Skip to content

Xuhong Li, Yves Grandvalet, and Franck Davoine. "Explicit Inductive Bias for Transfer Learning with Convolutional Networks." In ICML 2018.

License

Notifications You must be signed in to change notification settings

holyseven/TransferLearningClassification

Repository files navigation

TransferLearningClassification

Steps of reproducing the experiments:

  1. Download resnet_v1_101.ckpt from tensorflow slim with the script ./z_pretrained_weights/download_resnet_v1_101.sh.

  2. (optional) Verify the pre-trained weights, using ./network_verification/resnet_verification.py. ImageNet validation set is needed. This can be also used for verifying the lifelong learning setting. After training on other databases, use the finetuned model (for shared variables) and the pre-trained model (for the last classifier layer) to do the evaluation on ImageNet.

  3. a. Download tfRecords files (link).

    b. Donwload images and create tfRecords files with scripts in ./create_databases.

    Stanford Dogs 120: link. MIT Indoors 67: link. Caltech 256: link.

    c. Create new databases and add interface in ./database/dataset_reader.py.

  4. run the script ./run_classification/train.sh.

On the database of Stanford Dogs, L2-SP should have around 88~89% classification accuracy while L2 81~82%. Tested with Tensorflow 1.4, Python 2.7.

About

Xuhong Li, Yves Grandvalet, and Franck Davoine. "Explicit Inductive Bias for Transfer Learning with Convolutional Networks." In ICML 2018.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published