Protocol Documentation

Anchor height in pixels. Default: 256

Anchor width in pixels. Default: 256

Anchor stride in height dimension in pixels. Default: 16

Anchor stride in width dimension in pixels. Default: 16

Anchor height offset in pixels. Default: 0

Anchor width offset in pixels. Default: 0

List of scales for the anchors. 

List of aspect ratios for the anchors. 

minimum level in feature pyramid Default: 2

maximum level in feature pyramid Default: 6

Scale of anchor to feature stride 8 * 4 = 32 Default: 8

Aspect ratios for anchors at each grid point. 

Number of intermediate scale each scale octave Default: 1

Whether to produce anchors in normalized coordinates. Default: true

Number of grid layers to create anchors for. Default: 6

Scale of anchors corresponding to finest resolution. Default: 0.2

Scale of anchors corresponding to coarsest resolution Default: 0.95

Can be used to override min_scale->max_scale, with an explicitly defined
set of scales.  If empty, then min_scale->max_scale is used. 

Aspect ratios for anchors at each grid point. 

When this aspect ratio is greater than 0, then an additional
anchor, with an interpolated scale is added with this aspect ratio. Default: 1

When set true, intepolate scale=sqrt(max_size*min_size), aspect_ratios=1.0
in all layers, otherwise lowest layer will be ignored Default: false

Whether to use the following aspect ratio and scale combination for the
layer with the finest resolution : (scale=0.1, aspect_ratio=1.0),
(scale=min_scale, aspect_ration=2.0), (scale=min_scale, aspect_ratio=0.5). Default: true

 Default: false

The base anchor size in height dimension. Default: 1

The base anchor size in width dimension. Default: 1

Anchor stride in height dimension in pixels for each layer. The length of
this field is expected to be equal to the value of num_layers. 

Anchor stride in width dimension in pixels for each layer. The length of
this field is expected to be equal to the value of num_layers. 

Anchor height offset in pixels for each layer. The length of this field is
expected to be equal to the value of num_layers. 

Anchor width offset in pixels for each layer. The length of this field is
expected to be equal to the value of num_layers. 

Anchor length in pixels. Default: 8

Anchor stride in height dimension in pixels. Default: 8

Anchor height offset in pixels. Default: 0

List of scales for the anchors. 

List of Anchor groups, the number of groups must be equal to
the number of feature maps 

Threshold for positive matches. Default: 0.5

Threshold for negative matches. Default: 0.5

Whether to construct ArgMaxMatcher without thresholds. Default: false

If True then negative matches are the ones below the unmatched_threshold,
whereas ignored matches are in between the matched and umatched
threshold. If False, then negative matches are in between the matched
and unmatched threshold, and everything lower than unmatched is ignored. Default: true

Whether to ensure each row is matched to at least one column. Default: false

Force constructed match objects to use matrix multiplication based gather
instead of standard tf.gather Default: false

Threshold for positive matches. Default: 1

Force constructed match objects to use matrix multiplication based gather
instead of standard tf.gather Default: false

 Default: true

 Default: true

 Default: 0

 Default: true

dropout config, keep_prob of aspp out features Default: 1

Compression mode: one in [`prune`, `quantize`]. Defualt to be `prune`. Default: prune

Whether is finetuning from a compressed model. Default: false

The target speedup ratio. Default to 1.0. Default: 1

Path to pretrained model. 

Auto tuning mode.
For `prune`, one in [`RL`, `random`, `uniform`].
For `quantize`, one in [`KL`, `LSQ`] Default: uniform

Number fo trials for the automatic search of compression strategy. Default: 10

Number of steps for re-training steps of each compressed model. Default: 1000

Metric key to evaluate the model performance. 

Metric mode. Which is better, one on [`bigger`, `smaller`].
Default to be `bigger`. Default: bigger

Graph scopes that would to to be included. 

Graph scopes that would to to be excluded. 

