ergo

Keras is deprecated and tensorflow bumped to version 2.

Fixes

• 8f5c0ab fix: using model outputs to compute number of output classes
• a5d7ede fix: typo when checking non flat files
• 9654c03 fix: typo preventing the check of the validation set
• b8a7da0 fix: added .idea folder to .gitignore

New Features

• categorical_crossentropy is now used as default metric (allows binary and multiclass)

Fixes

• ergo serve can now read a json post request as input

New Features

• clustering algorithms run in parallel
• show error when no clusters are adjusted
• inertia analysis added for kmeans clustering
• 3D projection of clustering and PCA analysis
• number of clusters can be specified from as argument or infered from number of classes
• explore now performs clustering analysis
• --no-tests flag added to action to_fdeep
• ergo serve will now autoload classes.json if available
• unnormalized features are shown
• ergo info now reports sklearn version
• explore graphs and tables have individual options
• table with stats for each feature added to ergo explore
• feature correlation matrix added to ergo explore
• pca explained variance chart
• pca decomposition showed in ergo explore
• explore action added
• added calculation of correlation for each predictor with target
• new noisy data label added to dbscan plots
• new dbscan clustering algorithm added

Fixes

• skipping repeated correlation fixed
• typo in explore help menu
• redundant action removed
• test for pkl files when cheking existing datasets
• fixed requirements.txt ergo -> ergo-ai
• fixed an issue caused by newer sklearn.classification_report versions breaking backwards compatibility (fixes #5)
• added seaborn to requirements.txt
• better logs

Misc

• Readme updated
• Refactoring

New Features

• ergo info has now an optional -j/--to-json argument to print info in json format.
• ergo relevance now has a -w/--workers optional argument to run the algorithm with concurrent workers by using more memory.
• help menus now report default values for each argument

Fixes

• fixed setup.py to install dependencies (fixes #4)

New Features

• ergo create now has the new --inputs, --outputs, --layers, --epochs and --batch-size optional arguments.
• ergo serve has now an optional --profile argument to enable the profiling middleware.
• ergo relevance has now a --metric argument (default to precision) to specify which metric to use for relevance.
• ergo view now has an optional --img-only argument to avoid showing any UI.
• ergo view will now plot the ROC curve.
• ergo view will now generate the trainig_cm.png, test_cm.png, validation_cm.png and history.png files.
• ergo serve now supports GET queries, POST forms and also file uploads as x.
• ergo serve will now expose an /encoding API route to perform encoding only.

Fixes

• fixed a typo in ergo cmp.

Misc

• refactored ergo optimize-dataset, ergo clean, ergo create and ergo view code.
• removed sumpy connector to decouple the projects.

New Features

• ergo serve will now return class names (customizable via the --classes argument) with their probabilities as a response.
• new ergo encode action to process raw samples.
• added requirements.txt and .gitignore to the project templates.

Fixes

• ergo serve now processes X correctly.

New Features

• ergo view now shows the confusion matrices
• ergo train now generates both stats.txt and stats.json
• ergo train now has a new --no-shuffle option
• ergo relevance now has a new --to-json optional argument to save the results in a json file.
• ergo cmp now supports models with different input shapes
• each action has now its own help menu woth ergo <action> -h

Fixes

• training history is now correctly built taking into account past training processes
• renamed sumpy requirement to sum-py

New Features

• added conversion to tensorflow script as the to-tf <path> action.

Fixes

• fixed a bug which prevented ergo to load pre generated datasets.

New Features

• implemented sum database integration

Fixes

• number of evals/s is now correctly reported as number of single sample evaluations

Misc

• renamed ergo-nn to ergo-ai

New Features

• ergo now supports both scalar and multi dimensional inputs, as well as functional models.
• implemented relevance algorithm to compute how each input attribute affects the model accuracy.
• ergo can now be installed by sudo pip3 install ergo-nn

Misc

• to_fdeep conversion logic updated to support the latest changes in the library.

New Features

• added algorithm to compute relevance of each attribute by differential training
• addded --no-save argument for train action to skip dataset save on disk
• datasets saving is now performed asynchronously in order to start training sooner

Fixes

• increased to_csv chunksize parameter to speedup datasets saving

New Features

• printing model stats after training
• added colors to the cmp action and a new --to-json output argument
• new asciiart logo

New Features

• new cmp action to compare models

New Features

• prepare.py template now allows to preprocess inputs for evaluation

Fixes

• better model.stats file layout
• refactored dataset load and save logic and made both async