# MODEL MANAGER FLOW

**Step 1:** Training Event Topic:

Training Event object:

ModelTrainingEvent {
ModelDesignSchema design_schema = 1
string dataset_artifact_url = 2
}

**Step 2:**
Fetch ModelDesignSchema from Redis by passing 'modelKeyToken'.
Deserialize and unpack the ModelDesignSchema.

Redis Config: IP - 10.168.134.54:6379
No auth

**Step 3:**
Refer dataset from shared K8 PVC path 'datasetPath'.
'datasetPath' will be already mounted.

**Step 4:**
Initiate Training Process

- Create Docker Image of model training
- Create Kubernetes Job

**Step 5:**
Save training result:

1. Send status of training in kafka queue
Queue Details:
    Topic -
    Message - 
    {
      string design_schema_id
      ModelTrainingStatusEnum status; //Success/Failure
      repeated String errors;
    }

2. Send Training output in Model Store
Model Store Details:
    Artifact: retail_forecasting_model.pkl
    Output:

3. Send Training output to Inference Store

BLOB Storage(Minio) is used to store training result and its path will be pushed to inference queue.

Inference Store Queue Details:
    Topic -
    Message - 
    {
        string model_id;
        string dataset_artifact_url;
    }

# **Model Design Schema Proto:**

**ModelDesignSchema**

message ModelDesignSchema {
  core.base.Token design_token = 1;
  core.base.Token producer_token = 2;
  algorithms.ModelAlgorithms algorithm = 3;
  ModelDesignVersion version = 4;
  ModelDesignMetadata metadata = 5;
  ModelDesignInputSchema input = 6;
  ModelDesignOutputSchema output = 7;
  algorithms.ModelHyperParameter hyper_parameter = 8;
  ModelDesignPreprocessorSchema preprocessor = 9;
  ModelDesignPostprocessorSchema postprocessor = 10;
  ModelDesignCriteria criteria = 11;
  instance.ModelAssembly model_framework = 12;
  core.base.Token dataset_token = 13;
  algorithms.ModelLossCriteria loss_criteria = 14;
  algorithms.ModelOptimizer learning_function = 15;
  algorithms.ModelEvaluationCriteria eval_criteria = 16;
  ModelTrainingOutputStore output_store = 17;
}

**ModelAlgorithms**

message ModelAlgorithms {
  core.base.Token algorithm_token = 1;
  oneof cortex_model_algorithm_is {
    ClassificationModels classification = 10;
    // RegressionModels regression = 20;
    // ClusteringModels clustering = 30;
    // DensityModels density = 40;
  }
}

message ClassificationModels {
  oneof classification_model_type_is {
    BinaryClassificationModels binary = 1;
    // MulticlassClassificationModels multiclass = 2;
    // MultilabelClassificationModels multilabel = 3;
    // ExtremeClassificationModels extreme = 4;
  }
}

message BinaryClassificationModels {
  oneof binary_classification_model_sub_type_is {
    LogisticRegression logistic_regression           = 1;
    //ArtificialNeuralNetworkClassifier neural_network = 2;
    //SupportVectorMachineClassifier svm = 2;
    //LinearDiscriminantAnalysis lda = 3;
    //QuadraticDiscriminantAnalysis qda = 4;
    //DecisionTreeClassifier decision_tree = 6;
    //RandomForestClassifier random_forest = 7;
    //GbdtClassifier gbdt = 8;
    //XgboostClassifier xgboost = 9;
  }
}

message LogisticRegression {
    ModelHyperParameter hyper_parameters = 1;
    model.instance.LogisticRegressionParameter parameters = 2;
    // repeated string attributes                   = 3;
}

message LogisticRegressionParameter {
    repeated string classes_          = 1;
    repeated double coef_             = 2;      // feature weights e.g., m1, m2, m3, m4, etc.
    repeated double intercept_        = 3;      // y intercept (m0)
    int64 n_features_in_              = 4;
    // repeated int64 feature_names_in_  = 5;
    repeated int32 n_iter_            = 5;
}

