We will be trying different models for performing face recognition.
Attempting to build a model to perform face recognition. Taking Siamese Network for One Shot Image Verification paper's approach and applying it with a small face dataset.
In the first attempt the best validation accuracy lies around 0.225 and evaluation accuracy is 0.125 for the one shot tasks.
The dataset is small and FaceLoader is not very effective at the moment making the training very slow and not very accurate.
Modified Face Loader to preload every 100 batches, where to create the full dataset for the 200 batch it randomly picks 100 persons (in our case it will repeat 8 different person randomly)
Afterwards for each person it form pairs for one batch of data, half of it randomly paired with an image of itself and another half randomly picking from a different class.
At the end the full preloaded set is shuffled so not necessarily batch at the end contains the same split.
The model was trained like before which led to higher validation accruacy (0.4) but seemed to decrease the evalution accuracy to 0.1
Instead of having a grayscale single channel input modified model and loader to take rgb images. Also modified laoder to take data from hdf5 file instead which improved image read time significantly. Preoloading a large batches allows to utilize the GPU shortly more but doesnt really save a lot of time as larger the preload longer the wait.
Trained a model where best validation accuracy was 0.3 evaluation accracy was 0.325
Ran the original source code with normal SGD instead of the modified one, where the final mean global accuracy was around 0.6475. The model had around 0.125 accuracy for face recognition tasks.
Retraining that model on the face dataset helped achieve 0.4 mean evalulation accuracy in the one shot face recognition task.
Following the paper and initialized the weights and bias for the layers
using tf.random_normal_initializer. Training a model from scratch got around 0.4 evaluation accuracy. Pretraining a model and then training again using the face dataset had similar performance.
- Paper: https://www.cs.cmu.edu/~rsalakhu/papers/oneshot1.pdf
- Code taken and modified further from: https://github.com/tensorfreitas/Siamese-Networks-for-One-Shot-Learning
- Small section of data collected from: https://drive.google.com/drive/folders/0B5G8pYUQMNZnLTBVaENWUWdzR0E
- Trained Models (Face Recognition): https://drive.google.com/drive/folders/17dw7C6rhIUWAKpGcbRabJMKJP7P3ze-I?usp=sharing
- Trained Models (Omniglot Dataset): https://drive.google.com/drive/folders/1MYU2HJgV5LjqaoeeN_Eq8fLmBJ2FDoJ2?usp=sharing