decoding

游雁

2024-06-12 be26169447b2e5f8f38c97af8f5f6a201bc6ce40

 funasr/train_utils/load_pretrained_model.py

@@ -7,141 +7,91 @@
import torch
import torch.nn
import torch.optim


def filter_state_dict(
   dst_state: Dict[str, Union[float, torch.Tensor]],
   src_state: Dict[str, Union[float, torch.Tensor]],
):
   """Filter name, size mismatch instances between dicts.

   Args:
      dst_state: reference state dict for filtering
      src_state: target state dict for filtering

   """
   match_state = {}
   for key, value in src_state.items():
      if key in dst_state and (dst_state[key].size() == src_state[key].size()):
         match_state[key] = value
      else:
         if key not in dst_state:
            logging.warning(
               f"Filter out {key} from pretrained dict"
               + " because of name not found in target dict"
            )
         else:
            logging.warning(
               f"Filter out {key} from pretrained dict"
               + " because of size mismatch"
               + f"({dst_state[key].size()}-{src_state[key].size()})"
            )
   return match_state

def assigment_scope_map(dst_state: dict, src_state: dict, scope_map: str=None):
   """Compute the union of the current variables and checkpoint variables."""
   import collections
   import re

   # current model variables
   name_to_variable = collections.OrderedDict()
   for name, var in dst_state.items():
      name_to_variable[name] = var
	
   scope_map_num = 0
   if scope_map is not None:
      scope_map = scope_map.split(",")
      scope_map_num = len(scope_map) // 2
      for scope_map_idx in range(scope_map_num):
         scope_map_id = scope_map_idx * 2
         logging.info('assignment_map from scope {} to {}'.format(scope_map[scope_map_id], scope_map[scope_map_id+1]))
	
   assignment_map = {}
   for name, var in src_state.items():

      if scope_map:
         for scope_map_idx in range(scope_map_num):
            scope_map_id = scope_map_idx * 2
            try:
               idx = name.index(scope_map[scope_map_id])
               new_name = scope_map[scope_map_id+1] + name[idx + len(scope_map[scope_map_id]):]
               if new_name in name_to_variable:
                  assignment_map[name] = var
            except:
               continue
      else:
         if name in name_to_variable:
            assignment_map[name] = var
	
   return assignment_map
import pdb


def load_pretrained_model(
   path: str,
   model: torch.nn.Module,
   ignore_init_mismatch: bool,
   map_location: str = "cpu",
   oss_bucket=None,
   scope_map="module.:none",
   excludes=None,
    path: str,
    model: torch.nn.Module,
    ignore_init_mismatch: bool = True,
    map_location: str = "cpu",
    oss_bucket=None,
    scope_map=[],
    excludes=None,
    **kwargs,
):
   """Load a model state and set it to the model.
    """Load a model state and set it to the model.

   Args:
      init_param: <file_path>:<src_key>:<dst_key>:<exclude_Keys>
    Args:
            init_param: <file_path>:<src_key>:<dst_key>:<exclude_Keys>

   Examples:
    Examples:

   """
	
   obj = model
   dst_state = obj.state_dict()
	
   print(f"ckpt: {path}")
   if oss_bucket is None:
      src_state = torch.load(path, map_location=map_location)
   else:
      buffer = BytesIO(oss_bucket.get_object(path).read())
      src_state = torch.load(buffer, map_location=map_location)
   if "state_dict" in src_state:
      src_state = src_state["state_dict"]
	
   src_state = src_state["model"] if "model" in src_state else src_state
	
   if isinstance(scope_map, str):
      scope_map = scope_map.split(",")
	
   for k in dst_state.keys():
      # if not k.startswith("module.") and "module." + k in src_state.keys():
      #    k_ddp = "module." + k
      # else:
      #    k_ddp = k
      k_src = k
    """

      if scope_map is not None:
         src_prefix = ""
         dst_prefix = ""
         for i in range(0, len(scope_map), 2):
            src_prefix = scope_map[i] if scope_map[i].lower() != "none" else ""
            dst_prefix = scope_map[i+1] if scope_map[i+1].lower() != "none" else ""
    obj = model
    dst_state = obj.state_dict()

            if k.startswith(dst_prefix) and k.replace(dst_prefix, src_prefix) in src_state.keys():
               k_src = k.replace(dst_prefix, src_prefix)
               print(f"init param, map: {k} from {k_src} in ckpt")
					
      if k_src in src_state.keys():
         dst_state[k] = src_state[k_src]
				
      # if k_ddp.startswith("audio_encoder"):
      #    if k_ddp.replace("audio_encoder", "encoder.model") in src_state.keys():
      #       k_ddp = k_ddp.replace("audio_encoder", "encoder.model")
      # if k_ddp.startswith("adaptor"):
      #    if k_ddp.replace("adaptor", "encoder_projector") in src_state.keys():
      #       k_ddp = k_ddp.replace("adaptor", "encoder_projector")
      # if k_ddp in src_state:
      #    dst_state[k] = src_state[k_ddp]
      else:
         print(f"Warning, miss key in ckpt: {k}, mapped: {k_src}")
			
   flag = obj.load_state_dict(dst_state, strict=False)
   # print(flag)
    logging.info(f"ckpt: {path}")

    if oss_bucket is None:
        src_state = torch.load(path, map_location=map_location)
    else:
        buffer = BytesIO(oss_bucket.get_object(path).read())
        src_state = torch.load(buffer, map_location=map_location)

    src_state = src_state["state_dict"] if "state_dict" in src_state else src_state
    src_state = src_state["model_state_dict"] if "model_state_dict" in src_state else src_state
    src_state = src_state["model"] if "model" in src_state else src_state

    if isinstance(scope_map, str):
        scope_map = scope_map.split(",")
    scope_map += ["module.", "None"]
    logging.info(f"scope_map: {scope_map}")

    if excludes is not None:
        if isinstance(excludes, str):
            excludes = excludes.split(",")
    logging.info(f"excludes: {excludes}")

    for k in dst_state.keys():

        for k_ex in excludes:
            if k.startswith(k_ex):
                logging.info(f"key: {{k}} matching: {k_ex}, excluded")
                continue

        k_src = k

        if scope_map is not None:
            src_prefix = ""
            dst_prefix = ""
            for i in range(0, len(scope_map), 2):
                src_prefix = scope_map[i] if scope_map[i].lower() != "none" else ""
                dst_prefix = scope_map[i + 1] if scope_map[i + 1].lower() != "none" else ""

                if dst_prefix == "" and (src_prefix + k) in src_state.keys():
                    k_src = src_prefix + k
                    if not k_src.startswith("module."):
                        logging.info(f"init param, map: {k} from {k_src} in ckpt")
                elif (
                    k.startswith(dst_prefix)
                    and k.replace(dst_prefix, src_prefix, 1) in src_state.keys()
                ):
                    k_src = k.replace(dst_prefix, src_prefix, 1)
                    if not k_src.startswith("module."):
                        logging.info(f"init param, map: {k} from {k_src} in ckpt")

        if k_src in src_state.keys():
            if ignore_init_mismatch and dst_state[k].shape != src_state[k_src].shape:
                logging.info(
                    f"ignore_init_mismatch:{ignore_init_mismatch}, dst: {k, dst_state[k].shape}, src: {k_src, src_state[k_src].shape}"
                )
            else:
                dst_state[k] = src_state[k_src]

        else:
            logging.info(f"Warning, miss key in ckpt: {k}, mapped: {k_src}")

    flag = obj.load_state_dict(dst_state, strict=True)
    logging.info(f"Loading ckpt: {path}, status: {flag}")