/ImageNet-edge-cases

Keras implementation to test pre-trained models on ImageNet, as well as interactively show and explore failure cases.

Primary LanguageJupyter NotebookMIT LicenseMIT

ImageNet Edge Cases

Keras implementation to test pre-trained models on ImageNet, as well as interactively show and explore failure cases.
Li Ding
Oct. 20, 2017

Overview

  • imgnet-val-resnet50.py - load the pre-trained ResNet-50 and evaluate on ILSVRC'12 validation set, save the prediction result.
  • imgnet-val-explore.ipynb - use the prediction to interactively explore failure cases in Jupyter Notebook.

Quick Start

  • dependencies: Python 2.7 + tensorflow, keras, xmltodict, Queue, threading, tqdm, cv2
  • set data path: enter a folder path in both files, e.g. '/imgnet' that has following structure: 
    /imgnet  
.../ILSVRC2012_img_val/ (ILSVRC'12 validation images, 50,000 items, totalling 6.7GB)  
.../val/ (ILSVRC'12 validation annotations, 50,000 .xml files, totalling 31.2 MB)  
.../val_prob.csv (optional, the prediction created by imgnet-val-resnet50.py)
  • run imgnet-val-resnet50.py to obtain the result val_prob.csv.
  • in terminal, jupyter notebook to start jupyter, load imgnet-val-explore.ipynb, run and explore.

alt text