package config

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: DatasetConfig

• required string file_pattern = 1

• required string layer_name = 2

• optional string dataset_name = 3

  Name of the dataset within in a data collection. For example, within an hdf5 file.

• optional DataStreamConfig.DataType data_type = 4

• optional int32 raw_image_size = 5

  DEPRECATED: use raw_image_size_y, raw_image_size_x

• optional int32 image_size = 6

  DEPRECATED: use image_size_y, image_size_x

• optional bool can_translate = 7

• optional bool can_flip = 8

• optional bool pixelwise_normalize = 9

• optional float pca_noise_stddev = 10

• optional bool normalize = 11

  Normalize the data by subtracting mean and dividing by stddev.

• optional int32 gpu_id = 12

  The GPU on which this data needs to be loaded.

• optional int32 stride = 13

  For sliding windows.

• optional string mean_file = 14

  Mean, stddev, eig values etc.

• optional int32 num_colors = 15

• optional bool parallel_disk_access = 16

  Enable parallel disk access. Use multiple threads to access the disk. Makes sense for reading raw image files.

• optional bool jitter_raw_image = 17

• optional float random_rotate_max_angle = 18

• optional float min_scale = 19

• optional string noise_layer_name = 20

  This layer needs the noise added to 'noise_layer_name'. Used for communicating jitter noise to bounding box outputs.

• optional bool avg10_full_image = 21

  Applicable to IMAGE_RAW.

• optional string bbox_file = 22

  Applicable to IMAGE_RAW.

• optional float context_factor = 23

  Applicable to IMAGE_RAW.

• optional bool center_on_bbox = 24

  Applicable to IMAGE_RAW.

• optional bool warp_bbox = 25

• optional int32 gpu_image_size_y = 26

• optional int32 gpu_image_size_x = 27

• optional int32 raw_image_size_y = 28

• optional int32 raw_image_size_x = 29

• optional int32 image_size_y = 30

• optional int32 image_size_x = 31

• optional bool is_sequence = 32

• optional int32 seq_length = 33

• optional string boundary_file = 34

• optional bool pick_first = 35

• optional bool normalize_local = 36

Used in: DataStreamConfig

• DUMMY = 0

• HDF5 = 1

• IMAGE_RAW = 2

• SLIDING_WINDOW = 3

• TXT = 4

• BOUNDING_BOX = 5

• CROPS = 6

• VIDEO_RAW = 7

Used in: FeatureExtractorConfig, Model

• repeated DataStreamConfig data_config = 1

• optional int32 batch_size = 2

• optional int32 chunk_size = 3

• optional int32 max_reuse_count = 4

• optional bool pipeline_loads = 5

• optional bool randomize_cpu = 6

• optional bool randomize_gpu = 7

• optional int32 random_access_chunk_size = 8

• optional int32 max_dataset_size = 9

• optional int32 multiplicity = 10

Used in: Model

• required string source = 1

  Names of the source and destination layers.

• required string dest = 2

• optional Edge.EdgeType edge_type = 3

• optional int32 kernel_size = 4

  Hyperparameters for convolution-like edges.

• optional int32 stride = 5

• optional int32 padding = 6

• optional Edge.Initialization initialization = 7

• optional float init_wt = 8

  Initial weight scale. This will be divided by sqrt fan-in, depending on initialization.

• optional float init_bias = 9

  Initial biases (constant).

• optional Optimizer weight_optimizer = 10

• optional Optimizer bias_optimizer = 11

• optional bool shared_bias = 12

  Whether biases whould be tied for convolutional layers.

• optional bool block_backprop = 13

  Block backprop through this edge.

• optional string tied_to = 14

  name of the edge to which this edge is tied. Name is specified as "source:dest".

• optional bool has_no_bias = 15

  Make the edge have no bias - This is useful when multiple edges feed into one layer and only one of them needs to have a bias.

