[examples/flax] add adafactor optimizer (#12544)

* add adafactor

* Update examples/flax/language-modeling/run_mlm_flax.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

examples/flax/language-modeling/requirements.txt

@@ -2,4 +2,4 @@ datasets >= 1.1.3
 jax>=0.2.8
 jaxlib>=0.1.59
 flax>=0.3.4
-optax>=0.0.8
+optax>=0.0.9

examples/flax/language-modeling/run_clm_flax.py

@@ -489,17 +489,24 @@ def main():
         return traverse_util.unflatten_dict(flat_mask)
 
     # create adam optimizer
-    adamw = optax.adamw(
-        learning_rate=linear_decay_lr_schedule_fn,
-        b1=training_args.adam_beta1,
-        b2=training_args.adam_beta2,
-        eps=training_args.adam_epsilon,
-        weight_decay=training_args.weight_decay,
-        mask=decay_mask_fn,
-    )
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
 
     # Setup train state
-    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
 
     def loss_fn(logits, labels):
         shift_logits = logits[..., :-1, :]

examples/flax/language-modeling/run_mlm_flax.py

@@ -513,17 +513,24 @@ if __name__ == "__main__":
         return traverse_util.unflatten_dict(flat_mask)
 
     # create adam optimizer
-    adamw = optax.adamw(
-        learning_rate=linear_decay_lr_schedule_fn,
-        b1=training_args.adam_beta1,
-        b2=training_args.adam_beta2,
-        eps=1e-8,
-        weight_decay=training_args.weight_decay,
-        mask=decay_mask_fn,
-    )
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
 
     # Setup train state
-    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw)
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
 
     # Define gradient update step fn
     def train_step(state, batch, dropout_rng):

examples/flax/language-modeling/run_t5_mlm_flax.py

@@ -635,16 +635,23 @@ if __name__ == "__main__":
         return traverse_util.unflatten_dict(flat_mask)
 
     # create adam optimizer
-    adamw = optax.adamw(
-        learning_rate=linear_decay_lr_schedule_fn,
-        b1=training_args.adam_beta1,
-        b2=training_args.adam_beta2,
-        weight_decay=training_args.weight_decay,
-        mask=decay_mask_fn,
-    )
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
 
     # Setup train state
-    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw)
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
 
     # Define gradient update step fn
     def train_step(state, batch, dropout_rng):