Add some lr schedulers with warmup (#5256)

* add -linear-schedule-with-warmup

* update lr scheduler usage in run_glue

* add import info

* add other 5 scheduler with warmup

* add copyright, update usage in run_glue

* simplify argument, make warmup arg support float and int

* Add some LambdaDecay scheduler with warmup, update usage in run_glue

* update classname and unify two cosine decays into one

* update usage in run_glue

* fix typo

* update WarmUp to Warmup, and update class name about Const, and update doc, and usage

* update usage of decay class

* update usage of decay class

PaddleNLP/benchmark/bert/run_glue.py

@@ -29,6 +29,7 @@ from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.data.sampler import SamplerHelper
 from paddlenlp.transformers import BertForSequenceClassification, BertTokenizer
 from paddlenlp.transformers import ErnieForSequenceClassification, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.metrics import Mcc, PearsonAndSpearman
 from paddlenlp.utils.log import logger
 
@@ -381,17 +382,11 @@ def do_train(args):
 
     # Create the training-backward program, this pass will not be
     # executed in the validation
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
     with paddle.static.program_guard(main_program, startup_program):
-        lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-            args.learning_rate,
-            lambda current_step, num_warmup_steps=args.warmup_steps,
-            num_training_steps=args.max_steps if args.max_steps > 0 else
-            (len(train_data_loader) * args.num_train_epochs): float(
-                current_step) / float(max(1, num_warmup_steps))
-            if current_step < num_warmup_steps else max(
-               0.0,
-               float(num_training_steps - current_step) / float(
-                  max(1, num_training_steps - num_warmup_steps))))
+        lr_scheduler = LinearDecayWithWarmup(
+            args.learning_rate, num_training_steps, args.warmup_steps)
         optimizer = paddle.optimizer.AdamW(
             learning_rate=lr_scheduler,
             epsilon=args.adam_epsilon,

PaddleNLP/benchmark/bert/run_pretrain.py

@@ -31,6 +31,7 @@ from paddle.io import DataLoader, Dataset
 
 from paddlenlp.transformers import BertForPretraining, BertModel, BertPretrainingCriterion
 from paddlenlp.transformers import BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from data import create_data_holder, create_pretraining_dataset
 
 MODEL_CLASSES = {"bert": (BertForPretraining, BertTokenizer)}
@@ -232,7 +233,8 @@ def dist_optimizer(args, optimizer):
     if args.use_amp:
         dist_strategy.amp = True
 
-        custom_black_list = ['lookup_table', 'lookup_table_v2'] if args.use_pure_fp16 else None
+        custom_black_list = ['lookup_table',
+                             'lookup_table_v2'] if args.use_pure_fp16 else None
         dist_strategy.amp_configs = {
             'custom_white_list': ['softmax', 'layer_norm', 'gelu'],
             'init_loss_scaling': args.scale_loss,
@@ -305,16 +307,11 @@ def do_train(args):
                      masked_lm_labels, next_sentence_labels, masked_lm_scale)
 
     # Define the dynamic learing_reate scheduler and optimizer
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,
@@ -327,7 +324,7 @@ def do_train(args):
         ],
         multi_precision=args.use_pure_fp16)
     if worker_num == 1 and args.use_amp:
-        custom_black_list=(['lookup_table', 'lookup_table_v2']
+        custom_black_list = (['lookup_table', 'lookup_table_v2']
                              if args.use_pure_fp16 else None)
         amp_list = paddle.static.amp.AutoMixedPrecisionLists(
             custom_white_list=['softmax', 'layer_norm', 'gelu'],

PaddleNLP/benchmark/bert/run_pretrain_single.py

@@ -28,6 +28,7 @@ from paddle.io import DataLoader, Dataset
 from paddlenlp.transformers import BertForPretraining, BertModel, BertPretrainingCriterion
 from paddlenlp.transformers import ErnieForPretraining, ErnieModel, ErniePretrainingCriterion
 from paddlenlp.transformers import BertTokenizer, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from data import create_data_holder, create_pretraining_dataset
 
 MODEL_CLASSES = {
@@ -222,16 +223,11 @@ def do_train(args):
                      masked_lm_labels, next_sentence_labels, masked_lm_scale)
 
