easy_ml 0.1.3 → 0.2.0.pre.rc1

data/app/models/easy_ml/model.rb

@@ -1,68 +1,602 @@
-require_relative "../../../lib/easy_ml/core/model"
+# == Schema Information
+#
+# Table name: easy_ml_models
+#
+#  id              :bigint           not null, primary key
+#  name            :string           not null
+#  model_type      :string
+#  status          :string
+#  dataset_id      :bigint
+#  model_file_id   :bigint
+#  configuration   :json
+#  version         :string           not null
+#  root_dir        :string
+#  file            :json
+#  sha             :string
+#  last_trained_at :datetime
+#  is_training     :boolean
+#  created_at      :datetime         not null
+#  updated_at      :datetime         not null
+#
+require_relative "models/hyperparameters"
+
 module EasyML
   class Model < ActiveRecord::Base
-    if ActiveRecord::Base.connection.data_source_exists?("easy_ml_models")
-      include EasyML::Core::ModelCore
-
-      self.table_name = "easy_ml_models"
-    else
-      # Placeholder if the table doesn't exist (keeps the file quiet)
-      def self.table_ready?
-        false
+    self.table_name = "easy_ml_models"
+    include Historiographer::Silent
+    historiographer_mode :snapshot_only
+
+    include EasyML::Concerns::Configurable
+    include EasyML::Concerns::Versionable
+
+    self.filter_attributes += [:configuration]
+
+    MODEL_OPTIONS = {
+      "xgboost" => "EasyML::Models::XGBoost",
+    }
+    MODEL_TYPES = [
+      {
+        value: "xgboost",
+        label: "XGBoost",
+        description: "Extreme Gradient Boosting, a scalable and accurate implementation of gradient boosting machines",
+      },
+    ].freeze
+    MODEL_NAMES = MODEL_OPTIONS.keys.freeze
+    MODEL_CONSTANTS = MODEL_OPTIONS.values.map(&:constantize)
+
+    add_configuration_attributes :task, :objective, :hyperparameters, :evaluator, :callbacks, :metrics
+    MODEL_CONSTANTS.flat_map(&:configuration_attributes).each do |attribute|
+      add_configuration_attributes attribute
+    end
+
+    belongs_to :dataset
+    belongs_to :model_file, class_name: "EasyML::ModelFile", foreign_key: "model_file_id", optional: true
+
+    has_one :retraining_job, class_name: "EasyML::RetrainingJob"
+    accepts_nested_attributes_for :retraining_job
+    has_many :retraining_runs, class_name: "EasyML::RetrainingRun"
+    has_many :deploys, class_name: "EasyML::Deploy"
+
+    scope :deployed, -> { EasyML::ModelHistory.deployed }
+
+    def latest_deploy
+      deploys.order(id: :desc).limit(1).last
+    end
+
+    after_initialize :bump_version, if: -> { new_record? }
+    after_initialize :set_defaults, if: -> { new_record? }
+    before_save :save_model_file, if: -> { is_fit? && !is_history_class? && model_changed? && !@skip_save_model_file }
+
+    VALID_TASKS = %i[regression classification].freeze
+
+    TASK_TYPES = [
+      {
+        value: "classification",
+        label: "Classification",
+        description: "Predict categorical outcomes or class labels",
+      },
+      {
+        value: "regression",
+        label: "Regression",
+        description: "Predict continuous numerical values",
+      },
+    ].freeze
+
+    validates :name, presence: true
+    validates :name, uniqueness: { case_sensitive: false }
+    validates :task, presence: true
+    validates :task, inclusion: {
+                       in: VALID_TASKS.map { |t| [t, t.to_s] }.flatten,
+                       message: "must be one of: #{VALID_TASKS.join(", ")}",
+                     }
+    validates :model_type, inclusion: { in: MODEL_NAMES }
+    validates :dataset_id, presence: true
+    validate :validate_metrics_allowed
+    before_save :set_root_dir
+
+    delegate :prepare_data, :preprocess, to: :adapter
+
+    STATUSES = %w[development inference retired]
+    STATUSES.each do |status|
+      define_method "#{status}?" do
+        self.status.to_sym == status.to_sym
       end
+    end
 
