AlphaTracker is tested in Linux systems broadly and those systems that have GPUs. However, if you would do not have Linux-compatible system or do not have access to consumer GPUs, you can use Google Colab for the purposes of training and tracking Alphatracker!
Google Colab is only useful for training and inference purposes. If you have a fully-labeled set of data using the Alphatracker annotation tool, continue on. If you have NOT labeled your data, please complete this step first!
Also, if you are trying out Alphatracker and would like to use the sample data, click here to download this data.
To access the Colab notebook, click here
Lastly, please run different experiments if you have a variable number of animals across the JSON files.
Find your folder/folders of data that contain the training images and the JSON annotated files.
Open your Google Drive, and upload the folder/folders, and the videos you want to label, into the Drive under the main My Drive folder. Ensure that your Drive has ample storage after uploading the files!!
Open the AlphaTrackerCOLAB_12-2-20.ipynb, which can be found at this link here. You have now opened your Python session
Click Runtime and then Change runtime type. From the dropdown menu, select GPU.
Now, we will connect our Python session to our Google Drive. Press the play button next to the code block to run the first cell. You will be prompted with an authorization link. Click this link and follow the instructions. Select the Google Drive account to which you have uploaded your data from Step 1. Paste the authorization code back into the prompt box.
Run the second code block. This is an important note: The main My Drive folder has the following path: /content/drive/My Drive. We are now inside this main folder
Run the following code block to download AlphaTracker into your Google Drive. Wait a minute or two, then go to your My Drive folder. You should notice a new folder by the name AlphaTracker has appeared.
In the following code block, enter in the variables that will define your experiment. Very important: Ensure that your training data for a specific experiment has the same number of animals and poses. If there are samples with different numbers, please separate them as run them with different experiment names!
In experiment_name, type in the name for the experiment inside the quotation marks.
In the following code block, in the variable image_root_list, enter the path to all your image folders.
In the variable json_path_list, enter the path to all the respective JSON training files in the same order as the image folders in image_root_list.
In number_of_animals, enter in the number of animals in each of the JSON files inside the brackets.
In number_of_poses, enter in the number of bodyparts that are being labeled per animal.
In image_suffix, enter the extension of the image, such as "jpg", or "png".
In video_path, enter in the path of where the videos to-be-tracked are located.
In start_frame, enter in the first frame of the video that you would like to label.
In end_frame, enter in the last frame of the video that you would like to label.
In number_of_animals_in_video, enter in the number of animals in the video.
In visualize_tracking_results, enter in either 'yes' or 'no' if you would like to visualize the tracking results.
The following are optional, but can be changed
In sppe_learning_rate, type in the learning rate for the SPPE model.
In sppe_iterations, type in the number of iterations for the SPPE model.
In sppe_batch_size, type in the batch size for the SPPE model.
In yolo_learning_rate, type in the learning rate for the YOLO model.
In yolo_iterations, type in the number of iterations for the YOLO model.
In yolo_batch_size, type in the batch size for the YOLO model.
Remember, the main My Drive folder has the path /content/drive/My Drive, therefore, a folder named Images that is under the main My Drive folder would have the path /content/drive/My Drive/Images. This is what you should type. For example, image_root_list = ['/content/drive/My Drive/Folder1', '/content/drive/My Drive/Folder2'] and etc...
Run the following code blocks back-to-back without any alterations. This will take about 6-10 minutes to complete!
Run the following code block to train AlphaTracker! This step can take anywhere from 30 minutes to 6 hours. It depends on how many iterations you are training for.
Run the following code block to perform tracking on the videos you listed in setting.py. Once this step is complete, you can go to the folder you designated in the result_folder variable in setting.py to find the location of the tracked results!
