使用上下文装饰器调试Pytorch的内存泄漏问题

def memleak_wrapper(func):
def wrap(*args, **kwargs):
print("num tensors start is ...")
out = func(*args, **kwargs)
print("num tensors end is ...")
return out
return wrap@memleak_wrapper
 def function_to_debug(x):
print(f"put line(s) of code here. Input is {x}")
out = x + 10
return outout = function_to_debug(x=1000)
 print(f"out is {out}")
 
 #输入类似这样
 #num tensors start is ...
 #put line(s) of code here. Input is 1000
 #num tensors end is ...
 #outis 1010

with open("file") as f:
…

 from contextlib import ContextDecorator
 
 class check_memory_leak_context(ContextDecorator):
def __enter__(self):
print('Starting')
return self
 
def __exit__(self, *exc):
print('Finishing')
return False

class check_memory_leak_context(ContextDecorator):
def __enter__(self):
self.start = get_n_tensors()
return self def __exit__(self, *exc):
self.end = get_n_tensors()
increase = self.end — self.start
 
if increase > 0:
print(f”num tensors increased with"
f"{self.end — self.start} !”)
else:
print(”no added tensors”)
return False

 import gc
 def get_n_tensors():
tensors= []
for obj in gc.get_objects():
try:
if (torch.is_tensor(obj) or
(hasattr(obj, ‘data’) and
torch.is_tensor(obj.data))):
tensors.append(obj)
except:
pass
return len(tensors)

 x = arbitrary_operation(x)
 ...
 with check_memory_leak_context():
y = x[0].permute(1, 2, 0).cpu().detach().numpy()
x = some_harmless_operation()
 ...
 x = another_arbitrary_operation(x)