Active Learning with the Nvidia TLT

Tutorial on active learning with the Nvidia Transfer Learning Toolkit (TLT).

In this tutorial, we will show you how you can do active learning for object detection with the Nvidia Transfer Learning Toolkit. The task will be object detection of apples in a plantation setting. Accurately detecting and counting fruits is a critical step towards automating harvesting processes. Furthermore, fruit counting can be used to project expected yield and hence to detect low yield years early on.

The structure of the tutorial is as follows:

1. Prerequisites
   1. Set up Lightly
   2. Set up Nvidia TLT
   3. Data
2. Active Learning
   1. Initial Sampling
   2. Training and Inference
   3. Active Learning Step
   4. Re-training

To get started, clone this repository to your machine and change the directory.

git clone https://github.com/lightly-ai/NvidiaTLTActiveLearning.git
cd NvidiaTLTActiveLearning

1 Prerequisites

For this tutorial, you require Python 3.6 or higher. You also need to install lightly, numpy and argparse.

pip install -r requirements.txt

1.1 Set up Lightly for Active Learning

To set up lightly for active learning, head to the Lightly Platform and create a free account by logging in. Make sure to get your token by clicking on your e-mail address and selecting "Preferences". You will need the token for the rest of this tutorial.

1.2 Set up Nvidia TLT

To install the Nvidia Transfer Learning Toolkit, follow these instructions. If you want to use custom scripts for training and inference, you can skip this part.

Setting up Nvidia TLT can be done in a few minutes and consists of the following steps:

1. Install Docker.
2. Install Nvidia GPU driver v455.xx or above.
3. Install nvidia docker2.
4. Get an NGC account and API key.

To make all relevant directories accessible to the Nvidia TLT, you need to mount the current working directory and the yolo_v4/specs directory to the Nvidia TLT docker. You can do so with the provided mount.py script.

python mount.py

Next, you need to specify all training configurations. The Nvidia TLT expects all training configurations in a .txt file which is stored in the yolo_v4/specs/ directory. For the purpose of this tutorial we provide an example in yolo_v4_minneapple.txt. The most important differences to the example script provided by Nvidia are:

• Anchor Shapes: We made the anchor boxes smaller since the largest bounding boxes in our dataset are only approximately 50 pixels wide.
• Augmentation Config: We set the output width and height of the augmentations to 704 and 1280 respectively. This corresponds to the shape of our images.
• Target Class Mapping: For transfer learning, we made a target class mapping from car to apple. This means that every time the model would now predict a car, it predicts an apple instead.

1.3 Data

In this tutorial, we will use the MinneApple fruit detection dataset. It consists of 670 training images of apple trees, annotated for detection and segmentation. The dataset contains images of trees with red and green apples.

Note: The Nvidia TLT expects the data and labels in the KITTI format. This means they expect one folder containing the images and one folder containing the annotations. The name of an image and its corresponding annotation file must be the same apart from the file extension. You can find the MinneApple dataset converted to this format attached to the first release of this tutorial. Alternatively, you can download the files from the official link and convert the labels yourself.

Create a data/ directory, move the downloaded minneapple.zip file there, and unzip it

cd data/
unzip minneapple.zip
cd ..

Here's an example of how the converted labels look like. Note how we use the label car instead of apple because of the target class mapping we had defined in section 1.2.

Car 0. 0 0. 1.0 228.0 6.0 241.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 5.0 228.0 28.0 249.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 30.0 238.0 46.0 256.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 37.0 214.0 58.0 234.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 82.0 261.0 104.0 281.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 65.0 283.0 82.0 301.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 82.0 284.0 116.0 317.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 111.0 274.0 142.0 306.0 0. 0. 0. 0. 0. 0. 0.
Car 0. 0 0. 113.0 308.0 131.0 331.0 0. 0. 0. 0. 0. 0. 0.

2 Active Learning

You can also upload thumbnails or even just metadata about the images. See this link for more information.

lightly-magic \
    input_dir=./data/raw/images \
    trainer.max_epochs=0 \
    loader.num_workers=8 \
    collate.input_size=512 \
    new_dataset_name="MinneApple" \
    token=MY_TOKEN

The above command will display the id of your dataset. You will need this later in the tutorial.

Once the upload has finished, you can visually explore your dataset in the Lightly Platform.

