Streaming Data

Streaming Export Introduction

Classically annotations were always exported into a static file. For example a JSON file.
Diffgram supports this: Export Walkthrough.

There is another new way. A new tool in your toolkit.

That's Streaming Data. Instead of generating an export as a one off thing, and then manually mapping that file, you stream the data directly into memory.

If you are just getting started we encourage you to consider starting with Streaming, becuase then you can more easily expand into large production scale use cases. Exporting the file still works fine too, so alternatively just awareness of the availability of streaming can be good.

Streaming Benefits

  1. Load only what you need, which may be a small percent of a JSON file.
    At scale it may be impractical to load all the data into a JSON file and so this may be a major benefit.
  2. Works better for large teams. Avoids waiting for static files. You can program and do work with the expected dataset before Annotation starts, or even while Annotation is happening.
  3. More memory efficient - becuase it's streaming, you need not ever load the entire dataset into memory. This is especially applicable for distributed training, and when marshalling a JSON file would be impractical on a local machine.
  4. Saves having "double mapping" e.g. mapping to another format, which will itself then be mapped to tensors. In some cases too parsing a JSON file can take even more effort then just updating a few tensors.
  5. More flexibility, no need for a pre-defined "format".

Streaming Drawbacks

  1. Defined in code. If the dataset changes may effect reproducibility unless additional steps are taken.
  2. Requires understanding mapping the tensors.
  3. Requires a network connection.
  4. Some legacy training systems / AutoML providers may not support loading directly from memory and may require static files.

Comparison Table



Use cases

Small Datasets

Large Datasets


All at once

Per item (e.g per image)


Load and map JSON file

Map tensors


Load all data first,
then construct query after the fact






!pip install diffgram

from diffgram import Project

# Public project example
coco_project = Project(project_string_id='coco-dataset')
default_dataset = = 'Default')
# Let's see how many images we got
print('Number of items in dataset: {}'.format(len(default_dataset)))

# Let's stream just the 8th element of the dataset
print('8th element: {}'.format(default_dataset[7]))

pytorch_ready_dataset = default_dataset.to_pytorch()

# EXAMPLE tensor mapping
# See AdjustForRCNN() example below
pytorch_ready_dataset = default_dataset.to_pytorch(
    transform = T.Compose([AdjustForRCNN()])
Example of a training data sample: a Tupple where first elm is the image and second elm is a dict with the label data.

(<PIL.Image.Image image mode=RGB size=650x417 at 0x7FE20F358908>,  
 {'area': tensor([141120.]),
 'boxes': tensor([[ 81.,  88., 522., 408.]]),   
 'image_id': tensor([0]),   
 'iscrowd': tensor([0]),   
 'labels': tensor([1])}

Notebook Example

Notebook example

Example Process

  1. Query the data desired with slice()
  2. Map tensors as required with to_pytorch() or to_tensorflow()
  3. Run it!

Tensors Mapping Example

See the notebook for complete walkthrough Notebook example.

def get_transform():
  transforms = []
  transforms.append(AdjustForRCNN()) # Our previously created transform
  # converts the image, a PIL image, into a PyTorch Tensor
  return ComposeTransforms(transforms)

class AdjustForRCNN:
      Adjust a sample from the Diffgram SDK generated pytorch
      Dataset to be used for the RCNN model.


    def __call__(self, sample):
        result = []
        image = sample['image']
        image_id = torch.tensor([sample['diffgram_file'].id])
        project = sample['diffgram_file'].client
        project.get_label_file_dict(use_session = False)
        label_id_name_map = coco_project.name_to_file_id
        label_id_list = [val for k, val in label_id_name_map.items()]
        boxes_array = []
        for i in range(len(sample['x_min_list'])):
            # Let's build the box array with the shape
            # [x_min,y_min,x_max, y_max]
        boxes = torch.as_tensor(boxes_array, dtype = torch.float32)
        num_objs = len(boxes_array)
        labels_list = sample['label_id_list']
        label_numbers_list = [label_id_list.index(x) for x in labels_list]
        labels = torch.as_tensor(label_numbers_list, dtype = torch.int64)
        if num_objs > 0:
            area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0])
            area = torch.as_tensor([], dtype = torch.float32)
        # suppose all instances are not crowd for this tutorial
        iscrowd = torch.zeros((num_objs,), dtype = torch.int64)
        target = {}
        target["boxes"] = boxes
        target["labels"] = labels
        target["image_id"] = image_id
        target["area"] = area
        target["iscrowd"] = iscrowd
        image = result[0]
        target = result[1]
        return image, target


There's two main reasons

  1. It's a requirement for large use cases

  2. Diffgram supports all media types, attributes etc. Even within the most well defined problem spaces, such as bounding box for images, there are multiple popular formats. When you consider say event level attributes on a Video the concept of a "format" becomes undefined. By refocusing the access around queries, the SDK, and so forth it help provide a best in class experience regardless of the complexity of the case.

SDK functions


Related Material

Accessing Media
Notebook example

Did this page help you?