Jax scheduler lr decay

deepxde/model.py

@@ -118,6 +118,15 @@ def compile(
                     - `exponential_decay
                       <https://optax.readthedocs.io/en/latest/api/optimizer_schedules.html#optax.schedules.exponential_decay>`_:
                       ("exponential", transition_steps, decay_rate)
+                    - `warmup_cosine_decay_schedule
+                      <https://optax.readthedocs.io/en/latest/api/optimizer_schedules.html#optax.schedules.warmup_cosine_decay_schedule>`_:
+                      ("warmup_cosine", peak_value, warmup_steps, decay_steps, end_value)
+                    - `warmup_exponential_decay_schedule
+                      <https://optax.readthedocs.io/en/latest/api/optimizer_schedules.html#optax.schedules.warmup_exponential_decay_schedule>`_:
+                      ("warmup_exponential", peak_value, warmup_steps, transition_steps, decay_rate)
+
+                    Only required arguments are listed. Optional positional arguments can be provided in Optax order. 
+                    `init_value` is not in the tuple/list, since it is always supplied by the `lr` argument.
 
             loss_weights: A list specifying scalar coefficients (Python floats) to
                 weight the loss contributions. The loss value that will be minimized by

deepxde/optimizers/jax/optimizers.py

@@ -2,6 +2,7 @@
 
 import jax
 import optax
+from optax import schedules
 
 
 apply_updates = optax.apply_updates
@@ -37,12 +38,17 @@ def _get_learningrate(lr, decay):
     if decay is None:
         return lr
     if decay[0] == "linear":
-        return optax.linear_schedule(lr, decay[1], decay[2])
+        return schedules.linear_schedule(lr, *decay[1:])
     if decay[0] == "cosine":
-        return optax.cosine_decay_schedule(lr, decay[1], decay[2])
+        return schedules.cosine_decay_schedule(lr, *decay[1:])
     if decay[0] == "exponential":
-        return optax.exponential_decay(lr, decay[1], decay[2])
+        return schedules.exponential_decay(lr, *decay[1:])
+    if decay[0] == "warmup_cosine":
+        return schedules.warmup_cosine_decay_schedule(lr, *decay[1:])
+    if decay[0] == "warmup_exponential":
+        return schedules.warmup_exponential_decay_schedule(lr, *decay[1:])
 
     raise NotImplementedError(
-        f"{decay[0]} learning rate decay to be implemented for backend jax."
+        f"Unknown decay schedule '{decay[0]}' for JAX backend. "
+        f"Supported: 'linear', 'cosine', 'exponential', 'warmup_cosine', 'warmup_exponential'."
     )

examples/pinn_forward/Poisson_Dirichlet_1d_exactBC.py

@@ -49,8 +49,10 @@ def output_transform(x, y):
 net.apply_output_transform(output_transform)
 
 model = dde.Model(data, net)
+# Most backends
 model.compile("adam", lr=1e-4, decay=("inverse time", 1000, 0.3), metrics=["l2 relative error"])
-
+# Backend jax (does not support "inverse time" decay)
+# model.compile("adam", lr=1e-4, decay=("exponential", 1000, 0.9), metrics=["l2 relative error"])
 losshistory, train_state = model.train(iterations=30000)
 
 dde.saveplot(losshistory, train_state, issave=True, isplot=True)