Quick tip: converting Gluon models to symbolic format

Source: Deep Learning on Medium

As discussed earlier, Gluon is an imperative API available on top of the symbolic API implemented by Apache MXNet.

One of the cool features of Gluon is the extensive collection of pre-trained computer vision models available in Gluon CV. Using these models with the Gluon API is extremely easy, but sometimes we’d rather use the symbolic API instead.

One reason could be language support: Gluon is Python only whereas MXNet supports a whole list of languages including C++, which you may need to get the best prediction performance possible. Unfortunately, the MXNet model zoo is not synchronized with the Gluon model zoo, so you can’t just grab the same models.

One easy solution to this problem is to use the Gluon API to download models, export them to symbolic format and then load them using the MXNet API.

It goes like this.

from gluoncv import model_zoo
import mxnet as mx
import numpy as np
# Download the model from the Gluon model zoo
# You'll find it in ~/.mxnet/models
net = model_zoo.get_model('resnet50_v1', pretrained=True)
# Convert the model to symbolic format
# Build a fake image to run a single prediction
# This is required to initialize the model properly
x = np.zeros([1,3,224,244])
x = mx.nd.array(x)
# Predict the fake image
# Export the model

This will export the model weights and the JSON file containing the symbolic definition of the model.

$ ls -1 resnet50*

Now you can load this model as usual with the model.load_checkpoint() API. Just make sure you use a recent version of MXNet, as Gluon models could be incompatible with older ones. YMMV.

That’s it. Now you can enjoy all these models with the symbolic API :)