Reverted to use vectorized_update_gradients() instead of self.ftrl.pa…

…r_update_gradients() as len(user_items) is 20 or less and no enough room to utilize parallel processing
myui · Nov 17, 2024 · cb81476 · cb81476
1 parent bee9fe3
commit cb81476
Showing 1 changed file with 16 additions and 16 deletions.
diff --git a/src/slim.rs b/src/slim.rs
@@ -69,6 +69,20 @@ impl SlimMSE {
         }
     }
 
+    pub fn fit_identified(&mut self, user_interactions: Vec<(i32, i32, f32, f32)>, add_interaction: Option<bool>, update_interaction: Option<bool>) {
+        for (user_id, item_id, tstamp, rating) in user_interactions {
+            if let Err(e) = {
+                if add_interaction.unwrap_or(true) {
+                    self.interactions.add_interaction(user_id, item_id, tstamp, rating, update_interaction.unwrap_or(false));
+                }
+                self.update_weights(user_id, item_id);
+                Ok::<(), Box<dyn std::error::Error>>(())
+            } {
+                warn!("Failed to fit interaction: {}", e);
+            }
+        }
+    }
+
     #[pyo3(signature = (pydf, epochs = 1, random_seed = None))]
     pub fn bulk_fit(&mut self, pydf: PyDataFrame, epochs: usize, random_seed: Option<u64>) -> Result<()> {
         let df: DataFrame = pydf.into();
@@ -135,20 +149,6 @@ impl SlimMSE {
         Ok(())
     }
 
-    pub fn fit_identified(&mut self, user_interactions: Vec<(i32, i32, f32, f32)>, add_interaction: Option<bool>, update_interaction: Option<bool>) {
-        for (user_id, item_id, tstamp, rating) in user_interactions {
-            if let Err(e) = {
-                if add_interaction.unwrap_or(true) {
-                    self.interactions.add_interaction(user_id, item_id, tstamp, rating, update_interaction.unwrap_or(false));
-                }
-                self.update_weights(user_id, item_id);
-                Ok::<(), Box<dyn std::error::Error>>(())
-            } {
-                warn!("Failed to fit interaction: {}", e);
-            }
-        }
-    }
-
     /// Bulk identify users and items from the provided interactions.
     #[inline]
     pub fn bulk_identify(&mut self, user_interactions: Vec<(SerializableValue, SerializableValue)>) -> Vec<(i32, i32)> {
@@ -193,7 +193,7 @@ impl SlimMSE {
         //
         // Vectorized updates for better performance.
         let updates: Vec<_> = user_items
-            .par_iter()
+            .iter()
             .filter(|&&ui| ui != item_id) // Exclude the target item_id
             .map(|&ui| {
                 let rating = self.interactions.get_user_item_rating(user_id, ui, 0.0);
@@ -203,7 +203,7 @@ impl SlimMSE {
             .filter(|(_, grad)| grad.abs() > 1e-6) // Skip very small gradients
             .collect();
 
-        self.ftrl.par_update_gradients(item_id, &updates);
+        self.ftrl.vectorized_update_gradients(item_id, &updates);
     }
 
     pub fn predict_rating(&self, user: SerializableValue, item: SerializableValue) -> f32 {