     # Define the dynamic learing_reate scheduler and optimizer
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,
@@ -244,7 +240,7 @@ def do_train(args):
         ],
         multi_precision=args.use_pure_fp16)
     if args.use_amp:
-        custom_black_list=(['lookup_table', 'lookup_table_v2']
+        custom_black_list = (['lookup_table', 'lookup_table_v2']
                              if args.use_pure_fp16 else None)
         amp_list = paddle.static.amp.AutoMixedPrecisionLists(
             custom_white_list=['layer_norm', 'softmax', 'gelu'],

PaddleNLP/examples/dialogue/dgu/main.py

@@ -9,13 +9,13 @@ import paddle.nn as nn
 import paddle.nn.functional as F
 import paddle.distributed as dist
 from paddle.io import DataLoader, DistributedBatchSampler, BatchSampler
-from paddle.optimizer.lr import LambdaDecay
 from paddle.optimizer import AdamW
 from paddle.metric import Accuracy
 
 from paddlenlp.datasets import MapDatasetWrapper
 from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.transformers import BertTokenizer, BertForSequenceClassification, BertForTokenClassification
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 from args import parse_args, set_default_args
 import data
@@ -54,14 +54,6 @@ def save_ckpt(model, optimizer, output_dir, name):
     paddle.save(optimizer.state_dict(), opt_path)
 
 
-def compute_lr_factor(current_step, warmup_steps, max_train_steps):
-    if current_step < warmup_steps:
-        factor = float(current_step) / warmup_steps
-    else:
-        factor = 1 - float(current_step) / max_train_steps
-    return factor
-
-
 class DGULossFunction(nn.Layer):
     def __init__(self, task_name):
         super(DGULossFunction, self).__init__()
@@ -117,16 +109,14 @@ def print_logs(args, step, logits, labels, loss, total_time, metric):
 def train(args, model, train_data_loader, dev_data_loader, metric, rank):
     num_examples = len(train_data_loader) * args.batch_size * args.n_gpu
     max_train_steps = args.epochs * len(train_data_loader)
-    warmup_steps = int(max_train_steps * args.warmup_proportion)
     if rank == 0:
         print("Num train examples: %d" % num_examples)
         print("Max train steps: %d" % max_train_steps)
-        print("Num warmup steps: %d" % warmup_steps)
-    factor_fn = partial(
-        compute_lr_factor,
-        warmup_steps=warmup_steps,
-        max_train_steps=max_train_steps)
-    lr_scheduler = LambdaDecay(args.learning_rate, factor_fn)
+        print("Warmup proportion: %d" % args.warmup_proportion)
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, max_train_steps,
+                                         args.warmup_proportion)
+
     optimizer = AdamW(
         learning_rate=lr_scheduler,
         parameters=model.parameters(),

PaddleNLP/examples/glue/run_glue.py

@@ -32,6 +32,7 @@ from paddlenlp.data.sampler import SamplerHelper
 from paddlenlp.transformers import BertForSequenceClassification, BertTokenizer
 from paddlenlp.transformers import ElectraForSequenceClassification, ElectraTokenizer
 from paddlenlp.transformers import ErnieForSequenceClassification, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.metrics import AccuracyAndF1, Mcc, PearsonAndSpearman
 
 FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
@@ -355,17 +356,10 @@ def do_train(args):
 
     num_training_steps = args.max_steps if args.max_steps > 0 else (
         len(train_data_loader) * args.num_train_epochs)
-    warmup_steps = args.warmup_steps if args.warmup_steps > 0 else (
-        int(math.floor(num_training_steps * args.warmup_proportion)))
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=warmup_steps,
-        num_training_steps=num_training_steps : float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    warmup = args.warmup_steps if args.warmup_steps > 0 else args.warmup_proportion
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         warmup)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/language_model/bert/run_glue.py

@@ -32,6 +32,7 @@ from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.transformers import BertForSequenceClassification, BertTokenizer
 from paddlenlp.transformers import ElectraForSequenceClassification, ElectraTokenizer
 from paddlenlp.transformers import ErnieForSequenceClassification, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.metrics import AccuracyAndF1, Mcc, PearsonAndSpearman
 
 FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
@@ -162,11 +163,13 @@ def parse_args():
         type=str,
         default="gpu",
         help="Device for selecting for the training.")
-    parser.add_argument("--use_amp",
+    parser.add_argument(
+        "--use_amp",
         type=distutils.util.strtobool,
         default=False,
         help="Enable mixed precision training.")
-    parser.add_argument("--scale_loss",
+    parser.add_argument(
+        "--scale_loss",
         type=float,
         default=2**15,
         help="The value of scale_loss for fp16.")
@@ -360,17 +363,10 @@ def do_train(args):
 
     num_training_steps = args.max_steps if args.max_steps > 0 else (
         len(train_data_loader) * args.num_train_epochs)
-    warmup_steps = args.warmup_steps if args.warmup_steps > 0 else (
-        int(math.floor(num_training_steps * args.warmup_proportion)))
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=warmup_steps,
-        num_training_steps=num_training_steps : float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    warmup = args.warmup_steps if args.warmup_steps > 0 else args.warmup_proportion
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         warmup)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/language_model/bert/run_pretrain.py

@@ -34,6 +34,7 @@ from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.transformers import BertForPretraining, BertModel, BertPretrainingCriterion
 from paddlenlp.transformers import ErnieForPretraining, ErnieModel, ErniePretrainingCriterion
 from paddlenlp.transformers import BertTokenizer, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
 logging.basicConfig(level=logging.INFO, format=FORMAT)
@@ -147,11 +148,13 @@ def parse_args():
         type=str,
         default="gpu",
         help="Device for selecting for the training.")
-    parser.add_argument("--use_amp",
+    parser.add_argument(
+        "--use_amp",
         type=distutils.util.strtobool,
         default=False,
         help="Enable mixed precision training.")
-    parser.add_argument("--scale_loss",
+    parser.add_argument(
+        "--scale_loss",
         type=float,
         default=2**15,
         help="The value of scale_loss for fp16.")
@@ -301,17 +304,11 @@ def do_train(args):
 
     # If use defalut last_epoch, lr of the first iteration is 0.
     # Use `last_epoch = 0` to be consistent with nv bert.
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))),
-        last_epoch=0)
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(
+        args.learning_rate, num_training_steps, args.warmup_steps, last_epoch=0)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/language_model/electra/run_glue.py

@@ -31,6 +31,7 @@ from paddlenlp.datasets import GlueCoLA, GlueSST2, GlueMRPC, GlueSTSB, GlueQQP,
 from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.data.sampler import SamplerHelper
 from paddlenlp.transformers import ElectraForSequenceClassification, ElectraTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.utils.log import logger
 from paddlenlp.metrics import AccuracyAndF1, Mcc, PearsonAndSpearman
 
@@ -226,16 +227,9 @@ def do_train(args):
 
     num_training_steps = args.max_steps if args.max_steps > 0 else (
         len(train_data_loader) * args.num_train_epochs)
-    warmup_steps = int(math.floor(num_training_steps * args.warmup_proportion))
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=warmup_steps,
-        num_training_steps=num_training_steps : float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_proportion)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/language_model/electra/run_pretrain.py

@@ -32,6 +32,7 @@ from paddle.io import DataLoader, Dataset
 from paddlenlp.transformers import ElectraForTotalPretraining, ElectraModel, ElectraPretrainingCriterion
 from paddlenlp.transformers import ElectraDiscriminator, ElectraGenerator
 from paddlenlp.transformers import ElectraTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
 logging.basicConfig(level=logging.INFO, format=FORMAT)
@@ -470,15 +471,9 @@ def do_train(args):
 
     num_training_steps = args.max_steps if args.max_steps > 0 else (
         len(train_data_loader) * args.num_train_epochs)
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=num_training_steps: float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
     optimizer = paddle.optimizer.AdamW(

PaddleNLP/examples/machine_reading_comprehension/DuReader-robust/run_du.py

@@ -29,6 +29,7 @@ import paddlenlp as ppnlp
 from paddlenlp.datasets import SQuAD, DuReaderRobust, CMRC, DRCD
 from paddlenlp.data import Pad, Stack, Tuple
 from paddlenlp.transformers import BertForQuestionAnswering, BertTokenizer, ErnieForQuestionAnswering, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.metrics.squad import squad_evaluate, compute_predictions
 
