examples

Tensorpack Examples

Training examples with reproducible performance.

The word "reproduce" should always mean reproduce performance. With the magic of SGD, wrong deep learning code often appears to work, especially if you try it on toy datasets. Github is full of deep learning code that "implements" but does not "reproduce" methods, and you'll not know whether the implementation is actually correct. See Unawareness of Deep Learning Mistakes.

We refuse toy examples. Instead of showing tiny CNNs trained on MNIST/Cifar10, we provide training scripts that reproduce well-known papers.

We refuse low-quality implementations. Unlike most open source repos which only implement methods, Tensorpack examples faithfully reproduce experiments and performance in the paper, so you're confident that they are correct.

Getting Started:

These are the only toy examples in tensorpack. They are supposed to be just demos.

• An illustrative MNIST example with explanation of the framework
• Tensorpack supports any symbolic libraries. See the same MNIST example written with tf.layers, and with weights visualizations
• A tiny Cifar ConvNet and SVHN ConvNet
• A boilerplate file to start with, for your own tasks

Vision:

Name	Performance
Train ResNet, ShuffleNet and other models on ImageNet	reproduce 10 papers
Train Mask/Faster R-CNN on COCO	reproduce 7 papers
Unsupervised learning with Momentum Contrast (MoCo)	reproduce 2 papers
Generative Adversarial Network(GAN) variants, including DCGAN, InfoGAN, Conditional GAN, WGAN, BEGAN, DiscoGAN, Image to Image, CycleGAN	visually reproduce 8 papers
DoReFa-Net: training binary / low-bitwidth CNN on ImageNet	reproduce 4 papers
Adversarial training with state-of-the-art robustness	official code for the paper
Fully-convolutional Network for Holistically-Nested Edge Detection(HED)	visually reproduce
Spatial Transformer Networks on MNIST addition	reproduce the paper
Visualize CNN saliency maps	visually reproduce
Similarity learning on MNIST
Single-image super-resolution using EnhanceNet
Learn steering filters with Dynamic Filter Networks	visually reproduce
Load a pre-trained AlexNet, VGG, or Convolutional Pose Machines
Load a pre-trained FlowNet2-S, FlowNet2-C, FlowNet2

Reinforcement Learning:

Name	Performance
Deep Q-Network(DQN) variants on Atari games, including DQN, DoubleDQN, DuelingDQN.	reproduce the paper
Asynchronous Advantage Actor-Critic(A3C) on Atari games	reproduce the paper

Speech / NLP:

Name	Performance
LSTM-CTC for speech recognition	reproduce the paper
char-rnn for fun	fun
LSTM language model on PennTreebank	reproduce reference code