Number of iterations for layer-reconstruction. Default to be 1000. Default: 1000

Initial learning rate for layer-selection. Default to be 1e-10. Default: 1e-10

Learning rate increase ratio for layer-selection. Default to be 1.4. Default: 1.4

Learning rate decrease ratio for layer-selection. Default to be 0.7. Default: 0.7

Learning rate for layer-reconstruction with Adam. Default to be 1e-4. Default: 0.0001

List of op types that can be pruned.
Subset of [`Conv2D`, `MatMul`] 

Base number for remained channels number as a multiple of it. Default to be 4. Default: 4

Graph scopes that would to to be included. 

Graph scopes that would to to be excluded. 

Quantization bits options. Such as [4, 8, 16, 32]. Default: 4

Layer to be kept in INT8 when using INT4 quantization. 

Whether to use per-channel quant for specific op types. 

Number of data batches for calibration. 

Path to save calibration scale file. 

Whether to use per-channel quant for Conv2D. Default: true

Whether to use per-channel quant for MatMul. Default: false

backbone class name, such as resnet_v1_50 

weight decay factor Default: 0.0005

if set False, batchnorm parameters and moving mean/std will not be update Default: true

currently only resnet backbone support this parameter
if output_stride is set greater than 0, when the product of each layer's stride
is equal to output_stride, the stride of the upper conv layers will be set to 1, 
and use dilation conv instead. Default: -1

boolean flag to control the avgpooling before the
logits layer. If false or unset, pooling is done with a fixed window
that reduces default-sized inputs to 1x1, while larger inputs lead to
larger outputs. If true, any input size is pooled down to 1x1. Default: false

depth_multiplier used only for mobilenet, which is used to adjust network for different
computation cost, please refer to https://arxiv.org/abs/1704.04861 Default: 1

when image are padded, use_true_shape need to be set true,
then network will use true shape to global pool Default: true

use fc or not, when the number of parameters are too large in fully connect layers,
use this parameter to drop fc layer in finetune cases. Default: true

normalization layer type Default: BATCH

block connect survival prob when training in efficientnet, default is 0.8 

keep prob for dropout Default: 1

user-defined args 

 Default: false

 Default: 1024

 Default: 2

model class name 

model block name, e.g. block4 

deprecated 

deprecated 

stride of the block Default: 1

deprecated 

weight decay factor Default: 0

Force constructed match objects to use matrix multiplication based gather
instead of standard tf.gather Default: false

Scale factor for anchor encoded box center. Default: 10

 Default: 10

Scale factor for anchor encoded box height. Default: 5

Scale factor for anchor encoded box width. Default: 5

Scale factor for anchor encoded box center and keypoints. Default: 10

 Default: 10

Scale factor for anchor encoded box height. Default: 5

Scale factor for anchor encoded box width. Default: 5

The standard deviation used to encode and decode boxes. Default: 0.01

Scale factor for anchor encoded box center. Default: 10

 Default: 10

Scale factor for anchor encoded box length. Default: 5

Scale factor for anchor encoded box center. Default: 1

 Default: 1

Scale factor for anchor encoded box height. Default: 1

Scale factor for anchor encoded box width. Default: 1

Hyperparameters for convolution ops used in the box predictor. 

Minimum feature map depth prior to predicting box encodings and class
predictions, used collaborately with Default: 0

Maximum feature depth map prior to predicting box encodings and class
predictions. If max_depth is set to 0, no additional feature map will be
inserted before location and class predictions. Default: 0

Number of the additional conv layers before the predictor. Default: 0

Keep probability for dropout Default: 1

Size of final convolution kernel. If the spatial resolution of the feature
map is smaller than the kernel size, then the kernel size is set to
min(feature_width, feature_height). Default: 1

Size of the encoding for boxes. Default: 2

 Default: 0

Whether to use depthwise separable convolution for box predictor layers. Default: false

Hyperparameters for convolution ops used in the box predictor. 

Minimum feature map depth prior to predicting box encodings and class
predictions, used collaborately with Default: 0

Maximum feature depth map prior to predicting box encodings and class
predictions. If max_depth is set to 0, no additional feature map will be
inserted before location and class predictions. Default: 0

Number of the additional conv layers before the predictor. Default: 0

Keep probability for dropout Default: 1

Size of final convolution kernel. If the spatial resolution of the feature
map is smaller than the kernel size, then the kernel size is set to
min(feature_width, feature_height). Default: 1

Size of the encoding for boxes. Default: 4

 Default: 0

Whether to use depthwise separable convolution for box predictor layers. Default: false

Hyperparameters for fully connected ops used in the box predictor. 