2.1 Initial Sampling

Lightly offers different sampling strategies, the most prominent ones being CORESET and RANDOM sampling. RANDOM sampling will preserve the underlying distribution of your dataset well while CORESET maximizes the heterogeneity of your dataset. While exploring our dataset in the Lightly Platform, we noticed many different clusters therefore we choose CORESET sampling to make sure that every cluster is represented in the training data.

Use the active_learning_query.py script to make an initial selection:

python active_learning_query.py \
    --token YOUR_TOKEN \
    --dataset_id YOUR_DATASET_ID \
    --new_tag_name 'initial-selection' \
    --n_samples 100 \
    --method CORESET

The above script roughly performs the following steps:

It creates an API client to communicate with the Lightly API.

# create an api client
client = ApiWorkflowClient(
    token=YOUR_TOKEN,
    dataset_id=YOUR_DATASET_ID,
)

Then, it creates an active learning agent which serves as an interface to do active learning.

# create an active learning agent
al_agent = ActiveLearningAgent(client)

Finally, it creates a sampling configuration, makes an active learning query, and puts the annotated images into the data/train directory.

# make an active learning query
cofnig = SamplerConfig(
    n_samples=100,
    method=SamplingMethod.CORESET,
    name='initial-selection',
)
al_agent.query(config)

# simulate annotation step by copying the data to the data/train directory 
oracle.annotate_images(al_agent.added_set)

The query will automatically create a new tag with the name initial-selection in the Lightly Platform.

You can verify that the number of annotated images is correct like this:

ls data/train/images | wc -l
ls data/train/labels | wc -l

2.2 Training and Inference

mkdir -p ./yolo_v4/pretrained_resnet18
ngc registry model download-version nvidia/tlt_pretrained_object_detection:resnet18 \
    --dest ./yolo_v4/pretrained_resnet18

Finetuning the object detector on the sampled training data is as simple as the following command. Make sure to replace YOUR_KEY with the API token you get from your Nvidia account.

mkdir -p $PWD/yolo_v4/experiment_dir_unpruned
tlt yolo_v4 train \
    -e /workspace/tlt-experiments/yolo_v4/specs/yolo_v4_minneapple.txt \
    -r /workspace/tlt-experiments/yolo_v4/experiment_dir_unpruned \
    --gpus 1 \
    -k YOUR_KEY

Now that you have finetuned the object detector on your dataset, you can do inference to see how well it works.

Doing inference on the whole dataset has the advantage that you can easily figure out for which images the model performs poorly or has a lot of uncertainties.

tlt yolo_v4 inference \
    -i /workspace/tlt-experiments/data/raw/images/ \
    -e /workspace/tlt-experiments/yolo_v4/specs/yolo_v4_minneapple.txt \
    -m /workspace/tlt-experiments/yolo_v4/experiment_dir_unpruned/weights/yolov4_resnet18_epoch_050.tlt \
    -o /workspace/tlt-experiments/infer_images \
    -l /workspace/tlt-experiments/infer_labels \
    -k MY_KEY

Below you can see two example images after training. It's evident that the model does not perform well on the unlabeled image. Therefore, it makes sense to add more samples to the training dataset.

2.3 Active Learning Step

python active_learning_query.py \
    --token YOUR_TOKEN \
    --dataset_id YOUR_DATASET_ID \
    --preselected_tag_name 'initial-selection' \
    --new_tag_name 'al-iteration-1' \
    --n_samples 200 \
    --method CORAL

# create a scorer to calculate active learning scores based on model outputs
scorer = ScorerObjectDetection(model_outputs)

# create an api client
client = ApiWorkflowClient(
    token=YOUR_TOKEN,
    dataset_id=YOUR_DATASET_ID,
)

# create an active learning agent and set the preselected tag
al_agent = ActiveLearningAgent(
    client,
    preselected_tag_name='initial-selection',
)

# create a sampler configuration
config = SamplerConfig(
    n_samples=200,
    method=SamplingMethod.CORAL,
    name='al-iteration-1',
)

# make an active learning query
al_agent.query(config, scorer)

# simulate the annotation step
oracle.annotate_images(al_agent.added_set)

ls data/train/images | wc -l
ls data/train/labels | wc -l

2.4 Re-training

tlt yolo_v4 train \
    -e /workspace/tlt-experiments/yolo_v4/specs/yolo_v4_minneapple.txt \
    -r /workspace/tlt-experiments/yolo_v4/experiment_dir_unpruned \
    --gpus 1 \
    -k MY_KEY

If you're still unhappy with the performance after re-training the model, you can repeat steps 2.2 and 2.3 and then re-train the model again.