Upgrade faster ernie by using fused transformers (#1308)

* add faster-ernie-1.0 model params

* add amp for faster ernie seq cls task

Co-authored-by: Zeyu Chen <chenzeyu01@baidu.com>

examples/experimental/faster_ernie/seq_cls/train.py

@@ -17,6 +17,7 @@
 import os
 import random
 import time
+import distutils.util
 
 import numpy as np
 import paddle
@@ -39,6 +40,8 @@
 parser.add_argument("--init_from_ckpt", type=str, default=None, help="The path of checkpoint to be loaded.")
 parser.add_argument("--seed", type=int, default=1000, help="random seed for initialization")
 parser.add_argument('--device', choices=['cpu', 'gpu', 'xpu'], default="gpu", help="Select which device to train model, defaults to gpu.")
+parser.add_argument("--use_amp", type=distutils.util.strtobool, default=False, help="Enable mixed precision training.")
+parser.add_argument("--scale_loss", type=float, default=2**15, help="The value of scale_loss for fp16.")
 args = parser.parse_args()
 # yapf: enable
 
@@ -121,15 +124,20 @@ def do_train():
 
     criterion = paddle.nn.loss.CrossEntropyLoss()
     metric = paddle.metric.Accuracy()
+    if args.use_amp:
+        scaler = paddle.amp.GradScaler(init_loss_scaling=args.scale_loss)
 
     global_step = 0
     tic_train = time.time()
     for epoch in range(1, args.epochs + 1):
         for step, batch in enumerate(train_data_loader, start=1):
             texts, labels = batch["text"], batch["label"]
             texts = to_tensor(texts)
-            logits, predictions = model(texts)
-            loss = criterion(logits, labels)
+            with paddle.amp.auto_cast(
+                    args.use_amp,
+                    custom_white_list=["fused_feedforward", "fused_attention"]):
+                logits, predictions = model(texts)
+                loss = criterion(logits, labels)
             probs = F.softmax(logits, axis=1)
             correct = metric.compute(logits, labels)
             metric.update(correct)
@@ -142,8 +150,12 @@ def do_train():
                     % (global_step, epoch, step, loss, acc,
                        10 / (time.time() - tic_train)))
                 tic_train = time.time()
-            loss.backward()
-            optimizer.step()
+            if args.use_amp:
+                scaler.scale(loss).backward()
+                scaler.minimize(optimizer, loss)
+            else:
+                loss.backward()
+                optimizer.step()
             lr_scheduler.step()
             optimizer.clear_grad()
             if global_step % 100 == 0:

paddlenlp/experimental/ernie_model.py

@@ -17,6 +17,7 @@
 import paddle
 import paddle.fluid.core as core
 import paddle.nn as nn
+from paddle.incubate.nn import FusedTransformerEncoderLayer
 
 from paddlenlp.experimental import FasterTokenizer, FasterPretrainedModel
 from paddlenlp.transformers.model_utils import register_base_model
@@ -105,7 +106,7 @@ class FasterErniePretrainedModel(FasterPretrainedModel):
     pretrained_resource_files_map = {
         "model_state": {
             "ernie-1.0":
-            "https://paddlenlp.bj.bcebos.com/models/transformers/ernie/ernie_v1_chn_base.pdparams",
+            "https://paddlenlp.bj.bcebos.com/models/transformers/faster_ernie/faster_ernie_v1_chn_base.pdparams",
             "ernie-tiny":
             "https://paddlenlp.bj.bcebos.com/models/transformers/ernie_tiny/ernie_tiny.pdparams",
             "ernie-2.0-en":
@@ -245,14 +246,14 @@ def __init__(
         self.embeddings = ErnieEmbeddings(
             vocab_size, hidden_size, hidden_dropout_prob,
             max_position_embeddings, type_vocab_size, pad_token_id, weight_attr)
-        encoder_layer = nn.TransformerEncoderLayer(
+        encoder_layer = FusedTransformerEncoderLayer(
             hidden_size,
             num_attention_heads,
             intermediate_size,
-            dropout=hidden_dropout_prob,
+            dropout_rate=hidden_dropout_prob,
             activation=hidden_act,
-            attn_dropout=attention_probs_dropout_prob,
-            act_dropout=0,
+            attn_dropout_rate=attention_probs_dropout_prob,
+            act_dropout_rate=0,
             weight_attr=weight_attr, )
         self.encoder = nn.TransformerEncoder(encoder_layer, num_hidden_layers)
         self.pooler = ErniePooler(hidden_size, weight_attr)