Number of the additional fc layers before the predictor. Default: 0

Output depth for the fc ops prior to predicting box encodings
and class predictions. Default: 0

Keep probability for dropout. This is only used if use_dropout is true. Default: 1

Size of the encoding for the boxes. Default: 2

Whether to use one box for all classes rather than a different box for each
class. Default: true

Hyperparameters for fully connected ops used in the box predictor. 

Number of the additional fc layers before the predictor. Default: 0

Output depth for the fc ops prior to predicting box encodings
and class predictions. Default: 0

Keep probability for dropout. This is only used if use_dropout is true. Default: 1

Size of the encoding for the boxes. Default: 4

Whether to use one box for all classes rather than a different box for each
class. Default: true

Hyperparameters for convolution ops used in the box predictor. 

Bin sizes for RFCN crops. Default: 3

 Default: 3

Target depth to reduce the input image features to. Default: 1024

Size of the encoding for the boxes. Default: 4

Size to resize the rfcn crops to. Default: 12

 Default: 12

 Default: true

Hyperparameters for convolution ops used in the box predictor. 

Number of the additional conv layers before the predictor. Default: 0

Output depth for the convolution ops prior to predicting box encodings
and class predictions. Default: 0

Size of final convolution kernel. If the spatial resolution of the feature
map is smaller than the kernel size, then the kernel size is set to
min(feature_width, feature_height). Default: 3

Size of the encoding for boxes. Default: 4

Bias initialization for class prediction. It has been show to stabilize
training where there are large number of negative boxes. See
https://arxiv.org/abs/1708.02002 for details. Default: 0

Keep probability for dropout Default: 1

Whether to share the multi-layer tower between box prediction and class
prediction heads. Default: true

Whether to use depthwise separable convolution for box predictor layers. Default: false

 Default: true

Hyperparameters for convolution ops used in the box predictor. 

Number of the additional conv layers before the predictor. Default: 1

input width height, if not set, will use default input size instead 

Backbone configuration 

Number of classes 

Loss configuration for training 

Whether to summary training related info Default: true

label_id offset, will be used to subtract from groundtruth class
when calcuating loss amd evaluation Default: 0

hidden size of last fc, if assigned, original fc will be replaced Default: -1

Backbone configuration 

input layer name 

Number of classes 

Loss configuration for training 

label_id offset, will be used to subtract from groundtruth class
when calcuating loss amd evaluation Default: 0

global pooling after backbone Default: true

dense layer hidden size appended before logits 

Number of model variables level to be trainable
set to 0 or -1: all variables are trainable Default: -1

separator between classname and description, for example:
family_name#QING, # is the separator 

when there are no matched labels in class_map, 
and default_class is set, the label will be matched
default class 

all image of error_class, or has objects of error_class 
will be discarded 

all objects of ignore_class will be ignored
ignore_class is only used in detection tfrecord conversion 

specified a map from label to class_name
if class_name is not specified, class_name == label_name 

max height/width of all the images used in train
if not specified, its default value is 0,
images will not be resized to within max_image_size
these will result in slow training speed or even
OOM(out ouf memory) error Default: 0

 Default: jpg

 Default: CLASSIFICATION

converter class, some converter are already implemented, such
as QinceConverter,  user can also pass self-defined converter
using format like module.class_name Default: QinceConverter

number of generate process Default: 10

path to .osscredentials file 

custom key value parameters 

a csv contain two column ["original", "replaced"] for replace
special chars in text, such as complex chinese character to
simple chinese character, chinese punctuation to english punctuation etc. 