• optional float scale_gradients = 16

  Scale the gradients from this edge by this amount before computing updates. Useful when the edge is tied and we want the contributions from different sources to be scaled.

• optional int32 partial_sum = 17

  Used for convolutional layers. Makes outer product faster by doing partial sums and then adding up. Default = 0 means add up over all locations in the kernel (slowest, uses less memory). Must exactly divide number of convolution locations.

• optional int32 sample_factor = 18

  Factor to up/down sample.

• optional bool response_norm_in_blocks = 19

  Hyperparameters for RESPONSE_NORM layer.

• optional float add_scale = 20

• optional float pow_scale = 21

• optional float frac_of_filters_response_norm = 22

• optional int32 gpu_id = 24

• optional string pretrained_model = 25

• optional string pretrained_edge_name = 26

• optional bool display = 27

• optional string source_slice = 28

  Layer slices.

• optional string dest_slice = 29

• optional bool grad_check = 30

  Grad check.

• optional int32 grad_check_num_params = 31

• repeated float grad_check_epsilon = 32

• optional int32 kernel_size_y = 33

  More general Conv params.

• optional int32 kernel_size_x = 34

• optional int32 kernel_size_t = 35

• optional int32 stride_y = 36

• optional int32 stride_x = 37

• optional int32 stride_t = 38

• optional int32 padding_y = 39

• optional int32 padding_x = 40

• optional int32 padding_t = 41

Used in: Edge

• FC = 0

  Fully Connected.

• CONVOLUTIONAL = 1

  Convolutional.

• LOCAL = 2

  Locally connected.

• MAXPOOL = 3

  Maxpool.

• RESPONSE_NORM = 4

  Cross-Map response normalization.

• UPSAMPLE = 5

• DOWNSAMPLE = 6

• RGBTOYUV = 7

• AVERAGE_POOL = 8

• CONV_ONETOONE = 9

  BATCH_NORMALIZATION = 10; REMOVED : Use inplace batch normalization in layer.

Weight initialization.

Used in: Edge

• DENSE_GAUSSIAN = 0

• SPARSE_GAUSSIAN = 1

• CONSTANT = 2

• DENSE_GAUSSIAN_SQRT_FAN_IN = 3

• PRETRAINED = 4

• DENSE_UNIFORM = 5

• DENSE_UNIFORM_SQRT_FAN_IN = 6

• required string output_file = 1

• repeated FeatureStreamConfig feature = 2

• required DatasetConfig input = 3

Used in: FeatureExtractorConfig

• required string layer = 1

• optional int32 average_batches = 2

• optional int32 average_online = 3

Used in: Model

• required string name = 1

• optional int32 num_channels = 2

• optional int32 size = 3

• optional float dropprob = 4

• optional bool is_input = 5

  DEPRECATED

• optional Layer.Activation activation = 6

• optional int32 image_size_y = 7

• optional int32 image_size_x = 8

• optional bool display = 9

• optional bool is_output = 10

  DEPRECATED

• optional bool gaussian_dropout = 11

• optional float max_act_gaussian_dropout = 12

• optional int32 gpu_id = 13

• optional float hinge_margin = 14

• repeated LayerSlice layer_slice = 15

  Sublayers.

• optional Layer.LossFunction loss_function = 16

• optional Layer.LossFunction performance_metric = 17

• optional float loss_function_weight = 18

• optional string tied_data = 19

  The data (input or target) at this layer should be tied to the data at 'tied_data'.

• optional int32 image_size_t = 20

• optional bool batch_normalize = 21

  Batch Normalization.

• optional float bn_f = 22

  Running avg.

• optional float bn_epsilon = 23

  For stabilizing sigma.

• optional Optimizer gamma_optimizer = 24

• optional Optimizer beta_optimizer = 25

Used in: Layer

• LINEAR = 0

• LOGISTIC = 1

• RECTIFIED_LINEAR = 2

• SOFTMAX = 3

• SOFTMAX_DIST = 4

  not one-of-K, but a distribution over K.

