tencent-ml-images

Why we included this project

Image-classification projects often stall at the data stage, and this is one of the few openly published multi-label datasets at real scale: roughly 17.6 million image URLs annotated across up to 11,166 categories, collected from ImageNet and Open Images. Alongside the data it ships a ResNet-101 model pre-trained on ML-Images that reaches about 80.7% top-1 accuracy on ImageNet after fine-tuning, a practical transfer-learning starting point for teams that would rather not train from scratch. Researchers and engineers building classifiers or visual feature extractors get a full pipeline with it, including scripts to download the images, build TFRecord files, pretrain and fine-tune, and pull features from the released checkpoints. One caveat: the code targets Python 2 and TensorFlow 1.x, so expect to port the scripts before running them in a modern environment.