New Features

• User controlled SFrame pickling functionality (Python 3 only) #3327

Bug Fixes

• Drawing Classifer: automaticaly resize images #3313
• Decrease memory usage for ROC curve calculation #3323
• Remove tc.config.get_server_log_location() #3329
• Fixed tc.version_info.git_sha #3310
• Fixed export_coreml() tilda expansion #3309
• Better error message when using image deep features with wrong model #3332

New Features

• Python 3.8 support #3296
• Expose deep features for images #3210
• SArray median #3222

Feature Enhancements

• Drawing Classifier uses MLCompute #3279
• Object Detection uses MLCompute #3278
• Add Non-maximum Suppression layer to Object Detection #3274
• Relax TensorFlow requirement #3281
• Add L2 Regularization to Text Classifier #3285

Bug Fixes

• Stop disabling TensorFlow version 2 behavior #3273
• Empty SArray cannot contain audio data #3208
• Better handling of setting SSL certificate path #3211

New Features

• Created a set of minimal wheels designed for users that only want to use SFrame, SArray, or SGraph and none of the machine learning functionality. These minimal wheels significantly reduce the total number of dependencies which need to be installed. These wheels are included below but will not be put on PyPI.
• SArray shuffle. #3178

Feature Enhancements

• Style Transfer: include style and content loss in metrics, when available. #3131
• Better S3 error handling. #3116
• Sound Classifier: better accuracy, particularly for training on long audio clips, by shuffling stacked deep features prior to training. #3178

Bug Fixes

• Fixed bad formatting typo in User Guide install instructions. #3141
• Clarify the use of tc.config.set_num_gpus. #3181

Three Critical Bug Fixes for TensorFlow Implementation of Object Detection:

• Fix color augmentation in TensorFlow object detection #3176
• Fix TensorFlow object detector evaluate always returning zero accuracy #3148
• Fix TensorFlow object detector not working for more than two objects/classes #3031

Pinned 2nd-order dependency llvmlite to version 0.31.0 on python versions 2.7 and 3.5 to avoid install errors.

New Features

• SFrame shuffle operator (#942)
• Object Detector takes seed parameter (#2168)
• Image Classifier and Image Similarity use eGPU (#883)

Feature Enhancements

• Unfix numpy dependency (#2664)
• Rewrite S3 code path with AWS SDK (#3025)
• S3 code path uses AWS SigV4 (#2910)

Bug Fixes

• Fix tc.load_model and tc.load_sgraph APIs to respect the user’s home directory (#2685)
• Temporarily restrict TensorFlow version range on Linux (#3089)
• Add type checking for tc.recommender.util.compare_models (#2758)
• Clarify definition of “SFrame” in user docs (#2908)
• Fix broken links on the tc.aggregate API doc page (#3011)
• Improve logging of execution path for GPU/CPU use cases (#3037)
• Fix threading issue in column indexer of ml_data. (#2979)
• Fix typos in Sound Classifier user guide (#3020)

New Features

Remove duplicate columns of a an SFrame (#2540) SArray function: abs() (#1885)

Feature Enhancements

Add metadata to exported CoreML models (#2800) Unpin numpy dependency by using new coremltools version (#2664) Better error message for unsupported platform on TuriCreate install (#1163) Nearest Neighbors: better logging and model summary (#2460) Object Detection: add warning message if class label not in training set (#2551) Drawing Classifier: evaluate method is faster (#2932) One Shot Object Detection: better tar file management for background images (#2005) Style Transfer: parameter checking (#2790) Style Transfer: model loaded lazily (#2799) Sound Classifier: better parameter checking (#2882, #2884, #2886)

