search

conservationtechlab / animl-py

Animl comprises a variety of machine learning tools for analyzing ecological data. This Python package includes a set of functions to classify subjects within camera trap field data and can handle both images and videos.
MIT License
6 stars 3 forks source link

readme

animl-py

AniML comprises a variety of machine learning tools for analyzing ecological data. This Python package includes a set of functions to classify subjects within camera trap field data and can handle both images and videos. This package is also available in R: animl

Table of Contents

  1. Installation
  2. Usage
  3. Models

Installation Instructions

It is recommended that you set up a conda environment for using animl. See Dependencies below for more detail. You will have to activate the conda environment first each time you want to run AniML from a new terminal.

From GitHub

git clone https://github.com/conservationtechlab/animl-py.git
cd animl-py
pip install -e .

From PyPi

pip install animl

Dependencies

We recommend running AniML on GPU-enabled hardware. **If using an NVIDIA GPU, ensure driviers, cuda-toolkit and cudnn are installed. PyTorch will install these automatically if using a conda environment. The /models/ and /utils/ modules are from the YOLOv5 repository. https://github.com/ultralytics/yolov5

Python >= 3.9

Python Package Dependencies

Animl also depends on exiftool for accessing file metadata.

Verify Install

We recommend you download the examples folder within this repository. Download and unarchive the zip folder. Then with the conda environment active:

python -m animl /path/to/example/folder

This should create an Animl-Directory subfolder within the example folder.

Or, if using your own data/models, animl can be given the paths to those files: Download and unarchive the zip folder. Then with the conda environment active:

python -m animl /example/folder --detector /path/to/megadetector --classifier /path/to/classifier --classlist /path/to/classlist.txt

You can use animl in this fashion on any image directory.

Usage

Inference

The functionality of animl can be parcelated into its individual functions to suit your data and scripting needs. The sandbox.ipynb notebook has all of these steps available for further exploration.

  1. It is recommended that you use the animl working directory for storing intermediate steps.

    from animl import file_management
workingdir = file_management.WorkingDirectory('/path/to/save/data')

  2. Build the file manifest of your given directory. This will find both images and videos.

    files = file_management.build_file_manifest('/path/to/images', out_file = workingdir.filemanifest)

  3. If there are videos, extract individual frames for processing. Select either the number of frames or fps using the argumments. The other option can be set to None or removed.

    from animl import video_processing
allframes = video_processing.extract_frames(files, out_dir=workingdir.vidfdir, out_file=workingdir.imageframes,
                                        parallel=True, frames=3, fps=None)

  4. Pass all images into MegaDetector. We recommend MDv5a. The function parse_MD will convert the json to a pandas DataFrame and merge detections with the original file manifest, if provided.

from animl import detect, megadetector
detector = megadetector.MegaDetector('/path/to/mdmodel.pt')
mdresults = detect.detect_MD_batch(detector, allframes["Frame"], checkpoint_path=working_dir.mdraw, quiet=True)
detections = detect.parse_MD(mdresults, manifest=all_frames, out_file=workingdir.detections, parallel=True)

  1. For speed and efficiency, extract the empty/human/vehicle detections before classification.

    from animl import split
animals = split.get_animals(detections)
empty = split.get_empty(detections)

  2. Classify using the appropriate species model. Merge the output with the rest of the detections if desired.

    from animl import classifiers, inference
classifier, class_list = classifiers.load_model('/path/to/model', '/path/to/classlist.txt')
animals = inference.predict_species(animals, classifier, class_list, file_col="Frame",
                                batch_size=4, out_file=working_dir.predictions)
manifest = pd.concat([animals if not animals.empty else None, empty if not empty.empty else None]).reset_index(drop=True)

  3. (OPTIONAL) Save the Pandas DataFrame's required columns to csv and then use it to create json for TimeLapse compatibility

from animl import timelapse, animl_results_to_md_results
csv_loc = timelapse.csv_converter(animals, empty, imagedir, only_animl = True)
animl_results_to_md_results.animl_results_to_md_results(csv_loc, imagedir + "final_result.json")
  1. (OPTIONAL) Create symlinks within a given directory for file browser access. 
    linked_manifest = symlink_species(manifest, workingdir.linkdir, file_col="FilePath", copy=False)
pd.to_csv(linked_manifest, workindir.results)

Training

Training workflows are still under development. Please submit Issues as you come upon them.

  1. Assuming a file manifest of training data with species labels, first split the data into training, validation and test splits. This function splits each label proportionally by the given percentages, by default 0.7 training, 0.2 validation, 0.1 Test.

    from animl import split
train, val, test, stats = train_val_test(manifest, out_dir='path/to/save/data/', label_col="species",
               percentage=(0.7, 0.2, 0.1), seed=None)

  2. Set up training configuration file. Specify the paths to the data splits from the previous step. Example .yaml file:

    
seed: 28  # random number generator seed (long integer value)
device: cuda:0  # set to local gpu device 
num_workers: 8  # number of cores

dataset parameters

num_classes: 53 #might need to be adjusted based on the classes file training_set: "/path/to/save/train_data.csv" validate_set: "/path/to/save/validate_data.csv" test_set: "/path/to/save/test_data.csv" class_file: "/home/usr/machinelearning/Models/Animl-Test/test_classes.txt"

training hyperparameters

architecture: "efficientnet_v2_m" image_size: [299, 299] num_epochs: 100 batch_size: 16

learning_rate: 0.003 weight_decay: 0.001

overwrite .pt files

overwrite: False

experiment_folder: '/home/usr/machinelearning/Models/Animl-Test/'


class_file refers to a flle that contains index,label pairs. For example:<br>
test_class.txt

id,Code,Species,Common 1,cat, Felis catus, domestic cat 2,dog, Canis familiaris, domestic dog


3. (Optional) Update train.py to include MLOPS connection. 

4. Using the config file, begin training
```bash
python -m animl.train --config /path/to/config.yaml

Every 10 epochs, the model will be checkpointed to the 'experiment_folder' parameter in the config file, and will contain performance metrics for selection.

  1. Testing of a model checkpoint can be done with the "test.py" module. Add an 'active_model' parameter to the config file that contains the path of the checkpoint to test. This will produce a confusion matrix of the test dataset as well as a csv containing predicted and ground truth labels for each image. 
    python -m animl.test --config /path/to/config.yaml

Models

The Conservation Technology Lab has several models available for use.