sample_options.org

Sample Options for Rivet

{
  <<options>>
}

Training

"cuda"                : true,
"num_epochs"          : 200,
"report_frequency"    : 32,
"save_frequency"      : 1,

save_frequency represents every n^th epoch, when the model is checked for being saved.

Network

Base Network

"network"             : "resnet18",
"weights_file"        : "/home/bvr/data/pytocone/save-20180427-114144/model_best.pth.tar",
"weights_key"         : "state_dict",

Combining Model (Pairwise/ Triplet)

"model"               : "feature_pair", 
"model_params"        : {
  "fc"                : null, 
  "feat_len"          : 512
},

model may be one of:

"all_models"          : [ "feature_pair",
                          "feature_triple",
                          "concat_pair",
                          "concat_triple"
                        ],

fc stands for fully connected layers. The value may either be null which is interpreted as Identity, or a list of output sizes for a series of fully connected layers in an artificial neural network.

For example,

"fc" : [128, 2]

feat_len is the output size of the base network (expected to be flattened).

Data

"dataset"             : "pairwise",
"dataset_params"      : {
  "adjacency"         : "/path/to/adjacency.json",
  "image_list"        : "/path/to/image_list.json",
  "labels"            : [0, 1],
  "transform"         : "sketch_transform"
},

"dataloader_params"   : {
  "batch_size"        : 256,
  "shuffle"           : true,
  "num_workers"       : 7
},

dataset may be one of:

"all_dataset"         : [ "pairwise",
                          "triplet"],

Loss

"criterion"           : "contrastive",
"criterion_params"    : {
  "distance"          : "euclidean" ,
  "margin"            : 2.0
},

criterion may be one of:

"all_criteria"        : [ "contrastive",
                          "triplet",
                          "bce",
                          "bce_triplet"
                        ],

distance may be one of:

"all_distances"       : [ "euclidean",
                          "kldiv"
                        ],

where kldiv stands for KLDivergence

Optimizer

"optimizer"           : "Adam",
"optimizer_params"    : {
  "weight_decay"      : 0.1
},

optimizer may be one of:

"all_optimizers"      : [ "Adadelta",
                          "Adagrad",
                          "Adam",
                          "SparseAdam",
                          "Adamax",
                          "ASGD",
                          "LBFGS",
                          "RMSprop",
                          "Rprop",
                          "SGD"
                        ],

Learning Rate Adjusters

"lr_adjuster"         : "ReduceLROnPlateau",
"lr_adjuster_params"  : {
  "eps"                 : 1e-4
},

lr_adjuster may be one of :

"all_lr_adjusters"    : [ "LambdaLR",
                          "StepLR",
                          "MultiStepLR",
                          "ExponentialLR",
                          "CosineAnnealingLR",
                          "ReduceLROnPlateau"
                        ],

Accuracy

"accuracy_transform"        : "loss_interpreter",
"accuracy_transform_params" : {},

accuracy_transform may be one of:

"all_accuracy_transform" : [ "loss_interpreter"
                           ]

Saver

"save_location"       : ".",
"saver_current"       : "checkpoint.pth.tar",
"saver_best"          : "model_best.pth.tar",

Reporting

"reporters"           : [ "log_average"
                        ]

reporters may once include either of:

"all_reporters"       : [ "log_average",
                          "grapher"
                        ],