Naming of Gluon Parameter and Blocks

In gluon, each Parameter or Block has a name (and prefix). Parameter names are specified by users and Block names can be either specified by users or automatically created.

In this tutorial we talk about the best practices on naming. First, let’s import MXNet and Gluon:

from __future__ import print_function
import mxnet as mx
from mxnet import gluon

Naming Blocks

When creating a block, you can assign a prefix to it:

mydense = gluon.nn.Dense(100, prefix='mydense_')
print(mydense.prefix)
mydense_

When no prefix is given, Gluon will automatically generate one:

dense0 = gluon.nn.Dense(100)
print(dense0.prefix)
dense0_

When you create more Blocks of the same kind, they will be named with incrementing suffixes to avoid collision:

dense1 = gluon.nn.Dense(100)
print(dense1.prefix)
dense1_

Naming Parameters

Parameters within a Block will be named by prepending the prefix of the Block to the name of the Parameter:

print(dense0.collect_params())
dense0_ (
  Parameter dense0_weight (shape=(100, 0), dtype=<type 'numpy.float32'>)
  Parameter dense0_bias (shape=(100,), dtype=<type 'numpy.float32'>)
)

Name scopes

To manage the names of nested Blocks, each Block has a name_scope attached to it. All Blocks created within a name scope will have its parent Block’s prefix prepended to its name.

Let’s demonstrate this by first defining a simple neural net:

class Model(gluon.Block):
    def __init__(self, **kwargs):
        super(Model, self).__init__(**kwargs)
        with self.name_scope():
            self.dense0 = gluon.nn.Dense(20)
            self.dense1 = gluon.nn.Dense(20)
            self.mydense = gluon.nn.Dense(20, prefix='mydense_')

    def forward(self, x):
        x = mx.nd.relu(self.dense0(x))
        x = mx.nd.relu(self.dense1(x))
        return mx.nd.relu(self.mydense(x))

Now let’s instantiate our neural net.

  • Note that model0.dense0 is named as model0_dense0_ instead of dense0_.
  • Also note that although we specified mydense_ as prefix for model.mydense, its parent’s prefix is automatically prepended to generate the prefix model0_mydense_.
model0 = Model()
model0.initialize()
model0(mx.nd.zeros((1, 20)))
print(model0.prefix)
print(model0.dense0.prefix)
print(model0.dense1.prefix)
print(model0.mydense.prefix)
model0_
model0_dense0_
model0_dense1_
model0_mydense_

If we instantiate Model again, it will be given a different name like shown before for Dense.

  • Note that model1.dense0 is still named as dense0_ instead of dense2_, following dense layers in previously created model0. This is because each instance of model’s name scope is independent of each other.
model1 = Model()
print(model1.prefix)
print(model1.dense0.prefix)
print(model1.dense1.prefix)
print(model1.mydense.prefix)
model1_
model1_dense0_
model1_dense1_
model1_mydense_

It is recommended that you manually specify a prefix for the top level Block, i.e. model = Model(prefix='mymodel_'), to avoid potential confusions in naming.

The same principle also applies to container blocks like Sequential. name_scope can be used inside __init__ as well as out side of __init__:

net = gluon.nn.Sequential()
with net.name_scope():
    net.add(gluon.nn.Dense(20))
    net.add(gluon.nn.Dense(20))
print(net.prefix)
print(net[0].prefix)
print(net[1].prefix)
sequential0_
sequential0_dense0_
sequential0_dense1_

gluon.model_zoo also behaves similarly:

net = gluon.nn.Sequential()
with net.name_scope():
    net.add(gluon.model_zoo.vision.alexnet(pretrained=True))
    net.add(gluon.model_zoo.vision.alexnet(pretrained=True))
print(net.prefix, net[0].prefix, net[1].prefix)
sequential1_ sequential1_alexnet0_ sequential1_alexnet1_

Saving and loading

Because model0 and model1 have different prefixes, their parameters also have different names:

print(model0.collect_params(), '\n')
print(model1.collect_params())
model0_ (
  Parameter model0_dense0_weight (shape=(20L, 20L), dtype=<type 'numpy.float32'>)
  Parameter model0_dense0_bias (shape=(20L,), dtype=<type 'numpy.float32'>)
  Parameter model0_dense1_weight (shape=(20L, 20L), dtype=<type 'numpy.float32'>)
  Parameter model0_dense1_bias (shape=(20L,), dtype=<type 'numpy.float32'>)
  Parameter model0_mydense_weight (shape=(20L, 20L), dtype=<type 'numpy.float32'>)
  Parameter model0_mydense_bias (shape=(20L,), dtype=<type 'numpy.float32'>)
) 

model1_ (
  Parameter model1_dense0_weight (shape=(20, 0), dtype=<type 'numpy.float32'>)
  Parameter model1_dense0_bias (shape=(20,), dtype=<type 'numpy.float32'>)
  Parameter model1_dense1_weight (shape=(20, 0), dtype=<type 'numpy.float32'>)
  Parameter model1_dense1_bias (shape=(20,), dtype=<type 'numpy.float32'>)
  Parameter model1_mydense_weight (shape=(20, 0), dtype=<type 'numpy.float32'>)
  Parameter model1_mydense_bias (shape=(20,), dtype=<type 'numpy.float32'>)
)

As a result, if you try to save parameters from model0 and load it with model1, you’ll get an error due to unmatching names:

model0.collect_params().save('model.params')
try:
    model1.collect_params().load('model.params', mx.cpu())
except Exception as e:
    print(e)
Parameter 'model1_dense0_weight' is missing in file 'model.params', which contains parameters: 'model0_mydense_weight', 'model0_dense1_bias', 'model0_dense1_weight', 'model0_dense0_weight', 'model0_dense0_bias', 'model0_mydense_bias'. Please make sure source and target networks have the same prefix.

To solve this problem, we use save_parameters/load_parameters instead of collect_params and save/load. save_parameters uses model structure, instead of parameter name, to match parameters.

model0.save_parameters('model.params')
model1.load_parameters('model.params')
print(mx.nd.load('model.params').keys())
['dense0.bias', 'mydense.bias', 'dense1.bias', 'dense1.weight', 'dense0.weight', 'mydense.weight']

Replacing Blocks from networks and fine-tuning

Sometimes you may want to load a pretrained model, and replace certain Blocks in it for fine-tuning.

For example, the alexnet in model zoo has 1000 output dimensions, but maybe you only have 100 classes in your application.

To see how to do this, we first load a pretrained AlexNet.

  • In Gluon model zoo, all image classification models follow the format where the feature extraction layers are named features while the output layer is named output.
  • Note that the output layer is a dense block with 1000 dimension outputs.
alexnet = gluon.model_zoo.vision.alexnet(pretrained=True)
print(alexnet.output)
print(alexnet.output.prefix)
Dense(4096 -> 1000, linear)
alexnet0_dense2_

To change the output to 100 dimension, we replace it with a new block.

with alexnet.name_scope():
    alexnet.output = gluon.nn.Dense(100)
alexnet.output.initialize()
print(alexnet.output)
print(alexnet.output.prefix)
Dense(None -> 100, linear)
alexnet0_dense3_