Bug Fixes

Fixed ImportError: libbz2.so.1.0 error message (#2302) Stop calling tensorflow.compat.v1.reset_default_graph() (#2731) Protocol component of URL is now case insensitive (#2902) All toolkits: better parameter checking for summary method (#2872) Fix verbose=False flag for several toolkits (#2677) Activity Classifier: exporting a model to CoreML which uses just one feature (#2945) Activity Classifier: drop missing data (#2841) Activity Classifier: check for missing features (#2844) Activity Classifier: better parameter checking (#2837) Style Transfer: disable weight update for the VGG parameters (#2990) Style Transfer: weight initialization in style transfer to use uniform between -0.7 and 0.7 (#2988) Style Transfer: remove duplicated printed progress output (#2926) Style Transfer: improved docstings (#2789) Style Transfer: remove outdated information from user guide (#2620) Object Detection: fix inference regression between CPU and GPU (#2955) Object Detection: predict doesn’t depend on annotation column (#2830) One Shot Object Detection: fix user guide links (#2959) One Shot Object Detection: summary improvements (#2862, 2863) Sound Classifier: fix max_iterations=0 bug (2764) Image Classifier: fix docstring link (#931)

Release Notes

New Features

• Swapped out using TensorFlow for MXNet for Deep Learning toolkits.
• Added support for Python 3.7 on Linux.

Feature Enhancements

• Added support for coremltools 3.1. (#2774)
• Added user-defined metadata in Core ML model. (#2562)
• Added guide of how to use confidence thresholds with the Vision framework. (#2615)
• Enabled activity classifier to run on Metal Performance Shaders for GPU based training. (#919)
• Added SFrame plus operator. (#2569)
• Supported image visualization on Google Colab Jupyter Notebook. (#1269, #2483)
• Ignore type when appending empty SArray. (#2390)
• Supported SArray’s filter_by function with dictionary values and keys. (#2221)
• Improved error message in apply method when most values are None. (#2066)
• Changed Sketch.num_na() to Sketch.num_missing(). (#1759)
• Updated SFrame.read_csv()'s API documentation. (#1367)
• Enabled export for an empty SFrame. (#804)

Bug Fixes

• Suppressed warnings to users while calling non-deprecated API. (#2529)
• Removed unused state variables from the object detector model. (#2506)
• Fixed conflicting Numpy dependencies (#2249)
• Handled invalid path for loading audio files. (#2686)
• Updated user guide to incorporate the new AC export_coreml format. (#2676)
• Fixed segfault for text classifier with data with undefined values. (#2402)
• Fix Image Classifier and Drawing Classifier annotation bugs on Linux. (#2401)
• Reduced CPU architecture requirement to core2. (#2266)
• Output of drawing classifier evaluation object now does not contain additional keys. (#1974)
• Image classifier annotation function throws error with empty SFrame. (#1915)
• Preserved nan values when converting from float SArray into int SArray. (#1761)
• Fixed functionality for predict_id in Activity Classifier’s predict function. (#1706)
• Fixed bug for dictionary input to text_analytics’ count_words function. (#954)

Release Notes

New Features

• Style Transfer supports training with Mac GPU

Feature Enhancements

• Improved speed for groupby operations over common keys (#2210)
• Added new API for SArray's read_json initializer (#2331)
• Improved speed for SFrame sample function. (#2343)
• Added equality operator for image_type (#2347)
• Enabled Sound Classifier predict for clips shorter than 975 milliseconds (#2310)

Bug Fixes

• Enhanced logging in exception messages (#2145)
• Fixed typos in Userguide (#2268)
• Fixed issue with degree to radian conversion in One Shot Object Detector. (#2285)
• Added column existence check in Image Classifier (#2296)
• Enhanced seeding strategy in One Shot Object Detector (#2297)
• Fixed inline Jupyter notebook title (#2306)
• Enhanced One Shot Object Detector rotation angle distributions (#2316)
• Clarified userguide regarding differing bounding box representation between Turi Create and Vision (#2320)
• Added column type check in SFrame for all toolkits. (#2324)
• Enhanced error checking for input datatype for all toolkits. (#2328)
• Recognized exclude parameter in Recommender. (#2333)
• Fixed evaluate().explore() crashes and hangs on Python 3 (#2337)
• Suppressed warnings in libcurl dependency (#2366)
• Added support for single image input to One Shot Object Detector. (#2370)

Notes

• One Shot Object Detector is released out of Beta.
• Added deprecation warnings for 6.0 release (#2321)

Installation

To install Turi Create 5.7.1 into your virtual environment, use:

(venv) pip install turicreate==5.7.1

Bug Fixes

• Visualization: .explore() method breaks in Python3(#2264)