 TASK_CLASSES = {"dureader-robust": DuReaderRobust, "cmrc": CMRC, "drcd": DRCD}
@@ -177,16 +178,11 @@ def do_train(args):
     if paddle.distributed.get_world_size() > 1:
         model = paddle.DataParallel(model)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, warmup_proportion=args.warmup_proportion,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader)*args.num_train_epochs): float(
-            current_step) / float(max(1, warmup_proportion*num_training_steps))
-        if current_step < warmup_proportion*num_training_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - warmup_proportion*num_training_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_proportion)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/machine_reading_comprehension/DuReader-yesno/run_du.py

@@ -28,6 +28,7 @@ import paddlenlp as ppnlp
 
 from paddlenlp.data import Pad, Stack, Tuple
 from paddlenlp.transformers import BertForSequenceClassification, BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 MODEL_CLASSES = {"bert": (BertForSequenceClassification, BertTokenizer)}
 
@@ -210,16 +211,11 @@ def do_train(args):
     if paddle.distributed.get_world_size() > 1:
         model = paddle.DataParallel(model)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, warmup_proportion=args.warmup_proportion,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_ds.examples)//args.batch_size*args.num_train_epochs): float(
-            current_step) / float(max(1, warmup_proportion*num_training_steps))
-        if current_step < warmup_proportion*num_training_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - warmup_proportion*num_training_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_ds.examples) // args.batch_size * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_proportion)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/machine_reading_comprehension/DuReader/run_du.py

@@ -28,6 +28,7 @@ import paddlenlp as ppnlp
 
 from paddlenlp.data import Pad, Stack, Tuple
 from paddlenlp.transformers import BertForQuestionAnswering, BertTokenizer, ErnieForQuestionAnswering, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.metrics.dureader import dureader_evaluate, compute_predictions
 
 MODEL_CLASSES = {
@@ -168,16 +169,11 @@ def do_train(args):
     if paddle.distributed.get_world_size() > 1:
         model = paddle.DataParallel(model)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, warmup_proportion=args.warmup_proportion,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_dataset.examples)//args.batch_size*args.num_train_epochs): float(
-            current_step) / float(max(1, warmup_proportion*num_training_steps))
-        if current_step < warmup_proportion*num_training_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - warmup_proportion*num_training_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_dataset.examples) // args.batch_size * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_proportion)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/machine_reading_comprehension/SQuAD/run_squad.py

@@ -28,6 +28,7 @@ import paddlenlp as ppnlp
 
 from paddlenlp.data import Pad, Stack, Tuple
 from paddlenlp.transformers import BertForQuestionAnswering, BertTokenizer, ErnieForQuestionAnswering, ErnieTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.metrics.squad import squad_evaluate, compute_predictions
 
 MODEL_CLASSES = {
@@ -168,16 +169,11 @@ def do_train(args):
     if paddle.distributed.get_world_size() > 1:
         model = paddle.DataParallel(model)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, warmup_proportion=args.warmup_proportion,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_dataset.examples)//args.batch_size*args.num_train_epochs): float(
-            current_step) / float(max(1, warmup_proportion*num_training_steps))
-        if current_step < warmup_proportion*num_training_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - warmup_proportion*num_training_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_dataset.examples) // args.batch_size * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_proportion)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/model_compression/run_glue_ofa.py

@@ -28,6 +28,7 @@ from paddle.metric import Metric, Accuracy, Precision, Recall
 from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.data.sampler import SamplerHelper
 from paddlenlp.transformers import BertModel, BertForSequenceClassification, BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.utils.log import logger
 from paddlenlp.metrics import AccuracyAndF1, Mcc, PearsonAndSpearman
 import paddlenlp.datasets as datasets
@@ -459,16 +460,11 @@ def do_train(args):
         num_heads=model.bert.config['num_attention_heads'])
     reorder_neuron_head(ofa_model.model, head_importance, neuron_importance)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/named_entity_recognition/msra_ner/run_msra_ner.py

@@ -27,6 +27,7 @@ import paddlenlp as ppnlp
 from paddlenlp.datasets import MSRA_NER
 from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.transformers import BertForTokenClassification, BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.metrics import ChunkEvaluator
 
 
@@ -291,16 +292,11 @@ def do_train(args):
     if paddle.distributed.get_world_size() > 1:
         model = paddle.DataParallel(model)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/text_classification/pretrained_models/train.py

