DeepImageRanking

Image Similarity using Deep Ranking

Documentation

1. For this particular implementation of Deep Ranking we are using TinyImageNet as our training dataset. Download the TinyImageNet dataset here

2. Change the folder structure of the validation set by running:

   python transform_validation.py

   This will aid later when generating the image embeddings

3. Now to sample triplets we run:

   python triplet_sampler.py --training 1 --num_pos_images 1 --num_neg_images 1

   Our command-line arugments are as follows:

   • training: 0 or 1 to sample from train or validation set
   • num_pos_images: the number of positive images per query image
   • num_neg_images: the number of negative images per query image

   This should generate training_triplet_sample.csv or val_triplet_sample.csv based on the argument provided.

4. To implement we use a pre-trained ResNet-101 with two sub-sampling layers. We finetune the earlier layers of the ResNet and train the rest. The PyTorch implementation can be found in the file deep_rank_net.py

5. To train the network run:

    python train_net.py --epochs 1 --optim sgd

   Our command-line arugments are as follows:

   • epochs: the number of training epochs to run
   • optim: thee optimizer to use, can be sgd or adam or rms. Will default to sgd.

   Other hyperparameters are set inside the file. These include BATCH_SIZE=25 and LEARNING_RATE=0.001 This should generate a fully trained model file called deepranknet.model and other intermediate files called temp_*.model

6. To generate embeddings:

   python gen_emebeddings.py

   This will generate the train and test embeddings in the files train_embedding.txt and test_embedding.txt respectively.

7. To evaluate the model run:

   python eval.py

8. If you wish to change the architecture, change the hyperparameters or understand the implementation further see the report file

Presentation & Report

• Presentation
• Report

Further Resources:

• Deep Ranking paper
• Project proposal