Spaces:
Running
Running
feat: use optax for gradient accumulation
Browse files- dev/seq2seq/run_seq2seq_flax.py +25 -43
dev/seq2seq/run_seq2seq_flax.py
CHANGED
@@ -280,9 +280,7 @@ class DataTrainingArguments:
|
|
280 |
|
281 |
|
282 |
class TrainState(train_state.TrainState):
|
283 |
-
dropout_rng: jnp.ndarray
|
284 |
-
grad_accum: jnp.ndarray
|
285 |
-
optimizer_step: int
|
286 |
|
287 |
def replicate(self):
|
288 |
return jax_utils.replicate(self).replace(
|
@@ -502,9 +500,8 @@ def main():
|
|
502 |
with (Path(artifact_dir) / "training_state.json").open("r") as f:
|
503 |
training_state = json.load(f)
|
504 |
step = training_state["step"]
|
505 |
-
optimizer_step = step // training_args.gradient_accumulation_steps
|
506 |
|
507 |
-
return step,
|
508 |
|
509 |
# Set up wandb run
|
510 |
wandb.init(
|
@@ -512,6 +509,7 @@ def main():
|
|
512 |
project="dalle-mini",
|
513 |
job_type="Seq2Seq",
|
514 |
config=parser.parse_args(),
|
|
|
515 |
)
|
516 |
|
517 |
# set default x-axis as 'train/step'
|
@@ -722,7 +720,7 @@ def main():
|
|
722 |
train_batch_size = (
|
723 |
int(training_args.per_device_train_batch_size) * jax.device_count()
|
724 |
)
|
725 |
-
|
726 |
eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
|
727 |
if data_args.streaming:
|
728 |
len_train_dataset = data_args.len_train
|
@@ -743,12 +741,12 @@ def main():
|
|
743 |
len_eval_dataset = len(eval_dataset)
|
744 |
steps_per_epoch = len_train_dataset // train_batch_size
|
745 |
total_steps = steps_per_epoch * num_epochs
|
746 |
-
total_optimization_steps = (len_train_dataset //
|
747 |
|
748 |
# Create learning rate schedule
|
749 |
-
|
750 |
len_train_dataset,
|
751 |
-
|
752 |
training_args.num_train_epochs,
|
753 |
training_args.warmup_steps,
|
754 |
training_args.learning_rate,
|
@@ -783,11 +781,11 @@ def main():
|
|
783 |
# We use the default parameters here to initialize adafactor,
|
784 |
# For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
|
785 |
optimizer = optax.adafactor(
|
786 |
-
learning_rate=
|
787 |
)
|
788 |
else:
|
789 |
optimizer = optax.adamw(
|
790 |
-
learning_rate=
|
791 |
b1=training_args.adam_beta1,
|
792 |
b2=training_args.adam_beta2,
|
793 |
eps=training_args.adam_epsilon,
|
@@ -795,21 +793,24 @@ def main():
|
|
795 |
mask=decay_mask_fn,
|
796 |
)
|
797 |
|
|
|
|
|
|
|
|
|
|
|
|
|
798 |
# Setup train state
|
799 |
state = TrainState.create(
|
800 |
apply_fn=model.__call__,
|
801 |
params=model.params,
|
802 |
tx=optimizer,
|
803 |
dropout_rng=dropout_rng,
|
804 |
-
grad_accum=jax.tree_map(jnp.zeros_like, model.params),
|
805 |
-
optimizer_step=0,
|
806 |
)
|
807 |
if model_args.from_checkpoint is not None:
|
808 |
-
# restore optimizer state
|
809 |
-
step,
|
810 |
-
state = state.replace(
|
811 |
-
|
812 |
-
)
|
813 |
|
814 |
# label smoothed cross entropy
|
815 |
def loss_fn(logits, labels):
|
@@ -821,7 +822,7 @@ def main():
|
|
821 |
def train_step(state, batch):
|
822 |
dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
|
823 |
|
824 |
-
def compute_loss(params):
|
825 |
labels = batch.pop("labels")
|
826 |
logits = state.apply_fn(
|
827 |
**batch, params=params, dropout_rng=dropout_rng, train=True
|
@@ -830,35 +831,16 @@ def main():
|
|
830 |
return loss
|
831 |
|
832 |
grad_fn = jax.value_and_grad(compute_loss)
|
833 |
-
loss, grads = grad_fn(state.params)
|
834 |
-
|
835 |
-
|
836 |
-
def update_fn():
|
837 |
-
grads = jax.tree_map(
|
838 |
-
lambda x: x / training_args.gradient_accumulation_steps, grad_accum
|
839 |
-
)
|
840 |
-
grads = jax.lax.pmean(grads, "batch")
|
841 |
-
new_state = state.apply_gradients(
|
842 |
-
grads=grads,
|
843 |
-
grad_accum=jax.tree_map(jnp.zeros_like, grads),
|
844 |
-
optimizer_step=state.optimizer_step + 1,
|
845 |
-
)
|
846 |
-
return new_state
|
847 |
-
|
848 |
-
new_state = jax.lax.cond(
|
849 |
-
(state.step + 1) % training_args.gradient_accumulation_steps == 0,
|
850 |
-
lambda _: update_fn(),
|
851 |
-
lambda _: state.replace(grad_accum=grad_accum, step=state.step + 1),
|
852 |
-
None,
|
853 |
-
)
|
854 |
|
855 |
metrics = {