@@ -22,8 +22,9 @@ import numpy as np
 import paddle
 import paddle.nn.functional as F
 
-from paddlenlp.data import Stack, Tuple, Pad
 import paddlenlp as ppnlp
+from paddlenlp.data import Stack, Tuple, Pad
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 # yapf: disable
 parser = argparse.ArgumentParser()
@@ -214,19 +215,10 @@ def do_train():
     model = paddle.DataParallel(model)
 
     num_training_steps = len(train_data_loader) * args.epochs
-    num_warmup_steps = int(args.warmup_proportion * num_training_steps)
 
-    def get_lr_factor(current_step):
-        if current_step < num_warmup_steps:
-            return float(current_step) / float(max(1, num_warmup_steps))
-        else:
-            return max(0.0,
-                       float(num_training_steps - current_step) /
-                       float(max(1, num_training_steps - num_warmup_steps)))
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_proportion)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lr_lambda=lambda current_step: get_lr_factor(current_step))
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,
         parameters=model.parameters(),

PaddleNLP/examples/text_generation/ernie-gen/train.py

@@ -10,6 +10,7 @@ import paddle.nn as nn
 from paddle.io import DataLoader
 from paddlenlp.transformers import ErnieForGeneration
 from paddlenlp.transformers import ErnieTokenizer, ErnieTinyTokenizer, BertTokenizer, ElectraTokenizer, RobertaTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.datasets import Poetry
 from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.metrics import Rouge1, Rouge2
@@ -175,16 +176,10 @@ def train():
     if paddle.distributed.get_world_size() > 1:
         model = paddle.DataParallel(model)
 
-    max_steps = (len(train_data_loader) * args.num_epochs)
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=max_steps*args.warmup_proportion,
-        num_training_steps=max_steps: float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    max_steps = len(train_data_loader) * args.num_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, max_steps,
+                                         args.warmup_proportion)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/examples/text_matching/sentence_transformers/train.py

@@ -22,8 +22,9 @@ import numpy as np
 import paddle
 import paddle.nn.functional as F
 
-from paddlenlp.data import Stack, Tuple, Pad
 import paddlenlp as ppnlp
+from paddlenlp.data import Stack, Tuple, Pad
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 from model import SentenceTransformer
 
@@ -231,19 +232,10 @@ def do_train():
     model = paddle.DataParallel(model)
 
     num_training_steps = len(train_data_loader) * args.epochs
-    num_warmup_steps = int(args.warmup_proption * num_training_steps)
 
-    def get_lr_factor(current_step):
-        if current_step < num_warmup_steps:
-            return float(current_step) / float(max(1, num_warmup_steps))
-        else:
-            return max(0.0,
-                       float(num_training_steps - current_step) /
-                       float(max(1, num_training_steps - num_warmup_steps)))
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_proportion)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lr_lambda=lambda current_step: get_lr_factor(current_step))
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,
         parameters=model.parameters(),

PaddleNLP/legacy/benchmark/bert/run_glue.py

@@ -27,6 +27,7 @@ from paddlenlp.datasets import GlueQNLI, GlueSST2
 from paddlenlp.data import Stack, Tuple, Pad
 from paddlenlp.data.sampler import SamplerHelper
 from paddlenlp.transformers import BertForSequenceClassification, BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
 logging.basicConfig(level=logging.INFO, format=FORMAT)
@@ -332,16 +333,12 @@ def do_train(args):
     # Create the training-backward program, this pass will not be
     # executed in the validation
     with paddle.static.program_guard(main_program, startup_program):
-        lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-            args.learning_rate,
-            lambda current_step, num_warmup_steps=args.warmup_steps,
-            num_training_steps=args.max_steps if args.max_steps > 0 else
-            (len(train_data_loader) * args.num_train_epochs): float(
-                current_step) / float(max(1, num_warmup_steps))
-            if current_step < num_warmup_steps else max(
-               0.0,
-               float(num_training_steps - current_step) / float(
-                  max(1, num_training_steps - num_warmup_steps))))
+        num_training_steps = args.max_steps if args.max_steps > 0 else len(
+            train_data_loader) * args.num_train_epochs
+
+        lr_scheduler = LinearDecayWithWarmup(
+            args.learning_rate, num_training_steps, args.warmup_steps)
+
         optimizer = paddle.optimizer.AdamW(
             learning_rate=lr_scheduler,
             epsilon=args.adam_epsilon,

