Parameter Documentation

The following parameters are exposed to Python and Caffe.

  • General Streaming Parameters

    • int32 seed: randomization seed
    • bool loop: loop through records for endless streaming, instead of throwing StopIteration when all data are consumed.
    • bool shuffle: shuffle input records order before splitting the data.
    • bool reshuffle: shuffle input records in the beginning of every loop from the second loop on.
    • bool stratify: divide data by category, split and loop within each category, so any sliding window on the stream contain about same number of images for each category.
    • bool cache: cache all data in memory after they are read. Default is true, set to false when data do not fit in memory.
    • uint32 preload: pre-load queue size.
    • uint32 threads: number of decoding/pre-processing threads, default = 4.
    • uint32 batch: batch size
    • bool pad: if loop is false and there's not enough to produce a whole batch, throw StopIteration if not pad, otherwise return a whole batch with random padding. Number of padded samples are also returned.

  • Cross validation

    • uint32 split: equally split all data (or each category if stratify = true) into this number of partitions.
    • int32 split_fold: exclude examples from this split.
    • bool split_negate: use the rest of samples.

For example, 5-fold cross validation can be run with split = 5, and with split_fold = 0, 1, 2, 3, 4. For each value of split_fold, set split_netage = false when training, and true when validating.

  • Image Control

    • int32 channels: number of image channels, ignored if <= 0
    • int32 min_size:
    • int32 max_size:
    • int32 resize_width: resize to this, ignored if <= 0
    • int32 resize_height: resize to this, ignore if <= 0
    • int32 decode_mode: as in the second parameter of cv::imread. Set to -1 to load images as is.
    • bool bgr2rgb: default 3-channel images are stored as BGR (opencv behavior), set to true for RGB.

  • Image Annotation

    • string annotate: for annotation, set to "json" or "image"
    • int32 anno_type: annotation image type, as in cv::Mat::type()
    • bool anno_copy: copy image to label, for visualization. Set to false for training.
    • float anno_color1: annotation channel 1.
    • float anno_color2: annotation channel 2, typically not used.
    • float anno_color3: annotation channel 3, typically not used.
    • int32 anno_thickness: set to -1 to fill the annotation region with anno_color.

  • Image Augmentation

    • bool perturb: set to true to enable augmentation
    • float pert_color1:
    • float pert_color2
    • float pert_color3
    • float pert_angle: image/label rotated randomly within [-pert_angle, pert_angle], in degrees.
    • float pert_min_scale: image/label rescaled to a scale uniformly sampled from [pert_min_scale, pert_max_scale].
    • float pert_max_scale
    • bool pert_hflip: enable random horizontal flip.
    • bool pert_vflip: enable random vertical flip.

  • Label Encoding

    • uint32 onehot: one hot encoding of label. If onehot > 0, specify the number of categories.