**ModelDesignInputSchema**

message ModelDesignInputSchema {
  common.ModelInputType input_type = 1;
  // instance.ModelFramework model_framework = 3;
  ModelDatasetFramework dataset_framework = 2;
  oneof input_schema_is_one_of {
    ModelDesignMultivariateInputSchema multivariate = 10;
    ModelDesignStreamInputSchema time_series = 11;
    ModelDesignImageInputSchema image = 12;
  }
}

enum ModelInputType {
  MODEL_INPUT_UNSPECIFIED = 0;
  MODEL_INPUT_MULTIVARIATE = 2100;
  MODEL_INPUT_BASKET = 2200;
  MODEL_INPUT_BAG = 2300;
  MODEL_INPUT_SERIES = 2400;
  MODEL_INPUT_SEQUENCE = 2500;
  MODEL_INPUT_TEXT = 2600;
  MODEL_INPUT_IMAGE = 2700;
  MODEL_INPUT_VIDEO = 2800;
  MODEL_INPUT_AUDIO = 2900;
}

message ModelDatasetFramework {
  ModelDatasetFrameworkType dataset_framework_type = 1; // "TORCHVISION"
  string framework_dataset_name =2; // "CIFAR-10".
  string dataset_version =3;
}

enum ModelDatasetFrameworkType {
  DATASET_FRAMEWORK_GENERIC_DEFAULT = 0;
  DATASET_FRAMEWORK_TORCHVISION=1;
  DATASET_FRAMEWORK_KERAS=2;
  DATASET_FRAMEWORK_HUGGINGFACE=3;
}

message ModelDesignMultivariateInputSchema {
  // /healthcare/hrp/feature/bp or attribute token or model output/inference
  // /telecom/inference/customer/churn
  // /telecom/feature/customer/engagement
  // /telecom/knowledge/customer/gender
//  repeated telos.core.base.Token variable = 1;
 //  same as instance input schema
  core.type.StrVector design_input = 2; // Goes to Model (Output of Preprocessor)
}

message StrVector{
  repeated string element = 1;
}

**ModelDesignOutputSchema**

message ModelDesignOutputSchema {
  common.ModelOutputType design_output_type = 1;
  oneof output_is_one_of {
    ModelDesignClassificationOutputSchema classification_schema = 10;
    ModelDesignRegressionOutputSchema regression_schema = 11;
    ModelDesignProjectionOutputSchema projection_schema = 12;
    ModelDesignClusteringOutputSchema clustering_schema = 13;
    ModelDesignDensityOutputSchema density_schema = 14;
    ModelDesignRecommendationOutputSchema recommendation_schema = 15;
    ModelDesignForecastingOutputSchema forecasting_schema = 16;
    ModelDesignAnomalyOutputSchema anomaly_schema = 17;
    ModelDesignOptimizationOutputSchema optimization_schema = 18;
    ModelDesignDetectionOutputSchema detection_schema = 19;
    ModelDesignAnnotationOutputSchema annotation_schema = 20;
  }
}

enum ModelOutputType {
  OUTPUT_TYPE_UNSPECIFIED = 0;
  OUTPUT_TYPE_CLASSIFICATION = 2;
  OUTPUT_TYPE_REGRESSION = 3;
  OUTPUT_TYPE_PROJECTION = 4;
  OUTPUT_TYPE_CLUSTERING = 5;
  OUTPUT_TYPE_DENSITY = 6;
  OUTPUT_TYPE_RECOMMENDATION = 7;
  OUTPUT_TYPE_FORECASTING = 8;
  OUTPUT_TYPE_ANOMALY = 9;
  OUTPUT_TYPE_OPTIMIZATION = 10;
  OUTPUT_TYPE_DETECTION = 11;
  OUTPUT_TYPE_ANNOTATION = 12;
}

message ModelDesignClassificationOutputSchema {
  oneof classification_output_schema_is_one_of {
    ModelDesignBinaryClassificationOutputSchema binary = 1;
    ModelDesignMultiClassClassificationOutputSchema multi_class = 10;
    //ModelMultiLabelClassificationOutputSchema multi_label = 11;
    //ModelExtremeClassificationOutputSchema extreme = 12;
  }
}

message ModelDesignBinaryClassificationOutputSchema {
  string positive = 1;
  string negative = 2;
  telos.core.base.Token target_variable = 3;
}

message ModelDesignMultiClassClassificationOutputSchema {
  telos.core.type.Collection class_dictionary = 1;
}

message ModelDesignRegressionOutputSchema {
  //core.quantity.QuantitySchema target_variable = 1; // atomic numeric
  string label_schema = 1;
  string inference_schema = 2;
}

**ModelHyperParameter**

message ModelHyperParameter {
  core.base.Token design_token = 1;     // ModelDesignToken
  core.base.Token algorithm_token = 2;
  map<string, HyperParamater> hyper_parameter = 3; // TODO: can be changed to Quantity?
}

message HyperParamater {
  oneof hyperparameter_is_oneof {
    string str = 1;
    bool  bl = 2;
    float flt = 3;
    int32 i32 = 4;
    core.type.F32Vector fvector = 5;
    core.type.StrVector svector = 6;
  }
}