PaddleNLP/legacy/benchmark/bert/run_pretrain.py

@@ -28,6 +28,7 @@ from paddle.io import DataLoader, Dataset
 
 from paddlenlp.transformers import BertForPretraining, BertModel, BertPretrainingCriterion
 from paddlenlp.transformers import BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from data import create_data_holder, create_pretraining_dataset
 
 MODEL_CLASSES = {"bert": (BertForPretraining, BertTokenizer)}
@@ -198,17 +199,11 @@ def do_train(args):
     loss = criterion(prediction_scores, seq_relationship_score,
                      masked_lm_labels, next_sentence_labels, masked_lm_scale)
 
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
     # Define the dynamic learing_reate scheduler and optimizer
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/legacy/benchmark/bert/run_pretrain_single.py

@@ -27,6 +27,7 @@ import paddle
 from paddle.io import DataLoader, Dataset
 from paddlenlp.transformers import BertForPretraining, BertModel, BertPretrainingCriterion
 from paddlenlp.transformers import BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 from data import create_data_holder, create_pretraining_dataset
 
 MODEL_CLASSES = {"bert": (BertForPretraining, BertTokenizer)}
@@ -210,17 +211,11 @@ def do_train(args):
     loss = criterion(prediction_scores, seq_relationship_score,
                      masked_lm_labels, next_sentence_labels, masked_lm_scale)
 
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
     # Define the dynamic learing_reate scheduler and optimizer
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/paddlenlp/transformers/__init__.py

@@ -27,3 +27,4 @@ from .electra.modeling import *
 from .electra.tokenizer import *
 from .transformer.modeling import *
 from .ernie_gen.modeling import ErnieForGeneration
+from .optimization import *

PaddleNLP/paddlenlp/transformers/converter/run_glue_pp.py

@@ -31,6 +31,7 @@ from paddlenlp.data import *
 from paddlenlp.data.sampler import SamplerHelper
 from paddlenlp.transformers.model_bert import *
 from paddlenlp.transformers.tokenizer_bert import BertTokenizer
+from paddlenlp.transformers import LinearDecayWithWarmup
 
 TASK_CLASSES = {
     "qnli": (GlueQNLI, paddle.metric.Accuracy),  # (dataset, metric)
@@ -307,16 +308,11 @@ def do_train(args):
     if paddle.distributed.get_world_size() > 1:
         model = paddle.DataParallel(model)
 
-    lr_scheduler = paddle.optimizer.lr.LambdaDecay(
-        args.learning_rate,
-        lambda current_step, num_warmup_steps=args.warmup_steps,
-        num_training_steps=args.max_steps if args.max_steps > 0 else
-        (len(train_data_loader) * args.num_train_epochs): float(
-            current_step) / float(max(1, num_warmup_steps))
-        if current_step < num_warmup_steps else max(
-            0.0,
-            float(num_training_steps - current_step) / float(
-                max(1, num_training_steps - num_warmup_steps))))
+    num_training_steps = args.max_steps if args.max_steps > 0 else len(
+        train_data_loader) * args.num_train_epochs
+
+    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
+                                         args.warmup_steps)
 
     optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,

PaddleNLP/paddlenlp/transformers/optimization.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import math
+
+from paddle.optimizer.lr import LambdaDecay
+
+__all__ = [
+    'LinearDecayWithWarmup', 'ConstScheduleWithWarmup', 'CosineDecayWithWarmup',
+    'PolyDecayWithWarmup'
+]
+
+
+def is_integer(number):
+    if sys.version > '3':
+        return isinstance(number, int)
+    return isinstance(number, (int, long))
+
+
+class LinearDecayWithWarmup(LambdaDecay):
+    """
+    Create a learning rate scheduler, which increases learning rate linearly
+    from 0 to given `learning_rate`, after this warmup period learning rate
+    would be decreased linearly from the base learning rate to 0.
+
+    Args:
+        learning_rate (float): The base learning rate. It is a python float
+            number.
+        total_steps (int): The number of training steps.
+        warmup (int|float): If int, it means the number of steps for warmup.
+            If float, it means the proportion of warmup in total training steps. 
+        last_epoch (int, optional): The index of last epoch. It can be set to
+            restart training. If None, it means initial learning rate. 
+            Default: -1.
+        verbose (bool, optional): If True, prints a message to stdout for each
+            update. Default: False.
+
+    Examples:
+        
+        .. code-block:: python
+
+            from paddlenlp.transformers import LinearDecayWithWarmup
+            lr, warmup_steps, max_steps = 0.1, 100, 1000
+            lr_scheduler = LinearDecayWithWarmup(lr, max_steps, warmup_steps)
+
+    """
+
+    def __init__(self,
+                 learning_rate,
+                 total_steps,
+                 warmup,
+                 last_epoch=-1,
+                 verbose=False):
+        warmup_steps = warmup if is_integer(warmup) else int(
+            math.floor(warmup * total_steps))
+
+        def lr_lambda(current_step):
+            if current_step < warmup_steps:
+                return float(current_step) / float(max(1, warmup_steps))
+            return max(0.0,
+                       float(total_steps - current_step) /
+                       float(max(1, total_steps - warmup_steps)))
+
+        super(LinearDecayWithWarmup, self).__init__(learning_rate, lr_lambda,
+                                                    last_epoch, verbose)
+
+
+class ConstScheduleWithWarmup(LambdaDecay):
+    """
+    Create a learning rate scheduler, which increases learning rate linearly
+    from 0 to given `learning_rate` during warmup periods and keeps learning
+    rate a constant after that.
+
+    Args:
+        learning_rate (float): The base learning rate. It is a python float
+            number.
+        warmup (int|float): If int, it means the number of steps for warmup.
+            If float, it means the proportion of warmup in total training steps.
+        total_steps (int, optional): The number of training steps. If `warmup`
+            is a float number, `total_steps` must be provided.
+        last_epoch (int, optional): The index of last epoch. It can be set to
+            restart training. If None, it means initial learning rate. 
+            Default: -1.
+
+    Examples:
+        
+        .. code-block:: python
+
+            from paddlenlp.transformers import ConstScheduleWithWarmup
+            lr, warmup_steps = 0.1, 100
+            lr_scheduler = ConstScheduleWithWarmup(lr, warmup_steps)
+
+    """
+
+    def __init__(self,
+                 learning_rate,
+                 warmup,
+                 total_steps=None,
+                 last_epoch=-1,
+                 verbose=False):
+        warmup_steps = warmup if is_integer(warmup) else int(
+            math.floor(warmup * total_steps))
+        if is_integer(warmup):
+            warmup_steps = warmup
+        elif total_steps:
+            warmup_steps = int(math.floor(warmup * total_steps))
+        else:
+            raise ValueError(
+                "Please provide total steps if `warmup` is a float number , or provide integer for argument `warmup`."
+            )
+
+        def lr_lambda(current_step):
+            if current_step < warmup_steps:
+                return float(current_step) / float(max(1.0, warmup_steps))
+            return 1.0
+
+        super(ConstScheduleWithWarmup, self).__init__(learning_rate, lr_lambda,
+                                                      last_epoch, verbose)
+
+
+class CosineDecayWithWarmup(LambdaDecay):
+    """
+    Create a learning rate scheduler, which increases learning rate linearly
+    from 0 to given `learning_rate`, after this warmup period learning rate
+    would be decreased following the values of the cosine function. If
+    `with_hard_restarts` is True, the cosine function could have serveral hard
+    restarts.
+
+    Args:
+        learning_rate (float): The base learning rate. It is a python float
+            number.
+        total_steps (int): The number of training steps.
+        warmup (int|float): If int, it means the number of steps for warmup.
