A handy image preprocessing pipeline.

import imgflow

dataset = imgflow.fromDir(
    input_dir="data/object-detection-dataset", 
    dataset_format=imgflow.DATASET_COCO,
    img_format=imgflow.IMAGE_JPEG
)

dataset = imgflow.resize(width=1024, height=1024)(dataset)
subsets = imgflow.train_test_split(train_val_test_percent=(0.8, 0.1, 0.1))(dataset)

for subset in subsets:
    imgflow.convert(
        output_file="data/tfrecords/xxx.tfrecord",
        output_format=imgflow.CONVERT_TFRECORD
    )(subset).execute()

• DAG based lazy or eager execution.

• Convert ImgFlow into Cloud Dataflow and run in parallel.

• Cache on local storage.

• Global config.py.

• Multiple data source, e.g. local directory, cloud storage bucket.

• Classic object detection dataset format, e.g. MS COCO, PASCAL VOC.

• Custom dataset parser and loader.

• Attach shapes to images, e.g. bboxes.
• Attach extra info to images, e.g. labels.

• Basic stat anaylsis.
• Basic preprocessing, e.g. resize, pad, grayscale, RGB/BGR.
• Basic augmentations.
• Visualize bboxes, labels, etc.

• Save all intermediate results for manual validation.
• Export the processed dataset to JSON, MS COCO, PASCAL VOC, TFRecord format.

• Serialize/deserialize the pipeline.
• Model inference operation.
• Flask API integration.
• Dockerization.

• Smart dataset loading, automatically detect the directory hierarchy.