检查点序列化过程-2

save(k)

• save_str
• obj:’model_state_dict’
• n:16
• data:b’X\x10\x00\x00\x00model_state_dict’
• self.memoize(obj)
• idx:1 obj:’model_state_dict’
• id(obj):131443899919424

save(v)

• t:collections.OrderedDict
• f:None
• 检查一个可能存在的“私有分发表”（dispatch table）,找不到
• 是否是 Python 的类类型（type）
• reduce_ex 方法存在，rv = reduce(self.proto)
• rv = reduce(self.proto)：调用 reduce_ex 方法，并将返回值存储在 rv 中。
• rv是一个元祖tuple
• l = len(rv)
• 5
• 最终调用save_reduce

save_reduce

• func:<class ‘collections.OrderedDict’>
• args:() rv,tuple
• save(func)

save(func)

• save(type)

def save_type(self, obj):
       return self.save_global(obj)

   dispatch[FunctionType] = save_global
   dispatch[type] = save_type

• self.save_global(obj)

• obj/obj2:<class ‘collections.OrderedDict’>

• name/lastname:’OrderedDict’

• module_name:’collections’

• parent/module:import(module_name, level=0):<module ‘collections’ from ‘/home/dell/anaconda3/envs/torch_new_env/lib/python3.11/collections/__init__.py’>

• write(data):b’ccollections\nOrderedDict\n’

• memo(idx(2)):b’q\x02’

save(args)

• save_tuple(self, obj)
• obj:()

  def save_tuple(self, obj):
        if not obj: # tuple is empty
             if self.bin:
                 self.write(EMPTY_TUPLE)
             else:
                 self.write(MARK + TUPLE)
             return

• tuple是空，write(EMPTY_TUPLE):b’)’
• write(REDUCE)：b’R’

memo(obj)

• memo(idx(3)):b’q\x03’
• id(obj):132101266175680
• _batch_setitems

n=7 _batch_setitems

save(k)

• write(MARK):b’(‘
• save_str(embedding.weight)
• write(b’X\x10\x00\x00\x00embedding.weight’)
• memo(b’q\x04’)

save(v)

• rv = reduce(self.proto),看起来像TypedStorage

   -2.759326696395874
   -0.4928496479988098
   -0.9460598826408386
   0.12430991232395172
   -0.8162614107131958
   -0.46302852034568787
   2.887676239013672
   0.2881816625595093
   -0.3105534017086029
   -0.1721319556236267
   1.5114414691925049
   0.6535921096801758
  [torch.storage.TypedStorage(dtype=torch.float32, device=cpu) of size 10000], 0, (10000, 1), (1, 1), False, OrderedDict()))

• save_reduce()
• save(func):<function _rebuild_tensor_v2 at 0x7825cb0ff060>
• save(args) args看起来就是rv
• obj

  tensor([[ 1.1894],
          [-0.5411],
          [-2.4535],
          ...,
          [-0.1721],
          [ 1.5114],
          [ 0.6536]])

save(func)

• save_global
• name:’_rebuild_tensor_v2’
• module_name:’torch._utils’
• write(b’ctorch._utils\n_rebuild_tensor_v2\n_rebuild_tensor_v2\n’)
• memo(b’q\x05’)

save(args)

• save_tuple
• write(MARK):b’(‘
• save(element)

persistent_id

if isinstance(obj, torch.storage.TypedStorage):
                # TODO: Once we decide to break serialization FC, this case
                # can be deleted
                storage = obj._untyped_storage
                print(obj._untyped_storage)
                storage_dtype = obj.dtype
                storage_type_str = obj._pickle_storage_type()
                storage_type = getattr(torch, storage_type_str)
                storage_numel = obj._size()

• storage_dtype:torch.float32
• storage_type_str:’FloatStorage’
• storage_type:<class ‘torch.FloatStorage’>
• storage_numel:10000
• storage_key=0

  def save_pers(self, pid):
         # Save a persistent id reference
         if self.bin:
             self.save(pid, save_persistent_id=False)
             self.write(BINPERSID)