+            If float, it means the proportion of warmup in total training steps.
+        with_hard_restarts (bool) Whether cosine function has several hard
+            restarts. Default: False.
+        num_cycles (int|float optional): If `with_hard_restarts` is False, it
+            means the number of waves in cosine scheduler and should be an
+            integer number and defaults to 1. If `with_hard_restarts` is True,
+            it means the number of hard restarts to use and should be a float
+            number and defaults to be 0.5. Default: None.
+        last_epoch (int, optional): The index of last epoch. It can be set to
+            restart training. If None, it means initial learning rate. 
+            Default: -1.
+
+    Examples:
+        
+        .. code-block:: python
+
+            from paddlenlp.transformers import CosineDecayWithWarmup
+            lr, warmup_steps, max_steps = 0.1, 100, 1000
+            lr_scheduler = CosineDecayWithWarmup(lr, max_steps, warmup_steps)
+
+    """
+
+    def __init__(self,
+                 learning_rate,
+                 total_steps,
+                 warmup,
+                 with_hard_restarts=False,
+                 num_cycles=None,
+                 last_epoch=-1,
+                 verbose=False):
+        warmup_steps = warmup if is_integer(warmup) else int(
+            math.floor(warmup * total_steps))
+        # Input check
+        if num_cycles is not None:
+            assert not with_hard_restarts and isinstance(num_cycles, int) or with_hard_restarts and isinstance(num_cycles, float), \
+            "`num_circles` should be an integer while `with_hard_restarts` is False, an float while `with_hard_restarts` is True."
+        else:
+            num_cycles = 1 if not with_hard_restarts else 0.5
+
+        def lr_lambda(current_step):
+            if current_step < warmup_steps:
+                return float(current_step) / float(max(1, warmup_steps))
+
+            progress = float(current_step - warmup_steps) / float(
+                max(1, total_steps - warmup_steps))
+
+            if with_hard_restarts:
+                if progress >= 1.0:
+                    return 0.0
+                return max(0.0, 0.5 * (1.0 + math.cos(math.pi * (
+                    (float(num_cycles) * progress) % 1.0))))
+
+            return max(0.0, 0.5 * (
+                1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
+
+        super(CosineDecayWithWarmup, self).__init__(learning_rate, lr_lambda,
+                                                    last_epoch, verbose)
+
+
+class PolyDecayWithWarmup(LambdaDecay):
+    """
+    Create a learning rate scheduler, which increases learning rate linearly
+    from 0 to given `lr_init`, after this warmup period learning rate would
+    be decreased as a polynomial decay from the base learning rate to the end
+    learning rate `lr_end`. 
+
+    Args:
+        learning_rate (float): The base learning rate. It is a python float
+            number.
+        total_steps (int): The number of training steps.
+        warmup (int|float): If int, it means the number of steps for warmup.
+            If float, it means the proportion of warmup in total training steps.
+        lr_end (float, optional): The end learning rate. Default: 1e-7.
+        power (float, optional): Power factor. Default: 1.0.
+        last_epoch (int, optional): The index of last epoch. It can be set to
+            restart training. If None, it means initial learning rate.
+            Default: -1.
+
+    Examples:
+        
+        .. code-block:: python
+
+            from paddlenlp.transformers import PolyDecayWithWarmup
+            lr, lr_end, warmup_steps, max_steps = 0.1, 1e-6, 100, 1000
+            lr_scheduler = PolyDecayWithWarmup(lr, max_steps, warmup_steps, lr_end)
+
+    """
+
+    def __init__(self,
+                 learning_rate,
+                 total_steps,
+                 warmup,
+                 lr_end=1e-7,
+                 power=1.0,
+                 last_epoch=-1,
+                 verbose=False):
+        lr_init = learning_rate
+        assert lr_init > lr_end, f"`lr_end` must be be smaller than `learning_rate`. But `lr_end` is {lr_end} while `learning_rate` is {lr_init}."
+        warmup_steps = warmup if is_integer(warmup) else int(
+            math.floor(warmup * total_steps))
+
+        def lr_lambda(current_step):
+            if current_step < warmup_steps:
+                return float(current_step) / float(max(1, warmup_steps))
+            elif current_step > total_steps:
+                return lr_end / lr_init  # it multiplies by lr_init equals to lr_end
+            else:
+                lr_range = lr_init - lr_end
+                decay_steps = total_steps - warmup_steps
+                pct_remaining = 1 - (current_step - warmup_steps) / decay_steps
+                decay = lr_range * pct_remaining**power + lr_end
+                return decay / lr_init  # it multiplies by lr_init equals to decay
+
+        super(PolyDecayWithWarmup, self).__init__(lr_init, lr_lambda,
+                                                  last_epoch, verbose)