Can't pickle MXNet Modules

[fedorzh]

Description

Can't pickle mxnet Modules

Environment info (Required)

print pickle.__version__
print mx.__version__
$Revision: 72223 $
0.12.0

Error Message:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-36-024323561f98> in <module>()
      1 import pickle
----> 2 pickle.dumps(mlp_model)

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in dumps(obj, protocol)
   1378 def dumps(obj, protocol=None):
   1379     file = StringIO()
-> 1380     Pickler(file, protocol).dump(obj)
   1381     return file.getvalue()
   1382 

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in dump(self, obj)
    222         if self.proto >= 2:
    223             self.write(PROTO + chr(self.proto))
--> 224         self.save(obj)
    225         self.write(STOP)
    226 

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    329 
    330         # Save the reduce() output and finally memoize the object
--> 331         self.save_reduce(obj=obj, *rv)
    332 
    333     def persistent_id(self, obj):

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in save_reduce(self, func, args, state, listitems, dictitems, obj)
    423 
    424         if state is not None:
--> 425             save(state)
    426             write(BUILD)
    427 

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    284         f = self.dispatch.get(t)
    285         if f:
--> 286             f(self, obj) # Call unbound method with explicit self
    287             return
    288 

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in save_dict(self, obj)
    653 
    654         self.memoize(obj)
--> 655         self._batch_setitems(obj.iteritems())
    656 
    657     dispatch[DictionaryType] = save_dict

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in _batch_setitems(self, items)
    667             for k, v in items:
    668                 save(k)
--> 669                 save(v)
    670                 write(SETITEM)
    671             return

/home/ubuntu/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    304             reduce = getattr(obj, "__reduce_ex__", None)
    305             if reduce:
--> 306                 rv = reduce(self.proto)
    307             else:
    308                 reduce = getattr(obj, "__reduce__", None)

/home/ubuntu/anaconda2/lib/python2.7/copy_reg.pyc in _reduce_ex(self, proto)
     68     else:
     69         if base is self.__class__:
---> 70             raise TypeError, "can't pickle %s objects" % base.__name__
     71         state = base(self)
     72     args = (self.__class__, base, state)

TypeError: can't pickle module objects

Minimum reproducible example

net = mx.sym.Variable('data')
net = mx.sym.flatten(net)
net  = mx.sym.FullyConnected(net, num_hidden=128)
net = mx.sym.Activation(net, act_type="relu")
net = mx.sym.FullyConnected(net, num_hidden = 64)
net = mx.sym.Activation(net, act_type="relu")
net = mx.sym.FullyConnected(net, num_hidden=10)
net = mx.sym.SoftmaxOutput(net, name='softmax')
mlp_model = mx.mod.Module(symbol=net, context=mx.gpu())
import pickle
pickle.dumps(mlp_model)

[edmBernard]

why did you want to pickle the whole module class ? you can saved your symbol with: https://mxnet.incubator.apache.org/api/python/symbol.html#mxnet.symbol.Symbol.save

[fedorzh]

For example, I use parallel processing to distribute my training jobs. joblib uses pickle in multiprocessing

[edmBernard]

Honestly I don't think module can be pickle. Mxnet have lot's of C++ inside. For multi cpu core processing (if you don't use GPU) Mxnet support configuration with environnement variable: https://mxnet.incubator.apache.org/how_to/env_var.html More you can also use NNPACK to parallelize training operation over cpu.

[fedorzh]

If you take the old interface with mxnet.model, it can be pickled. Training of the model is actually not the longest part of my pipeline (sometimes, adds an insignificant part), a lot of other numpy-based machinery is happening outside of it, and parallelization helps immensely with that - I have to run multiple processes with different seeds

[leleamol]

Proposed Labels:"Feature Request", "Module","Python"

[sliawatimena]

I am using windows 7, anaconda navigator 1.9.2, Python 3.6.6, Jupyter notebook 5.7.0, try to learn code from Gluon crash course chapter 5.

I already add: import pickle

I got stuck at

for data, label in train_data:
    print(data.shape, label.shape)
    break
---
AttributeError                            Traceback (most recent call last)
<ipython-input-8-91a66f98d1d2> in <module>()
----> 1 for data, label in train_data:
      2     print(data.shape, label.shape)
      3     break

E:\Anaconda\envs\mxnet\lib\site-packages\mxnet\gluon\data\dataloader.py in __iter__(self)
    282         # multi-worker
    283         return _MultiWorkerIter(self._num_workers, self._dataset,
--> 284                                 self._batchify_fn, self._batch_sampler)
    285 
    286     def __len__(self):

E:\Anaconda\envs\mxnet\lib\site-packages\mxnet\gluon\data\dataloader.py in __init__(self, num_workers, dataset, batchify_fn, batch_sampler)
    142                 args=(self._dataset, self._key_queue, self._data_queue, self._batchify_fn))
    143             worker.daemon = True
--> 144             worker.start()
    145             workers.append(worker)
    146 

E:\Anaconda\envs\mxnet\lib\multiprocessing\process.py in start(self)
    103                'daemonic processes are not allowed to have children'
    104         _cleanup()
--> 105         self._popen = self._Popen(self)
    106         self._sentinel = self._popen.sentinel
    107         # Avoid a refcycle if the target function holds an indirect

E:\Anaconda\envs\mxnet\lib\multiprocessing\context.py in _Popen(process_obj)
    221     @staticmethod
    222     def _Popen(process_obj):
--> 223         return _default_context.get_context().Process._Popen(process_obj)
    224 
    225 class DefaultContext(BaseContext):

E:\Anaconda\envs\mxnet\lib\multiprocessing\context.py in _Popen(process_obj)
    320         def _Popen(process_obj):
    321             from .popen_spawn_win32 import Popen
--> 322             return Popen(process_obj)
    323 
    324     class SpawnContext(BaseContext):

E:\Anaconda\envs\mxnet\lib\multiprocessing\popen_spawn_win32.py in __init__(self, process_obj)
     63             try:
     64                 reduction.dump(prep_data, to_child)
---> 65                 reduction.dump(process_obj, to_child)
     66             finally:
     67                 set_spawning_popen(None)

E:\Anaconda\envs\mxnet\lib\multiprocessing\reduction.py in dump(obj, file, protocol)
     58 def dump(obj, file, protocol=None):
     59     '''Replacement for pickle.dump() using ForkingPickler.'''
---> 60     ForkingPickler(file, protocol).dump(obj)
     61 
     62 #

AttributeError: Can't pickle local object 'Dataset.transform_first.<locals>.base_fn'

please help. thanks.

[piyushghai]

@sliawatimena Can you file a separate issue on this repository and also can you provide a minimum reproducible example to help debug this issue ?

From the stacktrace that you've posted, it seems unclear as to where you are using pickle ? Also, are you using pickle.dump or pickle.load ?

[sliawatimena]

Dear @piyushghai,

I just copy from 5. Train the neural network, from step 1 - 5 are okay. In step 6, the error message are as previous post.

From googling results: this looks like a Windows-specific problem with Python multiprocessing and Jupyter Notebook. Please help.

Thanks.

Suryadi

[egeaydin]

This https://github.com/apache/incubator-mxnet/issues/10562 issue helped me.

train_data = gluon.data.DataLoader(
    mnist_train, batch_size=batch_size, shuffle=True, num_workers=0)

Change num_works=4 to num_works=0.

Also, do the same for validation data.

Hope this helps.