a txt contain a list of characters for the characters used in
model training, each character for one line.
if the default_char_dict_path is empty, output_char_dict
will infer from the input data 

number of parallel prefetch thread Default: 10

number of parallel write thread Default: 1

The number of samples in each parts of tfrecords:
If Mod(total_num, part_record_num) < part_record_num / 2:
  the rest samples are pad to the end of each tfrecords
  although the size is large than part_record_num;
If Mod(total_num, part_record_num) > part_record_num / 2:
  the rest samples will be placed into a new tfrecord Default: -1

train/test dataset split ratio of test dataset Default: 0

max image size of test dataset images, if not set, it will use max_image_size 

video decode parameters
decode type Default: 4

sample rate, default -1, full sampling. Default: -1

output size of decoded frames， -1 means no resize
if scalar is provided, height=width
otherwise output frame size is (heigh, width) Default: 112

 Default: 112

batch size of each decode phase. Default: 10

left size of last decode phase. Default: 0

flow calc algo
'' means not calculate optical flow
opencv means calculate optical flow using opencv
tvnet means calculate optical flow using tvnet 

minimum bounding box size, bounding box size less than this value will be filtered Default: 5

file path for self defined converter 

file path for self-defined generator 

class name for generator 

If false, do not rotate the image according to EXIF's orientation flag. Default: false

all objects of ignore recog class will not be recognized in TextEnd2End model 

size of input queue, one thread read .csv
file and feed input data into input queue Default: 1048576

size of prefetch queue, multithread thread prefetch
file data and feed into input queue Default: 1024

size of output queue, subproc write serialized
tf example into output queue, and main queue
acquire the data from output queue Default: 1024

padding cropped roi image with multiple of roi short edge Default: 0.75

padding cropped roi image with multiple of roi short edge Default: 0.75

label task id for PaiConverter 

marked class 

the class used in tf record 

label map path: specifying the mapping from class_name to class_ids 

label map path: specifying the mapping from class_name to class_ids 

 Default: false

dataset input path, support pattern filename patterns
 using tf.match_files(input_path) 

Effective batch size to use for training. Default: 32

Data augmentation options. 

whether to shuffle data Default: true

Buffer size to be used when shuffling. Default: 2048

Buffer size to be used when shuffling file names. Default: 100

The number of times a data source is read. If set to zero, the data source
will be reused indefinitely. Default: 0

Number of reader instances to create. Default: 1

Number of records to read from each reader at once. Default: 32

Number of decoded records to prefetch before batching. Default: 512

Number of parallel decode ops to apply. Default: 64

whether to use difficult samples Default: true

shard dataset to 1/num_workers in distribute mode Default: false

whether the last batch should be dropped in the case it has
fewer than batch_size elements Default: true

bucketing size for height and width of images, default is empty, no bucketing

specific settings to each of the dataset, such as voc
will extend to dataset in the future 

input class name if want to direct use one input class 

key points number Default: 4

label map path: specifying the mapping from class_name to class_ids 

dict_path: specifying the char dict 

transform label to upper case Default: false

key points number Default: 4

label map path: specifying the mapping from class_name to class_ids 

dict_path: specifying the char dict 

specify the maximal width/height of all the training images 

specify the minimal width/height of all the training images 

put data into similar-length buckets 

transform label to upper case Default: false

filter image with aspect ratio > max_input_ratio Default: true

specify the maximal text length of all the training images 

label map path: specifying the mapping from class_name to class_ids 

load optical flow or rgb frame
'rgb', 'flow', 'rgb+flow' Default: rgb

load multilabel data or not Default: false

label map path: specifying the mapping from class_name to class_ids 

Whether to load groundtruth instance masks. Default: false

Type of instance mask. Default: NUMERICAL_MASK_FORMAT

Number of groundtruth keypoints per object. Default: 0

embedding size Default: 256

decoder depth Default: 2

whether to add residual connections Default: true

beam width when using beam search decoder. If 0 (default), use standard  decoder with greedy helper Default: 0

length penalty for beam search Default: 0

the probability of sampling from the outputs instead of reading directly from the inputs when training Default: 0

attention mechanisms luong | scaled_luong | bahdanau | normed_bahdanau Default: normed_bahdanau

number of attention heads Default: 1

whether use attention as the cell output at each timestep Default: true

Visualize attentions type or not. choice: line | spatial 

whether to pass encoder's rnn state to decoder Default: true

attention type line | spatial Default: line

number of encoder layers 

hidden units size 

number of attention heads 

hidden size of FeedForwardLayer 

postprocess layer dropout Default: 0.1

attention layer dropout Default: 0.1

relu layer dropout Default: 0.1

beam search width Default: 1

length penalty for beam search Default: 0

Number of visualization images to generate. Default: 10

Number of examples to process of evaluation. Default: 0

How often to run evaluation. Default: 300

Maximum number of times to run evaluation. If set to 0, will run forever. Default: 0