**ModelDesignPreprocessorSchema**

message ModelDesignPreprocessorSchema {
  oneof preprocessor_is_one_of {
    ModelDesignMultivariatePreprocessorSchema multivariate = 1;
   // CortexImagePreprocessor image = 2;
   // CortexVideoPreprocessor video = 3;
   // CortexSeriesPreprocessor series = 4;
   // CortexSequencePreprocessor sequence = 5;
   // CortexBagPreprocessor bag = 6;
  }
}

message ModelDesignMultivariatePreprocessorSchema {
  //telos.core.expression.ExpressionProgram program = 1;
  repeated PreprocessorExpression preprocess = 1;
}

message PreprocessorExpression {
  string expression = 1;
  string output=2;
  map<string, core.type.Primitive> constants = 3; // tbd: right place or not ??
}

message Primitive {
  oneof primitive_is_one_of {
    float float_data = 1;
    double double_data = 2;
    uint32 uint32_data = 3;
    sint32 sint32_data = 4;
    uint64 uint64_data = 5;
    sint64 sint64_data = 6;
    string string_data = 7;
    bool boolean_data = 8;
    Vector list_data = 9;
    //NumberVector number_list_data = 10;
    //StrStrMap mapdata = 10;
  }
}

**ModelDesignPostprocessorSchema**

message ModelDesignPostprocessorSchema {
  oneof postprocessor_schema_is {
    ModelDesignClassificationProbabilityPostprocessorSchema probability_postprocessor_schema = 1;
  }
}

message ModelDesignClassificationProbabilityPostprocessorSchema {
  telos.core.expression.ExpressionProgram program = 1;
}

message ExpressionProgram {
  ExpressionMap program = 1;
  ExpressionProgramInitialization initialization = 2;
  ExpressionProgramDependency dependency = 3; // gets computed at the time of validation/initialization
  repeated string topological_order = 4;
}

**ModelAssembly**

message ModelAssembly {
  ModelFramework model_framework = 1;  //variant name(MODEL_FRAMEWORK_SKLEARN) & variantVersion(0.24.2)
  ModelDependencies dependencies = 2;
  ModelCompute compute_dependencies = 3;
  ModelDeploymentRPC deployment_rpc = 4; // can be removed - this was added for v1
}

message ModelFramework {
  ModelFrameworkType framework_type = 1; // "scikit-learn"
  string framework_version = 2; //0.24.2
  string framework_algorithm_name =3; // algorithm name within framework.
}


enum ModelFrameworkType {
  MODEL_FRAMEWORK_SKLEARN_DEFAULT = 0;
  MODEL_FRAMEWORK_SKLEARN = 1;
  MODEL_FRAMEWORK_TENSORFLOW = 2;
  MODEL_FRAMEWORK_PYTORCH = 3;
  MODEL_FRAMEWORK_MXNET = 4;
  MODEL_FRAMEWORK_CAFFE = 5;
  MODEL_FRAMEWORK_TRITON_SERVER = 6;
  MODEL_FRAMEWORK_LIGHTGBM = 7;
  MODEL_FRAMEWORK_CUSTOM = 8;
  MODEL_FRAMEWORK_XGBOOST = 9;
  MODEL_FRAMEWORK_CORNAC = 10;
}


message ModelDependencies {
  string name = 1;
  string version = 2;
}


message ModelCompute {
  oneof compute_type {
    ModelCPUCompute cpu = 1;
    ModelGPUCompute gpu = 2;
  }
}


message ModelCPUCompute {
  CpuType cpu_type = 1;
  int32 no_of_cores = 2;
  float ram = 3;
}


enum GpuType {
  GPU_TYPE_NVIDIA_DEFAULT = 0;
  GPU_TYPE_NVIDIA = 1;
  GPU_TYPE_AMD = 2;
}


enum CpuType {
  CPU_TYPE_INTEL_DEFAULT = 0;
  CPU_TYPE_INTEL = 1;
  CPU_TYPE_AMD = 2;
  CPU_TYPE_M1 = 3;
  CPU_TYPE_M2 = 4;
}


message ModelGPUCompute {
  GpuType gpu_type = 1;
}

//TODO : This change is just for a v1 for cortex e2e. Going forward use ModelEndpoint
message ModelDeploymentRPC{
  string url = 1;
}

**ModelTrainingOutputStore**

message ModelTrainingOutputStore {
  oneof options {
    bool model_store = 1;
    bool inference_store = 2;
  }
}