mxnet dataiters.

Providing the mxnet dataiters as Stream or Conduit

mxnet-dataiter

Here is an example of making a Conduit from MNIST dataset.

mnistIter (add @"image" "data/train-images-idx3-ubyte" $ 
           add @"label" "data/train-labels-idx1-ubyte" $
           add @"batch_size" 128 
           nil) :: ConduitData IO (NDArray Float, NDArray Float)

The first argument is provides named parameters for the MXNet Data Iterators. Detailed specification can be found in MXNet API 's python document.

Below is a snapshot of current support in this package.

type CSVIter_Args = 
    '[ "data_csv" := String, "data_shape" := [Int], "label_csv" := String, "label_shape" := [Int]
     , "batch_size" := Int, "round_batch" := Bool, "prefetch_buffer" := Integer, "dtype" := String] 

type MNISTIter_Args = 
    '[ "image" := String, "label" := String, "batch_size" := Int, "shuffle" := Bool, "flat" := Bool
     , "seed" := Int, "silent" := Bool, "num_parts" := Int, "part_index" := Int
     , "prefetch_buffer" := Integer, "dtype" := String]

type ImageRecordIter_Args = 
    '[ "path_imglist" := String, "path_imgrec" := String, "path_imgidx" := String, "aug_seq" := String
     , "label_width" := Int, "data_shape" := [Int], "preprocess_threads" := Int, "verbose" := Bool
     , "num_parts" := Int, "part_index" := Int, "shuffle_chunk_size" := Integer
     , "shuffle_chunk_seed" := Int, "shuffle" := Bool, "seed" := Int, "batch_size" := Int
     , "round_batch" := Bool, "prefetch_buffer" := Integer, "dtype" := String, "resize" := Int
     , "rand_crop" := Bool, "max_rotate_angle" := Int, "max_aspect_ratio" := Float
     , "max_shear_ratio" := Float, "max_crop_size" := Int, "min_crop_size" := Int
     , "max_random_scale" := Float, "min_random_scale" := Float, "max_img_size" := Float
     , "min_img_size" := Float, "random_h" := Int, "random_s" := Int, "random_l" := Int, "rotate" := Int
     , "fill_value" := Int, "inter_method" := Int, "pad" := Int, "mirror" := Bool, "rand_mirror" := Bool
     , "mean_img" := String, "mean_r" := Float, "mean_g" := Float, "mean_b" := Float, "mean_a" := Float
     , "std_r" := Float, "std_g" := Float, "std_b" := Float, "std_a" := Float, "scale" := Float
     , "max_random_contrast" := Float, "max_random_illumination" := Float]
     
type ImageDetRecordIter_Args = 
    '[ "path_imglist" := String, "path_imgrec" := String, "aug_seq" := String, "label_width" := Int
     , "data_shape" := [Int], "preprocess_threads" := Int, "verbose" := Bool, "num_parts" := Int
     , "part_index" := Int, "shuffle_chunk_size" := Integer, "shuffle_chunk_seed" := Int
     , "label_pad_width" := Int, "label_pad_value" := Float, "shuffle" := Bool, "seed" := Int
     , "batch_size" := Int, "round_batch" := Bool, "prefetch_buffer" := Integer, "dtype" := String
     , "resize" := Int, "rand_crop_prob" := Float, "min_crop_scales" := [Float]
     , "max_crop_scales" := [Float], "min_crop_aspect_ratios" := [Float]
     , "max_crop_aspect_ratios" := [Float], "min_crop_overlaps" := [Float], "max_crop_overlaps" := [Float]
     , "min_crop_sample_coverages" := [Float], "max_crop_sample_coverages" := [Float]
     , "min_crop_object_coverages" := [Float], "max_crop_object_coverages" := [Float]
     , "num_crop_sampler" := Int, "crop_emit_mode" := String, "emit_overlap_thresh" := Float
     , "max_crop_trials" := [Int], "rand_pad_prob" := Float, "max_pad_scale" := Float
     , "max_random_hue" := Int, "random_hue_prob" := Float, "max_random_saturation" := Int
     , "random_saturation_prob" := Float, "max_random_illumination" := Int
     , "random_illumination_prob" := Float, "max_random_contrast" := Float, "random_contrast_prob" := Float
     , "rand_mirror_prob" := Float, "fill_value" := Int, "inter_method" := Int, "resize_mode" := String
     , "mean_img" := String, "mean_r" := Float, "mean_g" := Float, "mean_b" := Float, "mean_a" := Float
     , "std_r" := Float, "std_g" := Float, "std_b" := Float, "std_a" := Float, "scale" := Float]

type ImageRecordUInt8Iter_Args = 
    '[ "path_imglist" := String, "path_imgrec" := String, "path_imgidx" := String, "aug_seq" := String
     , "label_width" := Int, "data_shape" := [Int], "preprocess_threads" := Int, "verbose" := Bool
     , "num_parts" := Int, "part_index" := Int, "shuffle_chunk_size" := Integer
     , "shuffle_chunk_seed" := Int, "shuffle" := Bool, "seed" := Int, "batch_size" := Int
     , "round_batch" := Bool, "prefetch_buffer" := Integer, "dtype" := String, "resize" := Int
     , "rand_crop" := Bool, "max_rotate_angle" := Int, "max_aspect_ratio" := Float
     , "max_shear_ratio" := Float, "max_crop_size" := Int, "min_crop_size" := Int
     , "max_random_scale" := Float, "min_random_scale" := Float, "max_img_size" := Float
     , "min_img_size" := Float, "random_h" := Int, "random_s" := Int, "random_l" := Int, "rotate" := Int
     , "fill_value" := Int, "inter_method" := Int, "pad" := Int]

type LibSVMIter_Args = 
    '[ "data_libsvm" := String, "data_shape" := [Int], "label_libsvm" := String, "label_shape" := [Int]
     , "num_parts" := Int, "part_index" := Int, "batch_size" := Int, "round_batch" := Bool
     , "prefetch_buffer" := Integer, "dtype" := String]