AgML 0.7.0

This is a major update to AgML that includes a new agml.models API, new datasets, new tools, and a suite of bugfixes and improvements across the board.

Major Changes

• The agml.models.Detector API has been introduced. This API wraps the Ultralytics YOLO11 model (see ultralytics for more information) and provides an easy-to-use interface for training and inference with state-of-the-art object detection models.
  • Use agml.models.Detector.train(loader, 'yolo_model') to quickly train an AgML dataset using your choice of YOLO model - all right within the AgML API.
  • Use agml.models.Detector.load('model_name') to load a trained model easily, stored directly within the AgML local model repository.
  • Use agml.models.Detector.load_benchmark('dataset_name', 'yolo_name') to access a pre-trained agricultural object detection model from a set of available pretrained benchmarks.
  • Note: If you want fine-grained transfer learning and finetuning capabilities, use the standard agml.models.DetectionModel API, but the agml.models.Detector API is a much more user-friendly quick-start for quick training and inference.
• We have added four new datasets:
  • tomato_ripeness_detection: Object Detection
  • corn_maize_leaf_disease: Image Classification
  • tomato_leaf_disease: Image Classification
  • vine_virus_photo_dataset: Image Classification
• A new export tool has been added agml.data.extensions.restructure_cvat_annotations, that automatically reformats the exported COCO annotations from the CVAT tool into the format used in AgML. This increases AgML's interoperability with existing machine learning frameworks.

Major Bugfixes

• Helios compilation and execution is now fixed and works for Windows systems - Windows users can now use the agml.synthetic API to generate their own simulated data.

Other Changes and Improvements

• The documentation for AgML datasets has been improved to show more representative samples, and datasets without documentation now have theirs added.
• Improvements have been made to the agml.data.ImageLoader, and there is now a new method agml.data.AgMLDataLoader.take_images() which enables you to extract just the images from an AgMLDataLoader, useful for inference and other image-only applications.
• A bug which prevented saving and loading splits for object detection has been fixed: you can now load and save object detection data splits successfully.
• Some small errors with visualization have been fixed, including orientation and number of images in certain methods.
• The agml.data.tools.convert_bbox_format tool now has new formats xyxy and yxyx which wrap longer strings, for ease of use.

AgML v0.6.2

This release introduces a new export tool for AgML datasets, various expansions to methods, and some bugfixes.

Major Changes

• A new method agml.data.exporters.export_yolo() has been added (alongside a companion wrapper to a dataset, loader.export_yolo()), that enables exporting datasets to the Ultralytics YOLO format. The resulting datasets can be directly integrated in one line of code into Ultralytics/Darknet YOLO training pipelines, which augments AgML's object detection resources.

Improvements

• Splitting datasets, specifically with multi-dataset loaders, now works properly.

Bugfixes

• Added an export_tensorflow() method for multi-loaders, which was previously missing.
• Updated Helios compilation instructions to be compatible with the C++17 standard needed for updated Helios versions.
• All example notebooks have been updated to be consistent with new AgML methods.
• AgML visualization methods now no longer double-display in Jupyter notebooks.

AgML 0.6.1

This is a small update.

• A new dataset vegann_multicrop_presence_segmentation has been added: this is the AgML-incorporation of the VegAnn dataset.
• Fixed an issue with the class names in the rangeland_weeds_australia dataset: the awry "no_weeds" class has been removed and the dataset now follows the original specification of the DeepWeeds dataset.
• A bug that prevented the storage/loading of custom splits has been fixed.
• A bug that cut off text when visualizing image classification data samples has been fixed.

Full Changelog: v0.6.0...v0.6.1

AgML 0.6.0

This release introduces tools for training custom agricultural machine learning models using AgML.

Main Changes

agml.models

• The agml.models API has been extended with new features for training, namely the run_training method which enables quick training of image classification, semantic segmentation, and object detection models.
  • Simply instantiate a model with your number of classes, build an AgMLDataLoader with your data preprocessing, and pass it to the run_training method alongside other training hyperparameters to train a model.
  • Choose your level of customizability: for newer users, options like optimizers, loss, and other hyperparameters are auto-selected, but for experienced users, you can go as far as extending the training_step, validation_step, and other arguments for greater customizability over training.

Other Changes and Bugfixes

• A major bug which prevented recompliation of Helios without LiDAR has been fixed, enabling users to switch between using LiDAR-compiled Helios and standard Helios.
• A bug which caused Helios installation on basic inspection is now patched. Helios will no longer auto-install unless the agml.synthetic module is actively used for data generation.
• You can now correctly display image samples when using agml.viz.show_sample with the image_only option.
• Bugfixes have been done for agml.models.metrics.Accuracy and agml.models.metrics.MeanAveragePrecision to ensure that they work with training.

Read the Full Changelog: v0.5.2...v0.6.0

AgML 0.5.2

New Features and Bugfixes

• Added the ability to save and load data splits for individual datasets using loader.save_split and loader.load_split.
• Added the ability to load multiple datasets using a wildcard name instead of passing each individual dataset in a list (e.g., to load all of the apple detection datasets, use apple_detection*).
• Use agml.models.get_benchmark to get benchmark information for a certain model trained on a certain dataset.
• Fixed bugs with mAP calculation during training.
• Fixed bugs with image preprocessing for object detection and visualization of object detection predictions.

Full Changelog: v0.5.1...v0.5.2

AgML 0.5.1

Main Changes

• Added a new dataset, ghai_strawberry_fruit_detection.
• Added a new method agml.io.read_image for convenience.
• Added the ability to generate both RGB & LiDAR data using the option agml.synthetic.SimulationType.Both.