-      Rails.logger.info("Skipping EasyML::Model definition as the 'easy_ml_models' table doesn't exist.")
+    def training?
+      is_training == true
     end
 
-    scope :live, -> { where(is_live: true) }
-    attribute :root_dir, :string
-    after_initialize :apply_defaults
+    def train(async: true)
+      pending_run # Ensure we update the pending job before enqueuing in background so UI updates properly
+      update(is_training: true)
+      if async
+        EasyML::TrainingJob.perform_later(id)
+      else
+        actually_train
+      end
+    end
 
-    validate :only_one_model_is_live?
-    def only_one_model_is_live?
-      return if @marking_live
+    def get_retraining_job
+      if retraining_job
+        self.evaluator = retraining_job.evaluator
+        evaluator = self.evaluator.symbolize_keys
+      else
+        default_eval = Core::ModelEvaluator.default_evaluator(task)
+        self.evaluator = default_eval
+        evaluator = default_eval
+      end
+
+      retraining_job || create_retraining_job(
+        model: self,
+        active: false,
+        evaluator: evaluator,
+        metric: evaluator[:metric],
+        direction: evaluator[:direction],
+        threshold: evaluator[:threshold],
+        frequency: "month",
+        at: { hour: 0, day_of_month: 1 },
+      )
+    end
 
-      if previous_versions.live.count > 1
-        raise "Multiple previous versions of #{name} are live! This should never happen. Update previous versions to is_live=false before proceeding"
+    def pending_run
+      job = get_retraining_job
+      job.retraining_runs.find_or_create_by(status: "pending", model: self)
+    end
+
+    def actually_train(&progress_block)
+      lock_model do
+        run = pending_run
+        run.wrap_training do
+          best_params = nil
+          if run.should_tune?
+            best_params = hyperparameter_search(&progress_block)
+          end
+          fit(&progress_block)
+          save
+          [self, best_params]
+        end
+        update(is_training: false)
+        run.reload
+      ensure
+        unlock!
+      end
+    end
+
+    def unlock!
+      Support::Lockable.unlock!(lock_key)
+    end
+
+    def lock_model
+      with_lock do |client|
+        yield
+      end
+    end
+
+    def with_lock
+      EasyML::Support::Lockable.with_lock(lock_key, stale_timeout: 60, resources: 1) do |client|
+        yield client
+      end
+    end
+
+    def lock_key
+      "training:#{self.name}:#{self.id}"
+    end
+
+    def hyperparameter_search(&progress_block)
+      tuner = retraining_job.tuner_config.symbolize_keys
+      extra_params = {
+        evaluator: evaluator,
+        model: self,
+        dataset: dataset,
+      }.compact
+      tuner.merge!(extra_params)
+      tuner_instance = EasyML::Core::Tuner.new(tuner)
+      tuner_instance.tune(&progress_block).tap do |best_params|
+        best_params.each do |key, value|
+          self.hyperparameters.send("#{key}=", value)
+        end
       end
+    end
+
+    def deployment_status
+      status
+    end
+
+    def formatted_model_type
+      adapter.class.name.split("::").last
+    end
+
+    def formatted_version
+      return nil unless version
+      Time.strptime(version, "%Y%m%d%H%M%S").strftime("%B %-d, %Y at %-l:%M %p")
+    end
 
-      return unless previous_versions.live.any? && is_live
+    def last_run_at
+      last_run&.created_at
+    end
+
+    def last_run
+      retraining_runs.order(id: :desc).limit(1).last
+    end
+
+    def inference_version
+      latest_deploy&.model_version
+    end
+
+    alias_method :current_version, :inference_version
+    alias_method :latest_version, :inference_version
+    alias_method :deployed, :inference_version
 
-      errors.add(:is_live,
-                 "cannot mark model live when previous version is live. Explicitly use the mark_live method to mark this as the live version")
+    def hyperparameters
+      @hypers ||= adapter.build_hyperparameters(@hyperparameters)
     end
 
