Conversion from Gluon to Module and vice versa?


#1

I would like to know how to convert between the two versions because it seems the quantization capabilities are mainly for the syms, arg_params, aux_params tuple style passing, which can be wrapped around modules well, but not gluon models(correct me if I’m wrong).

Here’s a small code snippet that trains a cnn model:

batch_size = 64
num_inputs = 784
num_outputs = 10
data_iter = mx.io.NDArrayIter(x, y, batch_size=batch_size)

num_fc = 512
net = gluon.nn.HybridSequential()
with net.name_scope():
    net.add(gluon.nn.Conv2D(channels=20, kernel_size=5, activation='relu'))
    net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2))
    net.add(gluon.nn.Conv2D(channels=50, kernel_size=5, activation='relu'))
    net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2))
    net.add(gluon.nn.Flatten())
    net.add(gluon.nn.Dense(num_fc, activation="relu"))
    net.add(gluon.nn.Dense(num_outputs))

net.hybridize()
# Parameter initialization
net.collect_params().initialize(mx.init.Xavier(magnitude=2.24), ctx=ctx)
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': .1})
softmax_cross_entropy = gluon.loss.SoftmaxCrossEntropyLoss()
for i, batch in enumerate(data_iter):
    data = batch.data[0].as_in_context(ctx)
    label = batch.label[0].as_in_context(ctx)
    with autograd.record():
        output = net(data)
        loss = softmax_cross_entropy(output, label)
    loss.backward()
    trainer.step(data.shape[0])

If I want to quantize a gluon model, I would try to serialize gluon into disk, and then bring it back out as module. This may cause troubles:

import os
net.export('mxnet')
mod = mx.module.Module.load('mxnet', 0) # 0 epoch

FWIW, I got a warning during the loading step, but was not sure what this was about:

/Users/ray_zhang/anaconda3/envs/idp3/lib/python3.6/site-packages/mxnet/module/base_module.py:54: UserWarning: You created Module with Module(..., label_names=['softmax_label']) but input with name 'softmax_label' is not found in symbol.list_arguments(). Did you mean one of:
	data
  warnings.warn(msg)

and as per the module API:

mod.bind( data_shapes = data_iter.provide_data, 
          label_shapes = data_iter.provide_label)
mod.predict(x)

but it does not work upon bind(), with the following stacktrace:

----------------------------------------------
KeyError     Traceback (most recent call last)
<ipython-input-10-f53137bb5e95> in <module>()
      1 mod.bind( data_shapes = data_iter.provide_data, 
----> 2           label_shapes = data_iter.provide_label)
      3 mod.predict(x)

~/anaconda3/envs/idp3/lib/python3.6/site-packages/mxnet/module/module.py in bind(self, data_shapes, label_shapes, for_training, inputs_need_grad, force_rebind, shared_module, grad_req)
    434                                                      fixed_param_names=self._fixed_param_names,
    435                                                      grad_req=grad_req, group2ctxs=self._group2ctxs,
--> 436                                                      state_names=self._state_names)
    437         self._total_exec_bytes = self._exec_group._total_exec_bytes
    438         if shared_module is not None:

~/anaconda3/envs/idp3/lib/python3.6/site-packages/mxnet/module/executor_group.py in __init__(self, symbol, contexts, workload, data_shapes, label_shapes, param_names, for_training, inputs_need_grad, shared_group, logger, fixed_param_names, grad_req, state_names, group2ctxs)
    281 
    282         eprint(sys._getframe().f_lineno, data_shapes, label_shapes)
--> 283         self.bind_exec(data_shapes, label_shapes, shared_group)
    284 
    285     def decide_slices(self, data_shapes):

~/anaconda3/envs/idp3/lib/python3.6/site-packages/mxnet/module/executor_group.py in bind_exec(self, data_shapes, label_shapes, shared_group, reshape)
    388         if label_shapes is not None:
    389             self.label_names = [i.name for i in self.label_shapes]
--> 390         self._collect_arrays()
    391 
    392     def reshape(self, data_shapes, label_shapes):

~/anaconda3/envs/idp3/lib/python3.6/site-packages/mxnet/module/executor_group.py in _collect_arrays(self)
    324             self.label_arrays = [[(self.slices[i], e.arg_dict[name])
    325                                   for i, e in enumerate(self.execs)]
--> 326                                  for name, _ in self.label_shapes]
    327         else:
    328             self.label_arrays = None

~/anaconda3/envs/idp3/lib/python3.6/site-packages/mxnet/module/executor_group.py in <listcomp>(.0)
    324             self.label_arrays = [[(self.slices[i], e.arg_dict[name])
    325                                   for i, e in enumerate(self.execs)]
--> 326                                  for name, _ in self.label_shapes]
    327         else:
    328             self.label_arrays = None

~/anaconda3/envs/idp3/lib/python3.6/site-packages/mxnet/module/executor_group.py in <listcomp>(.0)
    323                 eprint(323, e.arg_dict.keys())
    324             self.label_arrays = [[(self.slices[i], e.arg_dict[name])
--> 325                                   for i, e in enumerate(self.execs)]
    326                                  for name, _ in self.label_shapes]
    327         else:

KeyError: 'softmax_label'

Which is complaining about me missing that label in my e.arg_dict.

I printed out e.arg_dict:

dict_keys(['data', 'hybridsequential1_conv0_weight', 'hybridsequential1_conv0_bias', 'hybridsequential1_conv1_weight', 'hybridsequential1_conv1_bias', 'hybridsequential1_dense0_weight', 'hybridsequential1_dense0_bias', 'hybridsequential1_dense1_weight', 'hybridsequential1_dense1_bias'])

And indeed, softmax_label is not in there. Where is this label coming from and how can I convert gluon to module correctly?


#2

With regards to your initial point on quantization, this can be done with Gluon.

You need to cast your Block (i.e. network) and also cast your inputs to your network:

net = net.cast(np.float16)
...
data = data.astype(np.float16)
...
net(data)

Check out the video here for more details.


#3

Hi there,

Thanks for the answer. I was not aware of this, but I was also referring to another implementation where the layers are artificially modified to have thresholding and casted to int8:

See: https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/contrib/quantization.py

and here for a tutorial: https://github.com/apache/incubator-mxnet/tree/master/example/quantization

This is what I am trying to achieve, since even though fp16 is a good win, int8 would be a much better win. (the symmetric int8 quantization technique cannot be simply done via a cast, and must compute KL divergence to compute the best ‘cast’)

I am looking into the source code for mxnet Module right now and do not understand what label_arrays and label_names are exactly(the documentations don’t describe this), and I think I may be able to understand the source of my problems if that was explained.