Whether the TensorFlow graph used for evaluation should be saved to disk. Default: false

Path to directory to store visualizations in. If empty, visualization
images are not exported (only shown on Tensorboard). 

BNS name of the TensorFlow master. 

Type of metrics to use for evaluation.
possible values: 
  pascal_voc_detection_metrics
  pascal_voc07_detection_metrics
  coco_detection_metrics 

Path to export detections to COCO compatible JSON format. 

Option to not read groundtruth labels and only export detections to
COCO-compatible JSON file. Default: false

Use exponential moving averages of variables for evaluation.
TODO(rathodv): When this is false make sure the model is constructed
without moving averages in restore_fn. Default: false

Whether to evaluate instance masks.
Note that since there is no evaluation code currently for instance
segmenation this option is unused. Default: false

Minimum score threshold for a detected object box to be visualized Default: 0.5

Maximum number of detections to visualize Default: 20

When drawing a single detection, each label is by default visualized as
<label name> : <label score>. One can skip the name or/and score using the
following fields: Default: false

 Default: false

Whether to show groundtruth boxes in addition to detected boxes in
visualizations. Default: false

Box color for visualizing groundtruth boxes. Default: black

Whether to keep image identifier in filename when exported to
visualization_export_dir. Default: false

Whether to retain original images (i.e. not pre-processed) in the tensor
dictionary, so that they can be displayed in Tensorboard. Default: true

If True, additionally include per-category metrics. Default: false

If True, will open coco analyze function Default: false

iou threshold used for evaluation Default: 0.5

If True, additionally include per-dataset metrics. Default: false

when include_metrics_per_dataset is true, eval dataset in this dataset_names 

batch size used for exported model, -1 indicates batch_size is None
which is only supported by classification model right now, while 
other models support static batch_size Default: -1

type of exporter [final | latest | none] when train_and_evaluation
final: performs a single export in the end of training
latest: regularly exports the serving graph and checkpoints
none: do not perform export Default: final

type if color format [bgr | rbg] Default: rgb

whether export preprocess graph Default: false

custom defined parameters 

backbone config 

rpn head config 

block reuse part of backbone to extract box feature in second stage 

rcnn head config 

rmask head config 

 Default: 256

Param extra_conv_layers are used to extend feature maps beyond backbone,
 so that larger anchors(larger than 256) could be placed on more coarsed
 features(stride>=64).
When param retina_net is set to true, then will use strided convolution(s=2).
For fpn, extra_conv_layers = 1, which means that the fpn feature maps
 will be P2(C2) P3(C3) P4(C4) P5(C5) P6(P5 pooled).
C2, C3, C4, C5 are backbone feature maps of level 2, 3, 4, 5, such as
 resnet/block1, resnet/block2, resnet/block3, resnet/block4.
The anchors placed on PXs will be 32,64,128,256,512. Default: 0

 Default: false

 Default: BILINEAR

level refers to feature map indices, level is associated with feature
map strides = 2 ^ level, usually:
  feature_map                level       stride
    conv1                      1            2
    conv2(resnet/block1)       2            4
    conv3(resnet/block2)       3            8
    conv4(resnet/block3)       4           16
    conv5(resnet/block4)       5           32
roi_min_level: refers to the roi level of lowest fpn feature map
  example: resnet-50/block1 => 2 Default: 2

roi_max_level: refers to the roi level of highest fpn feature map
  example: resnet-50/block4 => 5 Default: 5

roi_canonical_scale and roi_canonical_level specified the parameters
used in distribute proposals to feature maps:
    k = floor(k0 + log2(sqrt(wh)/224))
here, roi_canonical_scale = k0, roi_canonical_level = 224
see (https://arxiv.org/abs/1612.03144) for details. Default: 4

 Default: 224

Hyperparameters for convolution ops used in fpn. 

Number of steps to delay before quantization takes effect during training. Default: 500000

Number of bits to use for quantizing weights.
Only 8 bit is supported for now. Default: 8

Number of bits to use for quantizing activations.
Only 8 bit is supported for now. Default: 8

 Default: 0.999

 Default: true

 Default: false

 Default: 0.001

Whether to train the batch norm variables. If this is set to false during
training, the current value of the batch_norm variables are used for
forward pass but they are never updated. Default: true

 Default: CONV

Regularizer for the weights of the convolution op. 

Initializer for the weights of the convolution op. 

 Default: RELU

BatchNorm hyperparameters. If this parameter is NOT set then BatchNorm is
not applied! 

Whether depthwise convolutions should be regularized. If this parameter is
NOT set then the conv hyperparams will default to the parent scope. Default: false

 Default: 1

 Default: 1

 Default: 0

 Default: 1

 Default: 0

 Default: 1

 Default: 2

 Default: false

 Default: FAN_IN

 Default: true

Hyperparameters for convolution ops used in the keypoint predictor. 

Hyperparameters for fc ops used in the keypoint predictor. 

Number of resnet block before keypoint predictor, we use down sampling between block Default: 1

Number resnet units per resnet block Default: 1

The depth of first resnet block Default: 256

The rate of squeeze_and_excitation, less and equal than zeros for disable Default: 0

The number of fc layers before predictor Default: 2

The depth of fc layers Default: 1024

Interpolation weight between 0 and 1. 

Whether hard boot strapping should be used or not. If true, will only use
one class favored by model. Othewise, will use all predicted class
probabilities. Default: false

DEPRECATED, do not use.
Output loss per anchor. Default: false

Maximum number of hard examples to be selected per image (prior to
enforcing max negative to positive ratio constraint).  If set to 0,
all examples obtained after NMS are considered. Default: 64

Minimum intersection over union for an example to be discarded during NMS. Default: 0.7

 Default: BOTH

Maximum number of negatives to retain for each positive anchor. If
num_negatives_per_positive is 0 no prespecified negative:positive ratio is
enforced. Default: 0

Minimum number of negative anchors to sample for a given image. Setting
this to a positive number samples negatives in an image without any
positive anchors and thus not bias the model towards having at least one
detection per image. Default: 0

Localization loss to use. 

Classification loss to use. 

If not left to default, applies hard example mining. 

Classification loss weight. Default: 1

Localization loss weight. Default: 1

If not left to default, applies random example sampling. 

The desired fraction of positive samples in batch when applying random
example sampling. Default: 0.01

DEPRECATED, do not use. Default: false

modulating factor for the loss. Default: 2

alpha weighting factor for the loss. 

use label smoothing in loss
please refer to label_smoothing explanation in tf.losses.sigmoid_cross_entropy Default: 0

iou type [iou/giou/diou/ciou] Default: iou

DEPRECATED, do not use.
Output loss per anchor. Default: false

DEPRECATED, do not use.
Output loss per anchor. Default: false

use label smoothing in loss
please refer to label_smoothing explanation in tf.losses.sigmoid_cross_entropy Default: 0

DEPRECATED, do not use.
Output loss per anchor. Default: false

Delta value for huber loss. Default: 1

DEPRECATED, do not use. Default: false

Scale and softmax groundtruth logits before calculating softmax
classification loss. Typically used for softmax distillation with teacher
annotations stored as logits. Default: 1

DEPRECATED, do not use.
Output loss per anchor. Default: false

Scale logit (input) value before calculating softmax classification loss.
Typically used for softmax distillation. Default: 1

use label smoothing in loss
please refer to label_smoothing explanation in tf.losses.sigmoid_cross_entropy Default: 0

Hyperparameters for convolution ops used in the box predictor. 

The depth for the first conv2d_transpose op applied to the
image_features in the mask prediction branch. If set to 0, the value
will be set automatically based on the number of channels in the image
features and the number of classes. Default: 256

The height and the width of the predicted mask. Default: 15

 Default: 15

The number of convolutions applied to image_features in the mask prediction
branch. Default: 2

 Default: false

Whether to apply convolutions on mask features before upsampling using
nearest neighbor resizing.
By default, mask features are resized to [`mask_height`, `mask_width`]
before applying convolutions and predicting masks. Default: false

Backbone configuration 

multi-label classification head 

whether display class-specific evaluation metric Default: false

input layer 

Number of classes 

loss weight Default: 1

global pooling type, max for max_pooling, avg for average pooling Default: max

extra conv layer hidden size 

classification loss 

 Default: 0.9

 Default: 0.999

 Default: 0.002

 Default: 0.002

 Default: 4000000

 Default: 0.0002

 Default: 10000

 Default: 0

 Default: 0.002

 Default: 4000000

 Default: 0.95

 Default: true

 Default: 0

 Default: 0

 Default: 0

 Default: 0.002

Whether to linearly interpolate learning rates for steps in
[0, schedule[0].step]. Default: false

 Default: 0.002

 Default: 0.9

 Default: false

 Default: 0.9999

 Default: 0

 Default: 0.9

 Default: 0.9

 Default: 1

 Default: 1

task type, such as CLASSIFICATION | DETECTION etc. Default: CLASSIFICATION

dataset data directory example
+ data/
   + tfrecord/
       - name_train_1.tfrecord
       - name_train_1_info.json
       - name_test.tfrecord
       - name_label_map.pbtxt
       - name_char_dict
train data path prefix, e.g. data/tfrecord/name 

param space preference Default: ACCURATE

pretrain model directory for incremental training 

param space size Default: 1

train config, including optimizer, weight decay, num_steps and so on 

cv model config 

in local mode user_resource_path should be
set to the directory containing customized code 

auto-compression options. 

Scalar threshold for score used in evaluation (low scoring boxes are removed). Default: 0

workaround setup for score_threshold used for prediction
to avoid revising config file before exporting models Default: 0.5

Scalar threshold for IOU (boxes that have high IOU overlap
with previously selected boxes are removed). Default: 0.6

Maximum number of detections to retain per class. Default: 100

Maximum number of detections to retain across all classes. Default: 100

Class agnostic set in nms. Default: false

Non max suppression parameters. 

Score converter to use. Default: IDENTITY

Scale logit (input) value before conversion in post-processing step.
Typically used for softmax distillation, though can be used to scale for
other reasons. Default: 1

name of predictor or predictor class path 

predictor model path 

evaluator config 

evaluation data config 

video length Default: 768

video input size Default: 112

resize height and width Default: 128

 Default: 171

means 

flip and random crop indicator Default: true

 Default: true

 Default: 32

 Default: 32

 Default: false

 Default: false

central crop fraction Default: 0.875

the number of augmentation transformations to apply
sequentially to an image Default: 2

shared magnitude across all augmentation operations Default: 10

Aspect ratio bounds of cropped image. Default: 0.75

 Default: 1.33

Allowed area ratio of cropped image to original image. Default: 0.1

 Default: 1

 Default: 513

if model_name is set, output_width and output_width will use default for
this model, e.g., efficientnet-b0: 224, efficientnet-b1: 240 

 Default: 224

 Default: 224

 Default: false

the name of the augmentation method to apply to the image.
`autoaugment` if AutoAugment is to be used
`randaugment` if RandAugment is to be used Default: randaugment

the number of augmentation transformations to apply
sequentially to an image in randaugment Default: 2

shared magnitude across all augmentation operations in randaugment Default: 10

 Default: 224

 Default: 224

 Default: false

 Default: false

 Default: 0.875

 Default: 16

 Default: 3

scale parameters Default: 1

 Default: 5

 Default: 0.840896428

spatial crop type Default: corner

sample size Default: 112

 Default: 1

 Default: true

 Default: true

 Default: 28

 Default: 28

 Default: false

target image aspect ratio 

constant value to pad 

 Default: 0

 Default: 1

 Default: 0.2

 Default: 0.8

 Default: 1.25

 Default: 0.02

 Default: 0.8

 Default: 1.25

The maximum number of black patches to add. Default: 10

The probability of a black patch being added to an image. Default: 0.5

Ratio between the dimension of the black patch to the minimum dimension of
the image (patch_width = patch_height = min(image_height, image_width)). Default: 0.1

