EM Loader

em-loader loads an spectrogram dataset into a tf.data.Dataset where each element is a tuple of Dense Tensors representing images and RaggedTensors representing the bounding boxes contained in the images.

em-loader supports all of the bounding box formats available in KerasCV, and loads bounding box Tensors as RaggedTensors by default. This format natively fits the format required by the KerasCV object detection API.

em-loader requires use of the Kaggle API.

Quickstart

Getting started with the em_loader loader is as easy as:

git clone https://github.com/lukewood/em-loader
cd em-loader
python setup.py develop

dataset = em_loader.load(
  bounding_box_format="xywh",
  split="train",
  batch_size=16
)

And fitting a model to the dataset is as easy as:

model = keras_cv.models.RetinaNet(
    classes=1,
    bounding_box_format="xywh",
    backbone="resnet50",
    backbone_weights="imagenet",
    include_rescaling=True,
)

API

The em-images API supports the following arguments:

• bounding_box_format: any KerasCV supported bounding box format, specified in the Keras documentation
• data_dir: the directory holding the data. If you do not have a data directory, please follow the Setup instructions.
• batch_size: batch size to use

Roadmap

• Guided setup (i.e. if the data directory is missing, give the user commands to run to set it up in Error messages)
• End to end Jupyter notebook using the loader in examples/
• Publish package on PyPi