Deep learning networks

This repo is used to research convolutional networks primarily for computer vision tasks. For this purpose, the repo contains (re)implementations of various classification, segmentation, detection, and pose estimation models and scripts for training/evaluating/converting.

The following frameworks are used:

• MXNet/Gluon (info),
• PyTorch (info),
• Chainer (info),
• Keras (info),
• TensorFlow 1.x/2.x (info).

For each supported framework, there is a PIP-package containing pure models without auxiliary scripts. List of packages:

• gluoncv2 for Gluon,
• pytorchcv for PyTorch,
• chainercv2 for Chainer,
• kerascv for Keras,
• tensorflowcv for TensorFlow 1.x,
• tf2cv for TensorFlow 2.x.

Currently, models are mostly implemented on Gluon and then ported to other frameworks. Some models are pretrained on ImageNet-1K, CIFAR-10/100, SVHN, CUB-200-2011, Pascal VOC2012, ADE20K, Cityscapes, and COCO datasets. All pretrained weights are loaded automatically during use. See examples of such automatic loading of weights in the corresponding sections of the documentation dedicated to a particular package:

• Gluon models,
• PyTorch models,
• Chainer models,
• Keras models,
• TensorFlow 1.x models,
• TensorFlow 2.x models.

Installation

To use training/evaluating scripts as well as all models, you need to clone the repository and install dependencies:

git clone git@github.com:osmr/imgclsmob.git
pip install -r requirements.txt

Table of implemented classification models

Some remarks:

• Repo is an author repository, if it exists.
• a, b, c, d, and e means the implementation of a model for ImageNet-1K, CIFAR-10, CIFAR-100, SVHN, and CUB-200-2011, respectively.
• A, B, C, D, and E means having a pre-trained model for corresponding datasets.

Table of implemented segmentation models

Some remarks:

• a/A corresponds to Pascal VOC2012.
• b/B corresponds to ADE20K.
• c/C corresponds to Cityscapes.
• d/D corresponds to COCO.
• e/E corresponds to CelebAMask-HQ.

Table of implemented object detection models

Some remarks:

• a/A corresponds to COCO.

Table of implemented human pose estimation models

Some remarks:

• a/A corresponds to COCO.

Table of implemented automatic speech recognition models

Some remarks:

• a/A corresponds to LibriSpeech.
• b/B corresponds to Mozilla Common Voice.