Used in: Layer

• SQUARED_ERROR = 0

• LINEAR_ERROR = 1

• CROSS_ENTROPY_MULTINOMIAL = 2

• CROSS_ENTROPY_BINARY = 3

• CROSS_ENTROPY_MULTINOMIAL_DISTRIBUTED = 4

• CLASSIFICATION_MULTINOMIAL = 5

• CLASSIFICATION_BINARY = 6

• HINGE_LINEAR = 7

• HINGE_QUADRATIC = 8

Used in: Layer

• required string name = 1

• optional int32 num_channels = 2

• required string name = 1

• repeated Layer layer = 2

• repeated Edge edge = 3

• required int32 seed = 4

• optional int32 max_iter = 5

• optional int32 display_after = 6

• optional int32 save_after = 7

• optional int32 image_size = 8

• optional int32 patch_size = 9

• optional int32 print_after = 10

• optional bool localizer = 11

• optional string checkpoint_dir = 12

• optional bool print_weights = 13

• repeated string timestamp = 15

• optional bool display = 17

• optional int32 validate_after = 18

• optional float reduce_lr_factor = 19

  For reducing learning rates.

• optional float reduce_lr_threshold = 20

• optional int32 reduce_lr_num_steps = 21

• optional int32 reduce_lr_max = 22

• optional bool smaller_is_better = 23

• optional int32 polyak_after = 24

  Polyak averaging.

• optional int32 polyak_queue_size = 25

• repeated Subnet subnet = 26

  Include other networks.

• optional DatasetConfig train_dataset = 27

  The dataset configs used to train this model.

• optional DatasetConfig valid_dataset = 28

• optional Optimizer default_weight_optimizer = 29

  The default optimizer to be used when no optimizer is specified for a parameter.

• optional Optimizer default_bias_optimizer = 30

• optional string reduce_lr_layer_name = 31

  Use this layer's performance metric for determining when to reduce learning rates. Useful when there are many output layers.

Used in: Edge, Layer, Model

• optional Optimizer.OptimizerType optimizer_type = 1

• optional float epsilon = 2

• optional int32 epsilon_decay_timescale = 3

• optional float initial_momentum = 4

• optional float final_momentum = 5

• optional int32 momentum_transition_timescale = 6

• optional float l2_decay = 7

• optional float weight_norm_limit = 8

• optional float weight_norm_constraint = 9

• optional Optimizer.Decay epsilon_decay = 10

• optional float minimum_epsilon = 11

• optional float decay_factor = 12

  Relevant only for EXPONENTIAL_STEP decay.

• optional float gradient_clip = 13

  Clip gradients. Default : -1 (Don't clip).

• optional int32 lbfgs_memory = 14

  Number of steps to remember.

• optional int32 start_optimization_after = 15

• optional float adagrad_delta = 16

  Adagrad delta.

• optional float rms_prop_factor = 17

  Running avg factor for rms_prop. Default (0) same as rprop.

• optional bool nesterov_momentum = 18

  Use Nesterov momentum.

• optional bool shared_prior = 19

• optional float shared_prior_cost = 20

• optional string shared_prior_file = 21

Used in: Optimizer

• NONE = 0

• INVERSE_T = 1

• EXPONENTIAL = 2

• LINEAR = 3

• EXPONENTIAL_STEP = 4

Used in: Optimizer

• STOCHASTIC_GRADIENT_DESCENT = 0

• LBFGS = 1

• ADAGRAD_SGD = 2

• RMSPROP_SGD = 3

Used in: Model

• required string name = 1

• required string model_file = 2

• optional string parameters_file = 3

• repeated Subnet.MergeLayer merge_layer = 4

• optional bool block_backprop = 5

• optional int32 start_optimization_after = 6

• optional int32 gpu_id_offset = 7

• optional int32 num_channels_multiplier = 8

• repeated string remove_layer = 9

Used in: Subnet

• required string subnet_layer = 1

• required string net_layer = 2