Bugfixes and Updates

• You can now instantiate a collection of synthetic datasets by passing a list of them to the initialization of an AgMLDataLoader.
• A bug which prevented the visualization of predictions when using a SegmentationModel has now been patched.

Full Changelog: v0.5.0...v0.5.1

AgML 0.5.0

This release introduces synthetic LiDAR data generation through Helios, a rework of the agml.viz module with better functionality, and a number of new available datasets, amongst other features.

Major Changes

• The agml.io module has been added, with a few convenience functions for working with file and directory structures.
  • Currently available functions include get_file_list and get_dir_list which also work with nested structures, as well as recursive_dirname, parent_path, and random_file.

agml.data

• Three new datasets have been introduced:
  • Object Detection: ghai_iceberg_lettuce_detection, ghai_broccoli_detection
  • Image Classification: riseholme_strawberry_classification_2021
• The agml.data.ImageLoader has been added, which is essentially a simple loader designed only for images.
  • Enables loading images from a nested directory structure.
  • Enables easy resizing and transforms of the loaded images.
• The AgMLDataLoader now has a new method show_sample which can be used to visualize samples from the dataset directly from the loader.

agml.synthetic

• LiDAR Data Generation: You can generate LiDAR-based synthetic data now, using opt.simulation_type = agml.synthetic.SimulationType.LiDAR.
• You can recompile Helios with LiDAR enabled, as well as in parallel (for Linux and MacOS systems) using recompile_helios(lidar_enabled = True) and recompile_helios(parallel = True) [note that parallel compilation is enabled by default].
• A new loader agml.synthetic.LiDARLoader has been added, which can be used to load point clouds from a generated directory in the same format as ImageLoader, and can be used to get point clouds and visualize them.

agml.viz

• The entire agml.viz module has been reworked with new methods and a functional visualization backend which enables both cv2 and matplotlib displays, depending on what the user desires.
• The following is a mapping of the old functions to new functions:
  • visualize_images_and_labels -> show_images_and_labels
  • output_to_mask -> convert_mask_to_colored_image
  • overlay_segmentation_masks -> annotate_semantic_segmentation
  • visualize_image_and_mask -> show_image_and_mask
  • visualize_overlaid_masks -> show_image_and_overlaid_mask
  • visualize_image_mask_and_predicted -> show_semantic_segmentation_truth_and_prediction
  • annotate_bboxes_on_image -> annotate_object_detection
  • visualize_image_and_boxes -> show_image_and_boxes
  • visualize_real_and_predicted_bboxes -> show_object_detection_truth_and_prediction
  • visualize_images -> show_images
• To swap between viz backends, use get_viz_backend and set_viz_backend.
• In order to simply display an image, you can use display_image.
• To visualize a point cloud (either in Open3D if installed, or matplotlib), use show_point_cloud.

Other Changes

• Major Change for New Users: torch, torchvision, and other related modeling packages for the agml.models package are no longer distributed with the AgML requirements -- you must install these on your own if you want to use them.

Minor/Functional Changes + Bugfixes

• Fixed backend swapping between TensorFlow and PyTorch when using clashing transforms.
• Added the ability to run prediction with a classification and a segmentation model without normalizing the input image using model.predict(..., normalize = False).
• Images no longer auto-show as matplotlib figures when using agml.viz.show_* methods, instead they are returned as image arrays and can be displayed in a desired format.
• Improved access and setting of information in the backend config.json file, so information is not accidentally overwritten.

Full Changelog: v0.4.7...v0.5.0

AgML 0.4.7

Main Changes

• Introduced two new public object detection datasets from the Western Growers Global Harvest Automation Initiative (https://github.com/AxisAg/GHAIDatasets): ghai_romaine_detection, and ghai_green_cabbage_detection.
• Added the ability to generate multiple different types of annotations when generating synthetic datasets with Helios.

Other Changes

• Fixed errors with different visualization utilities as well as synthetic data annotation conversions.

Full Changelog: v0.4.6...v0.4.7

AgML 0.4.6

Main Changes

• You can now manually generate synthetic data with Helios using your own C++ generation file and CMake parameters/XML style file, using the method agml.synthetic.manual_generate_data. This allows for a greater deal of customizability and control over how Helios is compiled, as well as which methods are being used.
• A convenience method agml.viz.show_sample has been added which enables quick visualization of an image and its annotations: this works for all types of annotations, simply pass a loader as the argument.

Other Improvements

• If you want to overwrite files when generating synthetic data with the HeliosDataGenerator, you can now pass the argument clear_existing_files = True to the generate method.
• You can now compile Helios using the standard API in both the 'Release' and 'Debug' modes.

Read the Full Changelog Here.

AgML 0.4.5

This feature adds a couple of new features as well as bugfixes for the existing API.

Main Changes

• You can now use image classification models (agml.models.ClassificationModel) as image regression models by passing regression = True upon instantiation. This drops the final argmax computation and returns the softmax regression values.
• Pass a custom set of RGB values to agml.viz.set_colormap to use a custom colormap.
• A new preprocessing function has been added to agml.models.preprocessing: agml.models.preprocessing.imagenet_preprocess, which prepares images for input to an ImageNet-backend model (image classification, semantic segmentation).

Bugfixes

• The MeanAveragePrecision metric has been fixed, and no longer throws errors for empty predictions (or for early-stage training results).
• Custom object detection datasets can now be auto-loaded and classes automatically inferred without throwing an error.

Read the Full Changelog Here.