请问一下文中设置的batch size和chunk size的区别

[plz583585760]

感谢作者的代码，我用的win10系统，batchsize可以和原文一样设置为16，但是chunksize只能设置为1才能运行，暂时没有找到解决方法，也仅有此处一处不同，导致模型准确率和原文相差甚远，能否释义一下chunksize运作含义以及对精度的影响，对windows系统如何才能修改为chunksize[16]运行呢，不然只能装虚拟机尝试了，麻烦作者了，辛苦！

[eeyrw]

是不是会遇上这样的错误：

training start...
  0%|                                   | 0/500000 [00:00<?, ?it/s]
Traceback (most recent call last):
  File ".\train.py", line 233, in <module>
    train(training_dbs, validation_db, args.start_iter, args.freeze) # 0
  File ".\train.py", line 158, in train
    = nnet.train(iteration, save, viz_split, **training)
  File "C:\Users\yuan\Desktop\LSTR\nnet\py_factory.py", line 115, in train
    ys)
  File "C:\Users\yuan\Desktop\LSTR\myvenv\lib\site-packages\torch\nn\modules\module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\data_parallel.py", line 66, in forward
    inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids, self.chunk_sizes)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\data_parallel.py", line 77, in scatter
    return scatter_kwargs(inputs, kwargs, device_ids, dim=self.dim, chunk_sizes=self.chunk_sizes)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 30, in scatter_kwargs
    inputs = scatter(inputs, target_gpus, dim, chunk_sizes) if inputs else []
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 25, in scatter
    return scatter_map(inputs)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 18, in scatter_map
    return list(zip(*map(scatter_map, obj)))
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 20, in scatter_map
    return list(map(list, zip(*map(scatter_map, obj))))
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 15, in scatter_map
    return Scatter.apply(target_gpus, chunk_sizes, dim, obj)
  File "C:\Users\yuan\Desktop\LSTR\myvenv\lib\site-packages\torch\nn\parallel\_functions.py", line 96, in forward
    outputs = comm.scatter(input, target_gpus, chunk_sizes, ctx.dim, streams)
  File "C:\Users\yuan\Desktop\LSTR\myvenv\lib\site-packages\torch\nn\parallel\comm.py", line 189, in scatter
    return tuple(torch._C._scatter(tensor, devices, chunk_sizes, dim, streams))
RuntimeError: start (0) + length (16) exceeds dimension size (1).

[Galaxy-Ding]

是不是会遇上这样的错误：

training start...
  0%|                                   | 0/500000 [00:00<?, ?it/s]
Traceback (most recent call last):
  File ".\train.py", line 233, in <module>
    train(training_dbs, validation_db, args.start_iter, args.freeze) # 0
  File ".\train.py", line 158, in train
    = nnet.train(iteration, save, viz_split, **training)
  File "C:\Users\yuan\Desktop\LSTR\nnet\py_factory.py", line 115, in train
    ys)
  File "C:\Users\yuan\Desktop\LSTR\myvenv\lib\site-packages\torch\nn\modules\module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\data_parallel.py", line 66, in forward
    inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids, self.chunk_sizes)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\data_parallel.py", line 77, in scatter
    return scatter_kwargs(inputs, kwargs, device_ids, dim=self.dim, chunk_sizes=self.chunk_sizes)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 30, in scatter_kwargs
    inputs = scatter(inputs, target_gpus, dim, chunk_sizes) if inputs else []
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 25, in scatter
    return scatter_map(inputs)
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 18, in scatter_map
    return list(zip(*map(scatter_map, obj)))
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 20, in scatter_map
    return list(map(list, zip(*map(scatter_map, obj))))
  File "C:\Users\yuan\Desktop\LSTR\models\py_utils\scatter_gather.py", line 15, in scatter_map
    return Scatter.apply(target_gpus, chunk_sizes, dim, obj)
  File "C:\Users\yuan\Desktop\LSTR\myvenv\lib\site-packages\torch\nn\parallel\_functions.py", line 96, in forward
    outputs = comm.scatter(input, target_gpus, chunk_sizes, ctx.dim, streams)
  File "C:\Users\yuan\Desktop\LSTR\myvenv\lib\site-packages\torch\nn\parallel\comm.py", line 189, in scatter
    return tuple(torch._C._scatter(tensor, devices, chunk_sizes, dim, streams))
RuntimeError: start (0) + length (16) exceeds dimension size (1).

请问解决了？我也遇到一样的问题

[atubo1024]

这是由于windows下子进程的system_configs和主进程的system_configs不相同导致的。在linux通过fork出来的子进程会共享主进程的初始状态，但windows并不会这样，所以在预加载数据的子进程中system_configs实际上是默认值，和主进程不一致。
解决的方法：可以考虑在windows环境中，将system_configs的内容通过参数传递给子进程，子进程初始化时再把参数覆盖到system_configs，例如：

# train.py
import platform
# ...
def init_parallel_jobs(dbs, queue, fn):
    if platform.system().lower() == 'windows':
        tasks = [Process(target=prefetch_data, args=(db, queue, fn, system_configs._configs)) for db in dbs]
    else:
        tasks = [Process(target=prefetch_data, args=(db, queue, fn, None)) for db in dbs]
    # ...
 
def prefetch_data(db, queue, sample_data, system_configs_data):
    if platform.system().lower() == 'windows':
        system_configs._configs = system_configs_data
    # ...