization.guided_training.lamb2 |
scalar | float |
| Lambda of l2 loss | quantization.guided_training.l2_loss_multiplier,quantization.guided_training.lamb3 |
scalar | float |
| Should QGT be applied to bias layers | quantization.guided_training.quantize_bias |
scalar | bool |
| Should QGT be applied to batchnorm layers | quantization.guided_training.quantize_batchnorm |
scalar | bool |
| Type of quantizer to use | quantization.quantizer |
choice | asymmetricasymmetricpcsymmetricsymmetricpc |
This component fits into¶
| Name | UUID | Synonyms |
|---|---|---|
| Full Recipe | 8b4b05... |
callbacks |