-    def mark_live
-      transaction do
-        self.class.where(name: name).where.not(id: id).update_all(is_live: false)
-        self.class.where(id: id).update_all(is_live: true)
+    def callbacks
+      @cbs ||= adapter.build_callbacks(@callbacks)
+    end
+
+    def predict(xs)
+      load_model!
+      adapter.predict(xs)
+    end
+
+    def save_model_file
+      raise "No trained model! Need to train model before saving (call model.fit)" unless is_fit?
+      return unless adapter.loaded?
+
+      model_file = get_model_file
+
+      bump_version(force: true)
+      path = model_file.full_path(version)
+      full_path = adapter.save_model_file(path)
+      model_file.upload(full_path)
+
+      model_file.save
+      self.model_file = model_file
+      cleanup
+    end
+
+    def feature_names
+      adapter.feature_names
+    end
+
+    def cleanup!
+      get_model_file&.cleanup!
+    end
+
+    def cleanup
+      get_model_file&.cleanup(files_to_keep)
+    end
+
+    def loaded?
+      model_file = get_model_file
+      return false if model_file.persisted? && !File.exist?(model_file.full_path.to_s)
+
+      file_exists = true
+      if model_file.present? && model_file.persisted? && model_file.full_path.present?
+        file_exists = File.exist?(model_file.full_path)
       end
+
+      loaded = adapter.loaded?
+      load_model_file unless loaded
+      file_exists && adapter.loaded?
     end
 
