code/pointcept_framework/pointcept/models/quantization/quant_utils.py

# pointcept/models/quantization/quant_utils.py

import torch
import torch.nn as nn

def convert_ptv3_to_bi_ptv3(model, verbose=False):
    """
    Convert a PTV3 model to Bi-PTV3.
    This version is COMPATIBLE with your existing checkpoint that has LSR scales but NO ema_offset.
    """
    from .binary_layers import BiLinearLSR

    exclusion_keywords = ['embedding', 'stem', 'seg_head']
    
    print("Starting the conversion from PTV3 to Bi-PTV3...")
    
    # Step 1: Replace Linear layers with BiLinearLSR (This is what you did during training)
    def _recursive_replace(module, parent_name=""):
        for child_name, child_module in module.named_children():
            full_name = f"{parent_name}.{child_name}" if parent_name else child_name
            is_excluded = any(keyword in full_name for keyword in exclusion_keywords)

            if isinstance(child_module, nn.Linear) and not is_excluded:
                old_layer = child_module
                new_layer = BiLinearLSR(
                    in_features=old_layer.in_features,
                    out_features=old_layer.out_features,
                    bias=(old_layer.bias is not None)
                )
                new_layer.weight.data.copy_(old_layer.weight.data)
                if old_layer.bias is not None:
                    new_layer.bias.data.copy_(old_layer.bias.data)
                setattr(module, child_name, new_layer)
                if verbose:
                    print(f"SUCCESS: Replaced '{full_name}' with BiLinearLSR.")
            else:
                if is_excluded and verbose:
                    print(f"INFO: Skipping module '{full_name}' due to exclusion rules.")
                _recursive_replace(child_module, parent_name=full_name)

    _recursive_replace(model)
    
    # Step 2: DO NOT inject ema_offset. Your checkpoint doesn't have it, so we skip it.
    # This makes the function compatible with your existing model_last.pth.
    # We assume EMA was not used in your training run, which is fine for FLOPs calculation.
    
    print("Conversion to Bi-PTV3 complete! (Compatible with checkpoint without ema_offset)")
    return model