Active Transfer Learning

CS772 (Probabilistic Machine Learning)

• Active Learning with Cross-Class Similarity Transfer - Guo, Yuchen, et al. "Active Learning with Cross-Class Similarity Transfer." AAAI. 2017.

• pretrained_model.py: Run a specified pretrained model on a specified dataset and generate feature vectors of dataset and store in a pickle file as <dataset>_<model>_features.pkl

python pretrained_model.py --dataset <dataset_name> --model <model_name> --cuda --ngpu <int>

Use -h option for more help.

• class_similarity.py: Compute similarity between classes using Word2Vec model trained on GoogleNews-vectors and store in a pickle file. Requires GoogleNews-vectors-negative300.bin in the same directory and specify classes inside the code

python compute_similarity.py

• active_transfer_learning_parallel.py: Run the Active Transfer Learning algorithm parallely on specified number of CPU cores. Set hyper-paramters manually in the code. Generate plots inside the plots folder with name as <dset>_<model_name>_atl.jpeg. Requires dataset file (feature vectors in pickle file) and class similarity matrix.

python active_transfer_learning_parallel.py
          -d, --dset <Path to dataset> (required)
          -g, --G <Path to class similarity matrix> (required)
          -m, --model <Model used to construct feature vectors> (default=alexnet)
          -l, --nlabels <Number of labels or classes in dataset> (default=10)
          -w, --workers <Number of CPU cores> (default=1)
          -s, --sigma <Sigma for heat kernel similarity> (default=0.0 and if not specified then will be calculated by code itself)

Use -h option for more help.

Requirements

• PyTorch
• Gensim
• Python-3.6 with packages: numpy, cvxopt, gensim