Intel Gaudi Media Loader

The habana_media_loader is used to dataload and pre-process inputs for deep learning frameworks. It consists of pre-enabled dataloaders for commonly used datasets and building blocks to assemble a generic dataloader. The loader decides internally if part of the operations can be offloaded to the Intel® Gaudi® AI accelerator. If the offload cannot be executed, it uses an alternative passing dataloader function to run.

habana_media_loader can operate in different modes. The optimal one is selected based on the underlying hardware:

  • In Intel® Gaudi® 2 AI accelerator, the dataloader uses hardware-based decoders for acceleration, lowering the load on the host CPU.

  • In first-gen Intel® Gaudi® AI accelerator, it uses either the default PyTorch Dataloader or habana_dataloader, depending on the use case. Both are done on the host CPU.

Setting Up the Environment

To install habana_media_loader, run the following command:

pip install habana_media_loader-1.17.1-40-py3-none-any.whl

Note

The above step is not required when running Intel Gaudi Docker image as habana_media_loader is already installed by default.

Using Media Loader with PyTorch

Follow the steps below to import habana_dataloader object. For the full example, refer to Torchvision model.

  1. Import habana_dataloader object.

    import habana_dataloader
    
  2. Create an instance of habana_dataloader object.

    habana_dataloader.HabanaDataLoader(
        dataset, batch_size=args.batch_size, sampler=train_sampler,
        num_workers=args.workers, pin_memory=True, drop_last=True)
    

The Intel Gaudi software selects the dataloader based on the underlying hardware and the dataset used:

  • In Gaudi 2, it uses hardware acceleration for ImageNet, COCO and Medical Segmentation Decathlon (BraTS) datasets.

  • In first-gen Gaudi, it uses software acceleration for ImageNet and COCO datasets.

Fallback

  • In Gaudi 2 - When the provided input parameters are not eligible for hardware acceleration (see Guidelines for Supported Datasets), the software accelerated habana_dataloader is activated. In such a case, the following message will be printed:

    Failed to initialize Habana media Dataloader, error: {error message}
    Fallback to aeon dataloader
    
  • In first-gen Gaudi - When the provided input parameters are not eligible for habana_dataloader (see Guidelines for Supported Datasets), the default PyTorch Dataloader is initialized and used. In such a case, the following message will be printed:

    Failed to initialize Habana Dataloader, error: {error message}
    Running with PyTorch Dataloader
    

Guidelines for Supported Datasets

The following lists the restrictions for the supported datasets using Gaudi 2:

  • Acceleration takes place only with the following parameters:

    • prefetch_factor=3

    • drop_last=False

    • dataset is torchvision.datasets.ImageFolder

  • The dataset should contain only .jpg or .jpeg files.

  • Acceleration can take place only with the following dataset torchvision transforms packed as transforms.Compose in Torchvision script/train.py.

    • RandomResizedCrop

    • CenterCrop

    • Resize

    • ToTensor

    • RandomHorizontalFlip, only with p=0.5

    • Normalize, only with mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]

  • Acceleration takes place only with the following parameters:

    • dataset is an instance of COCODetection. See SSD script/train.py.

    • drop_last=False

    • prefetch_factor=3

  • The dataset should be taken from the COCO Dataset webpage.

  • Acceleration can take place only with the following dataset transforms packed as transforms.Compose in SSD script/utils.py.

    • SSDCropping. See SSD script/train.py.

    • Resize

    • ColorJitter, only with brightness=0.125, contrast=0.5, saturation=0.5, hue=0.05

    • ToTensor

    • RandomHorizontalFlip, only with p=0.5

    • Normalize, only with mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]

    • Encoder. See SSD script/train.py.

  • Acceleration takes place only with the following parameters:

    • num_workers is 3 for Unet2D and 5 for Unet3D

    • Unet2D - Only nvol 1 is supported

    • Unet3D - Only batchsize 2 is supported

    • val_dataloader, test_dataloader are not supported by habana_media_loader

  • The dataset should be preprocessed as per script/preprocess.py.

  • Acceleration can take place only with the following dataset transforms:

    • Crop

    • Flip

    • Noise - Standard deviation range supported (0, 0.33)

    • Blur - Sigma’s range supported (0.5, 1.5)

    • Brightness - Brightness scale supported (0.7, 1.3)

    • Contrast - Contrast scale supported (0.65, 1.5)

    • Zoom transform is not supported

The following lists the restrictions for the supported datasets using first-gen Gaudi:

  • Acceleration takes place only with the following parameters:

  • Acceleration can take place only with the following dataset transforms packed as transforms.Compose in SSD script/utils.py.

    • SSDCropping. See SSD script/train.py.

    • Resize

    • ColorJitter, only with brightness=0.125, contrast=0.5, saturation=0.5, hue=0.05

    • ToTensor

    • RandomHorizontalFlip, only with p=0.5

    • Normalize, only with mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]

    • Encoder. See SSD script/train.py.

Model Examples

The following are full examples of models using Intel Gaudi Media Loader with PyTorch on the datasets mentioned above: