Visualizing the learned representations is extremely important in deep learning for interpretations. Additionally, tracing back to the original inputs from the features will provide a better understanding of the learned representations. In order to efficiently conduct this task and support research, I created a simple interactive tool using streamlit and bokeh.
The input file should be in .h5 format with following data:
- 'images' : Original Unnormalized images
- 'labels' : Labels (should be integers)
- 'tsne' : Array of the features after applying tsne
The following code block can be used as guide to save the files:
import h5py
hf = h5py.File("~/data.h5", 'w')
hf.create_dataset('images', data= "image data")
hf.create_dataset('labels', data= "label data")
hf.create_dataset('tsne', data= "tsne data")
hf.create_dataset('id', data='ID for each data')
hf.close()
pip install -r requirements.txt
streamlit run ~/visualizer.py --server.maxUploadSize 1000