Non-decreasing SInc(Hard) in same-day reviews (#376)

* Non-decreasing SInc(Hard) in same-day reviews

See: https://discord.com/channels/368267295601983490/1347982145418694747/1435352142973108285 and https://forums.ankiweb.net/t/fsrs-hard-acting-as-a-again-in-learning-relearning-phase/67334

* pass ci

---------

Co-authored-by: Jarrett Ye <jarrett.ye@outlook.com>

Author: Expertium

src/model.rs

@@ -115,7 +115,7 @@ impl<B: Backend> Model<B> {
         last_s
             * sinc
                 .clone()
-                .mask_where(rating.greater_equal_elem(3), sinc.clamp_min(1.0))
+                .mask_where(rating.greater_equal_elem(2), sinc.clamp_min(1.0))
     }
 
     fn mean_reversion(&self, new_d: Tensor<B, 1>) -> Tensor<B, 1> {
@@ -521,7 +521,7 @@ mod tests {
             .to_data()
             .to_vec::<f32>()
             .unwrap()
-            .assert_approx_eq([1.596818, 2.7470093, 5.0, 8.12961]);
+            .assert_approx_eq([1.596818, 5.0, 5.0, 8.12961]);
     }
 
     #[test]

src/training.rs

@@ -621,33 +621,33 @@ mod tests {
             Reduction::Sum,
         );
 
-        assert_eq!(loss.clone().into_scalar().to_f32(), 4.047898);
+        assert_eq!(loss.clone().into_scalar().to_f32(), 4.0466027);
         let gradients = loss.backward();
 
         let w_grad = model.w.grad(&gradients).unwrap();
 
         w_grad.to_data().to_vec::<f32>().unwrap().assert_approx_eq([
             -0.095688485,
             -0.0051607806,
-            -0.00080300873,
+            -0.0012249565,
             0.007462064,
-            0.03677408,
-            -0.084962785,
-            0.059571628,
-            -2.1566951,
-            0.5738574,
-            -2.8749206,
-            0.7123072,
+            0.03650761,
+            -0.082112335,
+            0.0593964,
+            -2.1474836,
+            0.57626534,
+            -2.8751316,
+            0.7154875,
             -0.028993709,
             0.0099172965,
             -0.2189217,
             -0.0017800558,
             -0.089381434,
             0.299141,
-            0.0708902,
-            -0.01219162,
-            -0.25424173,
-            0.27452517,
+            0.068104014,
+            -0.011605468,
+            -0.25398168,
+            0.27700496,
         ]);
 
         let config =
@@ -693,7 +693,7 @@ mod tests {
 
         let penalty =
             model.l2_regularization(init_w.clone(), params_stddev.clone(), 512, 1000, 2.0);
-        assert_eq!(penalty.clone().into_scalar().to_f32(), 0.6771115);
+        assert_eq!(penalty.clone().into_scalar().to_f32(), 0.67711145);
 
         let gradients = penalty.backward();
         let w_grad = model.w.grad(&gradients).unwrap();
@@ -757,7 +757,7 @@ mod tests {
             item.weights,
             Reduction::Sum,
         );
-        assert_eq!(loss.clone().into_scalar().to_f32(), 3.76888);
+        assert_eq!(loss.clone().into_scalar().to_f32(), 3.767796);
         let gradients = loss.backward();
         let w_grad = model.w.grad(&gradients).unwrap();
         w_grad
@@ -768,25 +768,25 @@ mod tests {
             .assert_approx_eq([
                 -0.040530164,
                 -0.0041278866,
-                -0.0006833144,
+                -0.0010157757,
                 0.007239434,
-                0.009416521,
-                -0.12156768,
-                0.039193563,
-                -0.86553144,
-                0.57743585,
-                -2.571437,
-                0.76415884,
+                0.009321215,
+                -0.120117955,
+                0.039143264,
+                -0.8628009,
+                0.5794302,
+                -2.5713828,
+                0.7669307,
                 -0.024242667,
                 0.0,
                 -0.16912507,
                 -0.0017008218,
                 -0.061857328,
                 0.28093633,
-                0.06636292,
-                0.0057900245,
-                -0.19041246,
-                0.6214733,
+                0.064058185,
+                0.0063592787,
+                -0.1903223,
+                0.6257775,
             ]);
         let grads = GradientsParams::from_grads(gradients, &model);
         model = optim.step(lr, model, grads);
@@ -802,9 +802,9 @@ mod tests {
             .to_vec::<f32>()
             .unwrap()
             .assert_approx_eq([
-                0.2882918, 1.3726242, 2.3862023, 8.215636, 6.339949, 0.9131501, 2.940647,
-                0.07696302, 1.7921939, 0.2464219, 0.71595156, 1.5631561, 0.001, 0.34230903,
-                1.7282416, 0.68038, 1.7929853, 0.46259063, 0.1426339, 0.14509763, 0.1,
+                0.2882918, 1.3726242, 2.3861322, 8.215636, 6.339965, 0.9130969, 2.940639,
+                0.07696985, 1.7921946, 0.2464217, 0.71595186, 1.5631561, 0.001, 0.34230903,
+                1.7282416, 0.68038, 1.7929853, 0.46258268, 0.14039303, 0.14509967, 0.1,
             ]);
     }