-    def previous_versions
-      EasyML::Model.where(name: name).order(id: :desc)
+    def model_changed?
+      return false unless is_fit?
+      return true if inference_version.nil?
+      return true if model_file.present? && !model_file.persisted?
+      return true if model_file.present? && model_file.fit? && inference_version.nil?
+
+      adapter.model_changed?(inference_version.sha)
+    end
+
+    def feature_importances
+      adapter.feature_importances
+    end
+
+    def fit_in_batches?
+      retraining_job.present? && retraining_job.batch_mode == true
+    end
+
+    def fit(tuning: false, x_train: nil, y_train: nil, x_valid: nil, y_valid: nil, &progress_block)
+      return fit_in_batches(**batch_args.merge!(tuning: tuning), &progress_block) if fit_in_batches?
+
+      dataset.refresh
+      adapter.fit(tuning: tuning, x_train: x_train, y_train: y_train, x_valid: x_valid, y_valid: y_valid, &progress_block)
+      @is_fit = true
+    end
+
+    def batch_args
+      defaults = {
+        batch_size: 1024,
+        batch_overlap: 3,
+        batch_key: nil,
+      }
+      overrides = { batch_size: retraining_job&.batch_size, batch_overlap: retraining_job&.batch_overlap, batch_key: retraining_job&.batch_key }.compact
+      defaults.merge!(overrides)
+    end
+
+    def batch_mode
+      retraining_job&.batch_mode || false
+    end
+
+    def prepare_callbacks(tune_started_at)
+      adapter.prepare_callbacks(tune_started_at)
+    end
+
+    def after_tuning
+      adapter.after_tuning
+    end
+
+    def fit_in_batches(tuning: false, batch_size: nil, batch_overlap: nil, batch_key: nil, checkpoint_dir: Rails.root.join("tmp", "xgboost_checkpoints"), &progress_block)
+      adapter.fit_in_batches(tuning: tuning, batch_size: batch_size, batch_overlap: batch_overlap, batch_key: batch_key, checkpoint_dir: checkpoint_dir, &progress_block)
+      @is_fit = true
+    end
+
+    attr_accessor :is_fit
+
+    def is_fit?
+      model_file = get_model_file
+      return true if model_file.present? && model_file.fit?
+
+      adapter.is_fit?
+    end
+
+    def deployable?
+      cannot_deploy_reasons.none?
+    end
+
+    def decode_labels(ys, col: nil)
+      dataset.decode_labels(ys, col: col)
+    end
+
+    def evaluate(y_pred: nil, y_true: nil, x_true: nil, evaluator: nil)
+      evaluator ||= self.evaluator
+      if y_pred.nil?
+        inputs = default_evaluation_inputs
+        y_pred = inputs[:y_pred]
+        y_true = inputs[:y_true]
+        x_true = inputs[:x_true]
+      end
+      EasyML::Core::ModelEvaluator.evaluate(model: self, y_pred: y_pred, y_true: y_true, x_true: x_true, evaluator: evaluator)
+    end
+
+    def evaluator
+      instance_variable_get(:@evaluator) || default_evaluator
+    end
+
+    def default_evaluator
+      return nil unless task.present?
+
+      EasyML::Core::ModelEvaluator.default_evaluator(task)
+    end
+
+    def get_params
+      @hyperparameters.to_h
+    end
+
+    def evals
+      last_run&.metrics || {}
+    end
+
+    def metric_accessor(metric)
+      metrics = last_run.metrics.symbolize_keys
+      metrics.dig(metric.to_sym)
+    end
+
+    EasyML::Core::ModelEvaluator.metrics.each do |metric_name|
+      define_method metric_name do
+        metric_accessor(metric_name)
+      end
+    end
+
+    EasyML::Core::ModelEvaluator.callbacks = lambda do |metric_name|
+      EasyML::Model.define_method metric_name do
+        metric_accessor(metric_name)
+      end
+    end
+
+    def allowed_metrics
+      EasyML::Core::ModelEvaluator.metrics(task).map(&:to_s)
+    end
+
+    def default_metrics
+      return [] unless task.present?
+
+      case task.to_sym
+      when :regression
+        %w[mean_absolute_error mean_squared_error root_mean_squared_error r2_score]
+      when :classification
+        %w[accuracy_score precision_score recall_score f1_score]
+      else
+        []
+      end
+    end
+
+    def self.constants
+      {
+        objectives: objectives_by_model_type,
+        metrics: metrics_by_task,
+        tasks: TASK_TYPES,
+        timezone: EasyML::Configuration.timezone_label,
+        retraining_job_constants: EasyML::RetrainingJob.constants,
+        tuner_job_constants: EasyML::TunerJob.constants,
+      }
+    end
+
+    def self.metrics_by_task
+      EasyML::Core::ModelEvaluator.metrics_by_task
+    end
+
+    def self.objectives_by_model_type
+      MODEL_OPTIONS.inject({}) do |h, (k, v)|
+        h.tap do
+          h[k] = v.constantize.const_get(:OBJECTIVES_FRONTEND)
+        end
+      end.deep_symbolize_keys
+    end
+
+    def attributes
+      super.merge!(
+        hyperparameters: hyperparameters.to_h,
+      )
+    end
+
+    class CannotdeployError < StandardError
+    end
+
+    def deploy(async: true)
+      last_run.deploy(async: async)
+    end
+
+    def actually_deploy
+      raise CannotdeployError, cannot_deploy_reasons.first if cannot_deploy_reasons.any?
+
+      # Prepare the inference model by freezing + saving the model, dataset, and datasource
+      # (This creates ModelHistory, DatasetHistory, etc)
+      save_model_file
+      self.sha = model_file.sha
+      save
+      dataset.upload_remote_files
+      snapshot.tap do
+        # Prepare the model to be retrained (reset values so they don't conflict with our snapshotted version)
+        bump_version(force: true)
+        dataset.bump_versions(version)
+        self.model_file = new_model_file!
+        save
+      end
+    end
+
+    def cannot_deploy_reasons
+      [
+        is_fit? ? nil : "Model has not been trained",
+        dataset.target.present? ? nil : "Dataset has no target",
+        !dataset.datasource.in_memory? ? nil : "Cannot perform inference using an in-memory datasource",
+      ].compact
+    end
+
+    def root_dir=(value)
+      raise "Cannot override value of root_dir!" unless value.to_s == root_dir.to_s
+
+      write_attribute(:root_dir, value)
+    end
+
+    def set_root_dir
+      write_attribute(:root_dir, root_dir)
+    end
+
+    def root_dir
+      EasyML::Engine.root_dir.join("models").join(underscored_name).to_s
+    end
+
+    def load_model(force: false)
+      download_model_file(force: force)
+      load_model_file
+    end
+
+    def metrics=(value)
+      value = [value] unless value.is_a?(Array)
+      value = value.map(&:to_s)
+      value = value.uniq
+      @metrics = value
+    end
+
+    def adapter
+      @adapter ||= begin
+          adapter_class = MODEL_OPTIONS[model_type]
+          raise "Don't know how to use model adapter #{model_type}!" unless adapter_class.present?
+
+          adapter_class.constantize.new(self)
+        end
     end
 
