Running TAO Toolkit on Google Colab

Google Colab provides access to free GPU instances for running compute jobs in the cloud. This page provides instructions for getting started with TAO Toolkit on Google Colab.

Google Colab has some restrictions with TAO based on the limitations of the hardware and software available with the Colab Instances. You can view the limitations in the Notes section.

To use Google Collab, you’ll need a Google Account with Google Drive access.

Log in to your Google account,or create one by following instructions on the Sign-Up for Gmail page

TAO Toolkit provides an extensive model zoo containing pretrained models for both computer-vision and conversational AI use cases. You can click on the “One-Click Deploy” link for the model of your use-case.

Refer to the Steps to Locate files in a Colab Notebook section for an example of accessing files within the Colab notebook.

General-Purpose Computer Vision Models

With general purpose models, you can train an image classification model, object detection model, or an instance segmentation model.

Purpose-Built Computer Vision Models

Purpose-built models are built for high accuracy and performance. You can deploy these models out of the box for applications such as smart city, retail, public safety, and healthcare. You can also retrain them with your own data.

Conversational AI Models

TAO Pre-trained Models (Inference Only)

In addition to training different models using the one-click deploy links, you can run inference with the Pre-trained Models TAO has published using this Notebook

If you have limited storage space, or want to iterate quickly through training experiments, we suggest you

• Download data on to your local system

• Run a utility script to generate a subset of the dataset

• Upload this smaller dataset to your Google Drive to run experiments on

TAO Toolkit provides utility scripts to generate such subsets for COCO dataset (which is around ~25 GB with ~120k images) and KITTI dataset (which is around ~12 GB with ~14k images)

To obtain subset for KITTI:

• The subset generation file is present here. Run this script in your local system (not in Colab as you might have storage limitations)

• Download and unzip kitti training and testing zip files here, you’ll have 2 folders training and testing

• To obtain subset for training:

  Copy

  Copied!

              
  
              
  python generate_kitti_subset.py --source-data-dir=path_to_training_folder --out-data-dir=path_to_save_subset_data/training/ --training True --num-images=num_of_images_in_subset
          

  Example

  Copy

  Copied!

              
  
              
  python generate_kitti_subset.py --source-data-dir=/home/user/data/training --out-data-dir=/home/user/subset_data/training/  --training True --num-images=100
          

• To obtain subset for testing:

  Copy

  Copied!

              
  
              
  python generate_kitti_subset.py --source-data-dir=path_to_training_folder --out-data-dir=path_to_save_subset_data/training/ --num-images=num_of_images_in_subset
          

  Example

  Copy

  Copied!

              
  
              
  python generate_kitti_subset.py --source-data-dir=/home/user/data/testing --out-data-dir=/home/user/subset_data/testing/ --num-images=100
          

The same procedure and arguments are applicable to obtain subset for pointpillars data, except the file to run is present here. For pointpillars, we need KITTI with velodyne and calib also downloaded

Dataset folder structure for kitti:

            

            
path_to_training_folder
|___calib (only for pointpillars)
|___image_2
|___label_2
|___velodyne (only for pointpillars)

path_to_testing_folder
|___calib (only for pointpillars)
|___image_2
|___velodyne (only for pointpillars)
        

To obtain subset for COCO:

• The subset generation file is present here. Run this script in your local system (not in Colab as you might have storage limitations)

• Download and unzip 2017 train and val images, 2017 train/val annotations from here

• To obtain subset for training:

  Copy

  Copied!

              
  
              
  python generate_coco_subset.py --source-image-dir=path_to_train2017_folder --source-annotation-file=path_to_instances_train2017.json_file --out-data-dir=path_to_save_subset_data --num-images=num_of_images_in_subset
          

  Example

  Copy

  Copied!

              
  
              
  python generate_coco_subset.py --source-image-dir=/home/user/data/train2017 --source-annotation-file=/home/user/data/annotations/instances_train2017.json --out-data-dir=/home/user/subset_data/ --num-images=100
          

• To obtain subset for validation:

  Copy

  Copied!

              
  
              
  python generate_coco_subset.py --source-image-dir=path_to_val2017_folder --source-annotation-file=path_to_instances_val2017.json_file --out-data-dir=path_to_save_subset_data --num-images=num_of_images_in_subset
          

  Example

  Copy

  Copied!

              
  
              
  python generate_coco_subset.py --source-image-dir=/home/user/data/val2017 --source-annotation-file=/home/user/data/annotations/instances_val2017.json --out-data-dir=/home/user/subset_data/ --num-images=100
          

Dataset folder structure for coco:

            

            
folder_into_which_downloaded_coco_files_are_unzipped
|___train2017
|___val2017
|___annotations
    |___instances_train2017.json
    |___instances_val2017.json
        

1. Mount the drive in the Colab Instance.

2. Click on the folder icon (shown within the green box)

3. Click on the ‘move up one folder’ icon (shown within the green box)

• Now you can view all folders that are tied to this instance.

• The google drive you mounted is present at content->drive->MyDrive.

• All the folders listed in the below screenshot are folders in the root directory “/”.

• Conversational AI pretrained backbones that use Megatron can’t run on Colab/ColabPro instances because they need more than 24GB of GPU RAM.

• The following TAO networks can’t run on Colab until Colab has updated its drivers to 515 or above: Detectnetv2, BPNET, FasterRCNN, FPNet, and Gazenet.

• Any folder that is not in /content/drive/MyDrive/ will be a volatile path: You’ll lose access to the contents when the session disconnects

• If you have a 15GB drive storage limit, try setting the data folder to /content/data/data_you_are_downloading instead of using the drive at /content/drive/MyDrive.

  • This is preferred because volatile storage is greater than 15G.B

  • You can also upload a file downloaded on your system to /content/data using the drag-and-drop method.

• The browser tab needs to be active for the duration of a Colab session. To keep the Colab instance active in the background, you can use a third party extension like Colab Alive.

• In Tensorflow notebooks, to maximize GPU usage, you can switch to GPU hardware only after completing the data download and preprocessing steps that doesn’t require a GPU:

  • At the point of switching to a GPU instance, it is fine to terminate the previous instance.

  • You can keep running successive cells without rerruning any of the previous cells, with the following exceptions:

    • If some data was saved in the volatile storage instead of the Google Drive storage, then switching instances will delete the data.

    • In such cases, the data must be backed up to Google Drive before terminating. This can be achieved with a simple Linux copy command: cp path_to_volatile_storage /content/drive/MyDrive/folder_to_be_copied_into

• With the free version of Colab, GPU instances can run a maximum of 8 hours.

• With the free version of Colab, only one GPU instance notebook can run at a time.

  • Colab Pro can run two GPU instance notebooks simultaneously.

• Colab does not allow you to choose the GPU type.

  • As of November 2022, the predominant GPU is a Tesla T4 with the free version and a P100 with Colab Pro.