feat: Output components with combined value

Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>
georgeguimaraes · Aug 29, 2024 · a8e04b2 · a8e04b2
1 parent 40407f6
commit a8e04b2
Show file tree

Hide file tree

Showing 3 changed files with 20 additions and 7 deletions.
diff --git a/lib/soothsayer.ex b/lib/soothsayer.ex
@@ -44,6 +44,11 @@ defmodule Soothsayer do
   end
 
   def predict(%Model{} = model, %Series{} = x) do
+    %{combined: combined} = predict_components(model, x)
+    combined
+  end
+
+  def predict_components(%Model{} = model, %Series{} = x) do
     processed_x =
       Preprocessor.prepare_data(DataFrame.new(%{"ds" => x}), nil, "ds", model.config.seasonality)
 
@@ -57,9 +62,10 @@ defmodule Soothsayer do
     x_normalized = normalize_with_params(x_input, model.config.normalization.x)
 
     predictions = Model.predict(model, x_normalized)
-    denormalized_predictions = denormalize(predictions, model.config.normalization.y)
 
-    Nx.to_flat_list(denormalized_predictions)
+    Map.new(predictions, fn {key, node} ->
+      {key, denormalize(node, model.config.normalization.y)}
+    end)
   end
 
   defp get_seasonality_input(data, seasonality, config) do

diff --git a/lib/soothsayer/model.ex b/lib/soothsayer/model.ex
@@ -34,7 +34,14 @@ defmodule Soothsayer.Model do
         Axon.constant(0)
       end
 
-    Axon.add([trend, yearly_seasonality, weekly_seasonality])
+    combined = Axon.add([trend, yearly_seasonality, weekly_seasonality])
+
+    Axon.container(%{
+      combined: combined,
+      trend: trend,
+      yearly_seasonality: yearly_seasonality,
+      weekly_seasonality: weekly_seasonality
+    })
   end
 
   def fit(model, x, y, epochs) do
@@ -44,7 +51,7 @@ defmodule Soothsayer.Model do
     trained_params =
       model.network
       |> Axon.Loop.trainer(
-        :mean_squared_error,
+        &Axon.Losses.huber(&1, &2.combined, reduction: :mean),
         Polaris.Optimizers.adam(learning_rate: model.config.learning_rate)
       )
       |> Axon.Loop.run(Stream.repeatedly(fn -> {x, y} end), initial_params,

diff --git a/test/soothsayer_test.exs b/test/soothsayer_test.exs
@@ -41,7 +41,7 @@ defmodule SoothsayerTest do
       future_end = Date.add(future_start, 29)
       future_dates = Date.range(future_start, future_end)
       x_test = Series.from_list(Enum.to_list(future_dates))
-      predictions = Soothsayer.predict(fitted_model, x_test)
+      predictions = Soothsayer.predict(fitted_model, x_test) |> Nx.to_flat_list()
 
       # Check if predictions follow the trend (with tolerance for noise)
       Enum.zip(predictions, future_dates)
@@ -89,7 +89,7 @@ defmodule SoothsayerTest do
       future_end = Date.add(future_start, 29)
       future_dates = Date.range(future_start, future_end)
       x_test = Series.from_list(Enum.to_list(future_dates))
-      predictions = Soothsayer.predict(fitted_model, x_test)
+      predictions = Soothsayer.predict(fitted_model, x_test) |> Nx.to_flat_list()
 
       # Check if predictions follow the seasonality pattern (with tolerance for noise)
       Enum.zip(predictions, future_dates)
@@ -140,7 +140,7 @@ defmodule SoothsayerTest do
       future_end = Date.add(future_start, 29)
       future_dates = Date.range(future_start, future_end)
       x_test = Series.from_list(Enum.to_list(future_dates))
-      predictions = Soothsayer.predict(fitted_model, x_test)
+      predictions = Soothsayer.predict(fitted_model, x_test) |> Nx.to_flat_list()
 
       # Check if predictions follow the trend and seasonality (with tolerance for noise)
       Enum.zip(predictions, future_dates)