MXNet Cookbook

About

The Cookbook.md and code examples are my personal notes for how to do things in MXNet. The repo is intended to be a "living" document that evolves over time. When possible, I try to have one-to-one code comparisons between imperative programming with Gluon and symbolic programming with Module.

Note: the symbol API is deprecated in mxnet. From now on, this repo will focus on Gluon and hybridize methods.

Custom Loss

Shows how to implement a custom loss function and custom metrics function.

• Module
  • MakeLoss needs to be written in symbol api
  • MakeLoss returns the gradient of the loss function. For the metrics you need the forward output of the network, and therefore you need to write an output that blocks the gradient and add it to the sym.Group() which contains the list of network outputs.
  • CustomMetric expects float values, not symbols. The function needs to be written with numpy arrays.

DCGAN (faces)

This consists of a simple Deep Convolutional Generative Adversarial Network (dcgan), implemented in Gluon. As training data, it uses the aligned images of the Labeled Faces in the Wild dataset.

Generative Adversarial Network (GAN)

In this folder are some basic and more advanced GAN implementations.

• gan_gluon.py Basic GAN where both the Generator and Discriminator use fully connected layers.
• dcgan_gluon.py Deep convolutional GAN. Note that the initialization values are important for stable training and good resuls. However, this version has no protection against mode collapse.
• wgan_gluon.py Deep convolutional GAN with Wasserstein Loss function and 1-Lipschitz Continuity Enforcement, to prevent mode collapse. (work in progress)

All files have support for training on multiple GPUs and plot the output to both console/matplotlib as well to TensorBoard.

Image Classifier

Code comparison between Module and Gluon for a simple image classifier, using a convolutional neural network. Some highlighted topics are how to load data to the GPU and run training of the model on GPU.

• mnist_module.py Module implementation for CPU or single GPU.
• mnist_gluon.py Gluon implementation for CPU or single GPU.
• mnist_gluon_mxboard.py Hybridized CPU or multi GPU implementation with TensorBoard writer.

Measurements are performed on the MNIST Fashion Image Classifier mnist_gluon_mxboard.py. Note that the model has been hybridized before training. The used system is a 8 core Intel i9-9900K @ 3.60GHz and 2 NVidia GeForce GTX 1070 Ti with 8Gb:

device	time per epoch (s)	total training time (s)
cpu	162.6	1716.7
gpu x 1	6.1	67.1
gpu x 2	3.7	41.1

Linear Regression 1D

Exploration of the autograd libraries and code comparision with a numpy implementation from scratch. This folder has multiple implementations of the linear regression, implemented from scratch. To make sure the numerical outputs are (almost) the same accross different implementations, al inputs and initializations have been fixed.

• lr_1d_numpy.py 1D Linear regression from scratch with implementation of the gradient for the MSE loss function with respect to W.

• lr_1d_gluon.py 1D Linear regression from scratch, using gluon autograd.

• lr_1d_torch.py 1D Linear Regression from scratch, using torch autograd.

• lr_1d_gluon_optim.py 1D Linear Regression, using build in loss function, optimizer and autograd.

• lr_1d_torch_optim.py 1D Linear Regression, using build in loss function, optimizer and autograd.

Linear Regression

Code comparison between Module and Gluon for a simple linear regression problem. There are 2 versions for the Module API. The file linear_regression_module.py uses the high-level fit() function. The file linear_regression_module_intermediate.py does the steps of fit() more verbose.

Pix2Pix

This is an implementation from the method described in the paper: "Image-to-Image Translation with Conditional Adversarial Networks", Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros. This script is based on the Pixel2Pixel example from Gluon, the straight dope New features are:

• Multiple GPU support
• New feature based loss function resulting in better and cleaner generated images.

Transfer Learning

Shows how to do transfer learning and combined transfer learning with fine tuning. Highlighted topics are how to replace the final layer(s) and how to freeze the feature layers, if needed.

Todo:

• Fix batch normalization.
• Freeze first layer in code.