     private
 
+    def default_evaluation_inputs
+      x_true, y_true = dataset.test(split_ys: true)
+      y_pred = predict(x_true)
+      {
+        x_true: x_true,
+        y_true: y_true,
+        y_pred: y_pred,
+      }
+    end
+
+    def underscored_name
+      name.gsub(/\s{2,}/, " ").gsub(/\s/, "_").downcase
+    end
+
+    def get_model_file
+      model_file || new_model_file!
+    end
+
+    def new_model_file!
+      build_model_file(
+        root_dir: root_dir,
+        model: self,
+        s3_bucket: EasyML::Configuration.s3_bucket,
+        s3_region: EasyML::Configuration.s3_region,
+        s3_access_key_id: EasyML::Configuration.s3_access_key_id,
+        s3_secret_access_key: EasyML::Configuration.s3_secret_access_key,
+        s3_prefix: prefix,
+      )
+    end
+
+    def prefix
+      s3_prefix = EasyML::Configuration.s3_prefix
+      s3_prefix.present? ? File.join(s3_prefix, name) : name
+    end
+
+    def load_model!
+      load_model(force: true)
+      load_dataset
+    end
+
+    def load_dataset
+      dataset.load_dataset
+    end
+
+    def load_model_file
+      return unless model_file&.full_path && File.exist?(model_file.full_path)
+
+      adapter.load_model_file(model_file.full_path)
+    end
+
+    def download_model_file(force: false)
+      return unless persisted?
+      return if loaded? && !force
+
+      get_model_file.download
+    end
+
     def files_to_keep
-      live_models = self.class.live
+      inference_models = EasyML::ModelHistory.deployed
+      training_models = EasyML::Model.all
 
-      recent_copies = live_models.flat_map do |live|
-        # Fetch all models with the same name
-        self.class.where(name: live.name).where(is_live: false).order(created_at: :desc).limit(live.name == name ? 4 : 5)
-      end
+      ([self] + training_models + inference_models).compact.map(&:model_file).compact.map(&:full_path).uniq
+    end
 
-      recent_versions = self.class
-                            .where.not(
-                              "EXISTS (SELECT 1 FROM easy_ml_models e2 WHERE e2.name = easy_ml_models.name AND e2.is_live = true)"
-                            )
-                            .where("created_at >= ?", 2.days.ago)
-                            .order(created_at: :desc)
-                            .group_by(&:name)
-                            .flat_map { |_, models| models.take(5) }
+    def underscored_name
+      name = self.name || self.class.name.split("::").last
+      name.gsub(/\s{2,}/, " ").gsub(/\s/, "_").downcase
+    end
 
-      ([self] + recent_versions + recent_copies + live_models).compact.map(&:file).map(&:path).uniq
+    def set_defaults
+      self.model_type ||= "xgboost"
+      self.status ||= :training
+      self.metrics ||= default_metrics
+    end
+
+    def validate_metrics_allowed
+      unknown_metrics = metrics.select { |metric| allowed_metrics.exclude?(metric) }
+      return unless unknown_metrics.any?
+
+      errors.add(:metrics,
+                 "don't know how to handle #{"metrics".pluralize(unknown_metrics)} #{unknown_metrics.join(", ")}, use EasyML::Core::ModelEvaluator.register(:name, Evaluator, :regression|:classification)")
     end
   end
 end
+
+require_relative "models/xgboost"