Data Reader is a function for reading input data. You can use 2 ways to set data_reader for dataset:
type:
for setting reader name. This approach gives opportunity to set additional parameters for adapter if it is required.In case, when you have model with several inputs which should use data stored in different format (e. g. images and json) you can use combine_reader
. combine_reader
allows specify reading scheme depends on file names. It use parameter scheme
for describing reading approaches as dictionary where keys are regular expressions for file names, values are reader_name.
AccuracyChecker supports following list of data readers:
opencv_imread
- read images using OpenCV library. Default color space is BGR.reading_flag
- (Optional) flag which specifies the way image should be read: color
- default, loads color image, gray
- loads image in grayscale mode, unchanged
- loads image as such including alpha channel.pillow_imread
- read images using Pillow library. Default color space is RGB.scipy_imread
- read images using similar approach as in scipy.misc.imread
skimage_imread
- read images using scikit-mage library. Default color space is RGB.tf_imread
- read images using TensorFlow. Default color space is RGB. Requires TensorFlow installation.opencv_capture
- read frames from video using OpenCV.json_reader
- read value from json file.key
- key for reading from stored in json dictionary.annotation_features_extractor
- read features from annotation.features
- list of features. All features should be fields of annotation representation.numpy_reader
- read numpy dumped files (npy or npz formats are supported for reading)keys
- comma-separated list of model input namesseparator
- separator symbol between input identifier and file identifierid_sep
- separator symbol between input name and record number in input identifierblock
- block mode (batch - oriented). In this mode reader returns whole variable.numpy_txt_reader
- read data stored in text format to numpy array.numpy_dict_reader
- read and unpack dictionaries saved in numpy files.nifti_reader
- read NifTI data formatchannels_first
- allows read nifti files and transpose in order where channels first (Optional, default False)wav_reader
- read WAV file into NumPy array. Also gets the samplerate.dicom_reader
- read images stored in DICOM format.pickle_reader
- read data stored in pickle file. Supported formats of pickle content:rawpy
- read raw image formats using rawpy library.postprocess
- allow image postprocessing and normalization (Optional, default True
).byte_reader
- read raw binary data and wrap them to numpy-array.