SCattering Ai aNalysis
This is a project that classifies the shape of nanoobject based on the intensity variation.
The software makes it possible to:
- train/improve the predictions models
- use trained model to perform predictions
There is a build-in help functionality.
Usage
Training
The app comes pre-trained on the included data. Training is only needed if new data is added.
The app responsible for training is train.py.
The data for training is provided in the cvs format and is stored in the data directory. Training data for a given shape is provided in a file with a name corresponding to that shape. In order to add the capability of recognizing a new shape, just add a properly named file with shape examples, the system will recognize this as a new shape.
To see the available training options please consult the current help (python train.py --help). It will always provide the most recent options. There are some notable features to know of:
- the command line switches can be combined, you may e.g. choose an arbitrary collection of prediction algorithms to be included
- if you select stacked switch then a variant with all algorithms stacked will be added
- you may decide to run PCA on input variables. PCA can be wither specified as a number of components (values from 1 up) or percentage of explained variance (0..1)
- you may decided to train a model on a continous subset of columns from the training set (for example just columns from 100 to 300)
- the trained models are stored in the models directory. They can be identified by specifying a tag. If you do not specify a tag, a default model will be replaced.
- there is an integration with Neptune.ai provided (upload option). If you wish to use it, specify NEPTUNE_API_TOKEN and NEPTUNE_PROJECT_NAME in the shell
Predictions
There are 3 ways to access the prediction models:
- command line - full access to all prediction models
- web interface - only default model available
The command line access is provided by predict.py. All available models are accessible this way. Please consult help (python predict.py--help) to find out the usage syntax. If you wish to you may package your commandline as docker image, the relevant Dockerfile is provided.
The web interface is based on the Streamlit integration and requires streamlit. To access it locally please run 'streamlit run onlinepredict.py'
Notes
SessionState used in the webinterface is based on the following gist: https://gist.github.com/tvst/036da038ab3e999a64497f42de966a92