Awesome

Interactive Multi-Label CNN Learning with Partial Labels

Overview

This repository contains the implementation of Interactive Multi-Label CNN Learning with Partial Labels.

In this work, we address efficient end-to-end learning a multi-label CNN classifier with partial labels using an interactive dependency learning scheme.

Prerequisites

• Python 3.x
• Tensorflow 1.x.x
• sklearn
• matplotlib
• skimage
• scipy

Data Preparation

Open Images

1. Please download pretrained Open Images model(https://storage.googleapis.com/openimages/2017_07/oidv2-resnet_v1_101.ckpt.tar.gz) into './model/resnet' folder

2. Please download Open Images urls and annotation into ./data/OpenImages folder according to the instructions within the folder ./data/OpenImages/2017_11.

3. To crawl images from the web, please run the script:

python ./download_imgs/asyn_image_downloader.py 					#`data_set` == `train`: download images into `./image_data/OpenImages/train/`
python ./download_imgs/asyn_image_downloader.py 					#`data_set` == `validation`: download images into `./image_data/OpenImages/validation/`
python ./download_imgs/asyn_image_downloader.py 					#`data_set` == `test`: download images into `./image_data/OpenImages/test/`

Please change the data_set variable in the script to train, validation, and test to download different data splits.

4. To extract features into TensorFlow storage format, please run:

python ./extract_data/extract_feature_2_TFRecords_OpenImages.py						#`data_set` == `train`: create `./TFRecords/train_feature.tfrecords`
python ./extract_data/extract_feature_2_TFRecords_OpenImages.py						#`data_set` == `validation`: create `./TFRecords/validation_feature.tfrecords`
python ./extract_data/extract_feature_2_TFRecords_OpenImages.py			        		#`data_set` == `test`:  create `./TFRecords/test_feature.tfrecords`

Please change the data_set variable in the extract_feature_2_TFRecords_OpenImages.py script to train, and validation to extract features from different data splits.

5. Please download:

• Sparse dictionary (https://drive.google.com/file/d/1he4omIq8N6SEuysFMNFeMkUlS7xpFIYE/view?usp=sharing) into the folder ./TFRecord/
• Label graph (https://drive.google.com/file/d/1kg0UeInk33cKadq8rEU3i-3wSm3Oq7Oh/view?usp=sharing) into the folder ./label_graph/
• Logistic model (https://drive.google.com/file/d/1our5lctu-H567VjX20X9TSgiGINIAPzz/view?usp=sharing) into the folder ./result/

CUB

1. Please download the following data files (https://drive.google.com/file/d/1gBQ_PQ0U8kzCaiiF7CvG92f1Ssfk8Zgq/view?usp=sharing), (https://drive.google.com/file/d/1fiNtiBj3hCj75eLHN1-02yWSDZrJ7GeN/view?usp=sharing), (https://drive.google.com/file/d/1O-0HTTFE9QpdTSQ8fdg31PsPAzJEiJga/view?usp=sharing) into ./TFRecord/ folder

MSCOCO

1. Please download MSCOCO images and annotation into ./image_data/MSCOCO folder according to the instructions within the folders./image_data/MSCOCO/train2014,./image_data/MSCOCO/val2014,./image_data/MSCOCO/annotation, and ./data/MSCOCO_1k.

2. To extract features into TensorFlow storage format, please run:

python ./extract_data/extract_train_img_2_TFRecords_MSCOCO.py						#create ./TFRecord/train_MSCOCO_img_ZLIB.tfrecords
python ./extract_data/extract_test_img_2_TFRecords_MSCOCO.py						#create ./TFRecord/test_MSCOCO_img_ZLIB.tfrecords
python ./extract_data/extract_dic_img_2_TFRecords_MSCOCO.py							#create ./TFRecord/dic_10_MSCOCO_img_ZLIB.tfrecords

Training and Evaluation

Open Images

1. To pretrain the logistic backbone network, please run the script:

python ./OpenImages_experiments/baseline_logistic_OpenImages.py					# fixed feature representation
python ./OpenImages_experiments/e2e_baseline_logistic_OpenImages.py				# end-to-end training

2. To train our method, please run the script:

python ./OpenImages_experiments/interactive_learning_OpenImages.py				# fixed feature representation
python ./OpenImages_experiments/e2e_interactive_learning_OpenImages.py			# end-to-end training

3. To evaluate the performance, please run the script:

python ./evaluation/evaluation_interactive_learning_OpenImages.py				# fixed feature representation
python ./evaluation/evaluation_e2e_interactive_learning_OpenImages.py			# end-to-end training

CUB

1. Please download the ImageNet ResNet backbone (http://download.tensorflow.org/models/resnet_v2_101_2017_04_14.tar.gz) into ./model/resnet_CUB

2. To pretrain the logistic backbone network, please run the script:

python ./CUB_experiments/e2e_baseline_logistic_CUB.py

3. To train and evaluate our method, please run the script:

python ./CUB_experiments/e2e_interactive_learning_CUB.py

MSCOCO

1. Please download the ImageNet VGG backbone (http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz) into ./model/vgg_ImageNet

2. To pretrain the logistic backbone network, please run the script:

python ./MSCOCO_experiments/e2e_baseline_logistic_MSCOCO_vgg.py

3. To train and evaluate our method, please run the script:

python ./MSCOCO_experiments/e2e_interactive_learning_MSCOCO.py

Pretrained Models

As the pretrained models are implemented with an older tensorflow version, it might not work with the current code-base.

• If you need the pretrained models, please reach out to me via huynh.dat@northeastern.edu

Citation

If this code is helpful for your research, we would appreciate if you cite the work:

@article{Huynh-mll:CVPR20,
  author = {D.~Huynh and E.~Elhamifar},
  title = {Interactive Multi-Label {CNN} Learning with Partial Labels},
  journal = {{IEEE} Conference on Computer Vision and Pattern Recognition},
  year = {2020}}