Cropped image must cover at least one box by this fraction. Default: 1

Aspect ratio bounds of cropped image. Default: 0.75

 Default: 1.33

Allowed area ratio of cropped image to original image. Default: 0.1

 Default: 1

Minimum overlap threshold of cropped boxes to keep in new image. If the
ratio between a cropped bounding box and the original is less than this
value, it is removed from the new image. Default: 0.3

Probability of keeping the original image. Default: 0

Cropping operation must cover at least one box by this fraction. Default: 1

Aspect ratio bounds of image after cropping operation. Default: 0.75

 Default: 1.33

Allowed area ratio of image after cropping operation. Default: 0.1

 Default: 1

Minimum overlap threshold of cropped boxes to keep in new image. If the
ratio between a cropped bounding box and the original is less than this
value, it is removed from the new image. Default: 0.3

Probability of keeping the original image during the crop operation. Default: 0

Maximum dimensions for padded image. If unset, will use double the original
image dimension as a lower bound. Both of the following fields should be
length 2. 

Color of the padding. If unset, will pad using average color of the input
image. This field should be of length 3. 

Cropped image must cover at least one box by this fraction. Default: 1

Aspect ratio bounds of cropped image. Default: 0.2

 Default: 5

Allowed area ratio of cropped image to original image. Default: 0.1

 Default: 1

Probability of keeping the original image. Default: 0.1

Aspect ratio. Default: 1

Minimum overlap threshold of cropped boxes to keep in new image. If the
ratio between a cropped bounding box and the original is less than this
value, it is removed from the new image. Default: 0.3

0 means first adjust brightness then adjust saturation, 1 otherwise Default: 0

in fast_mode, only adjust brightness and saturation
otherwise, adjust brightness, saturation, hue, contrast Default: false

Specifies a mapping from the original keypoint indices to horizontally
flipped indices. This is used in the event that keypoints are specified,
in which case when the image is horizontally flipped the keypoints will
need to be permuted. E.g. for keypoints representing left_eye, right_eye,
nose_tip, mouth, left_ear, right_ear (in that order), one might specify
the keypoint_flip_permutation below:
keypoint_flip_permutation: 1
keypoint_flip_permutation: 0
keypoint_flip_permutation: 2
keypoint_flip_permutation: 3
keypoint_flip_permutation: 5
keypoint_flip_permutation: 4 

 Default: 0.5

 Default: 2

 Default: 0.8

 Default: 1.2

 Default: BILINEAR

 Default: 0.05

Minimum dimensions for padded image. If unset, will use original image
dimension as a lower bound. 

Maximum dimensions for padded image. If unset, will use double the original
image dimension as a lower bound. 

Color of the padding. If unset, will pad using average color of the input
image. 

 Default: 0.9

 Default: 1.1

 Default: 0.1

 Default: BILINEAR

 Default: BILINEAR

 Default: -10

 Default: 10

random rotate image 90 degree or not Default: true

 Default: -10

 Default: 10

use keypoints to compute new bounding box or not Default: false

Specifies a mapping from the original keypoint indices to vertically
flipped indices. This is used in the event that keypoints are specified,
in which case when the image is vertically flipped the keypoints will
need to be permuted. E.g. for keypoints representing left_eye, right_eye,
nose_tip, mouth, left_ear, right_ear (in that order), one might specify
the keypoint_flip_permutation below:
keypoint_flip_permutation: 1
keypoint_flip_permutation: 0
keypoint_flip_permutation: 2
keypoint_flip_permutation: 3
keypoint_flip_permutation: 5
keypoint_flip_permutation: 4 

 Default: BILINEAR