Building deep neural networks using keras and tensorflow for various problems.
-
ASL:
- Image classification of sign language letters, dataset from https://www.kaggle.com/grassknoted/asl-alphabet.
- Train data - 2500 images for every class.
- Validation data - 500 images for every class.
- Used image augmentation and got around 97% accuracy with my cnn.
- With InceptionResNetV2() got 99.5%.
-
IMDb faces:
- Image classification of actors and actresses based on gender. Around 94% accuracy.
-
Rnn:
- LSTM network for generating new C code.
- Used limited number of C files from linux-kernel.
- After 50 epochs loss came down to 0.73.
-
neural_style_transfer:
- Algorithm for generating image in style of another image.