|
856 |
"loss": loss,
|
857 |
-
"learning_rate":
|
858 |
}
|
859 |
metrics = jax.lax.pmean(metrics, axis_name="batch")
|
860 |
-
|
861 |
-
return new_state.replace(dropout_rng=new_dropout_rng), metrics
|
862 |
|
863 |
# Define eval fn
|
864 |
def eval_step(params, batch):
|
|
|
280 |
|
281 |
|
282 |
class TrainState(train_state.TrainState):
|
283 |
+
dropout_rng: jnp.ndarray = None
|
|
|
|
|
284 |
|
285 |
def replicate(self):
|
286 |
return jax_utils.replicate(self).replace(
|
|
|
500 |
with (Path(artifact_dir) / "training_state.json").open("r") as f:
|
501 |
training_state = json.load(f)
|
502 |
step = training_state["step"]
|
|
|
503 |
|
504 |
+
return step, opt_state
|
505 |
|
506 |
# Set up wandb run
|
507 |
wandb.init(
|
|
|
509 |
project="dalle-mini",
|
510 |
job_type="Seq2Seq",
|
511 |
config=parser.parse_args(),
|
512 |
+
save_code=True,
|
513 |
)
|
514 |
|
515 |
# set default x-axis as 'train/step'
|
|
|
720 |
train_batch_size = (
|
721 |
int(training_args.per_device_train_batch_size) * jax.device_count()
|
722 |
)
|
723 |
+
batch_size_per_update = train_batch_size * training_args.gradient_accumulation_steps
|
724 |
eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
|
725 |
if data_args.streaming:
|
726 |
len_train_dataset = data_args.len_train
|
|
|
741 |
len_eval_dataset = len(eval_dataset)
|
742 |
steps_per_epoch = len_train_dataset // train_batch_size
|
743 |
total_steps = steps_per_epoch * num_epochs
|
744 |
+
total_optimization_steps = (len_train_dataset // batch_size_per_update) * num_epochs
|
745 |
|
746 |
# Create learning rate schedule
|
747 |
+
learning_rate_fn = create_learning_rate_fn(
|
748 |
len_train_dataset,
|
749 |
+
train_batch_size,
|
750 |
training_args.num_train_epochs,
|
751 |
training_args.warmup_steps,
|
752 |
training_args.learning_rate,
|
|
|
781 |
# We use the default parameters here to initialize adafactor,
|
782 |
# For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
|
783 |
optimizer = optax.adafactor(
|
784 |
+
learning_rate=learning_rate_fn,
|
785 |
)
|
786 |
else:
|
787 |
optimizer = optax.adamw(
|
788 |
+
learning_rate=learning_rate_fn,
|
789 |
b1=training_args.adam_beta1,
|
790 |
b2=training_args.adam_beta2,
|
791 |
eps=training_args.adam_epsilon,
|
|
|
793 |
mask=decay_mask_fn,
|
794 |
)
|
795 |
|
796 |
+
# add gradient accumulation
|
797 |
+
if training_args.gradient_accumulation_steps > 1:
|
798 |
+
optimizer = optax.chain(
|
799 |
+
optax.apply_every(training_args.gradient_accumulation_steps), optimizer
|
800 |
+
)
|
801 |
+
|
802 |
# Setup train state
|
803 |
state = TrainState.create(
|
804 |
apply_fn=model.__call__,
|
805 |
params=model.params,
|
806 |
tx=optimizer,
|
807 |
dropout_rng=dropout_rng,
|
|
|
|
|
808 |
)
|
809 |
if model_args.from_checkpoint is not None:
|
810 |
+
# restore optimizer state and step
|
811 |
+
step, opt_state = restore_state(state, artifact_dir)
|
812 |
+
state = state.replace(step=step, opt_state=opt_state)
|
813 |
+
# TODO: number of remaining training epochs/steps and dataloader state need to be adjusted
|
|
|
814 |
|
815 |
# label smoothed cross entropy
|
816 |
def loss_fn(logits, labels):
|
|
|
822 |
def train_step(state, batch):
|
823 |
dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
|
824 |
|
825 |
+
def compute_loss(params, batch):
|
826 |
labels = batch.pop("labels")
|
827 |
logits = state.apply_fn(
|
828 |
**batch, params=params, dropout_rng=dropout_rng, train=True
|
|
|
831 |
return loss
|
832 |
|
833 |
grad_fn = jax.value_and_grad(compute_loss)
|
834 |
+
loss, grads = grad_fn(state.params, batch)
|
835 |
+
grads = jax.lax.pmean(grads, "batch")
|
836 |
+
state = state.apply_gradients(grads=grads)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
837 |
|
838 |
metrics = {
|
839 |
"loss": loss,
|
840 |
+
"learning_rate": learning_rate_fn(state.step),
|
841 |
}
|
842 |
metrics = jax.lax.pmean(metrics, axis_name="batch")
|
843 |
+
return state.replace(dropout_rng=new_dropout_rng), metrics
|
|
|
844 |
|
845 |
# Define eval fn
|
846 |
def eval_step(params, batch):
|