HarshSulakhe/NNFL_OSL

Code for Siamese Networks for One Shot Recognition, as a part of our coursework in BITS F312 - Neural Networks and Fuzzy Logic

Instructions to Run

# download the dataset
git clone https://github.com/brendenlake/omniglot.git
cd omniglot/python
unzip images_evaluation.zip
unzip images_background.zip
cd ../..


# Now create a directory to save models
mkdir models


# Train the model using
python3 train.py --train_path omniglot/python/images_background \
                 --test_path  omniglot/python/images_evaluation \
                 --gpu_ids 0 \
                 --model_path models


# Evaluate the model using
python3 evaluate.py

#Plot graph of loss using
python3 plot_loss.py

#Save examples of affine transformations using
python3 affine.py

File Descriptions

• train.py - Consists of the main training code
• mydataset.py - Consists of custom datasets used for training and testing
• plot_loss.py - Code to plot and save the graph for training loss v/s iterations
• affine.py - Code to create and save affine transformations of custom images, with and without random sampling.
• model.py - Consists of the main model architecture of the convolutional siamese twins.
• evaluate.py - Code to evaluate the pre-trained model on 20-way one shot trials.

Some Results

Training loss

Example of affine transformations