modify ci accroding to PaddlePaddle#4398 and adjust predictor

Author: Yam0214

model_zoo/ernie-3.0/deploy/simple_serving/README.md

@@ -6,9 +6,7 @@
 - [其他参数设置](#其他参数设置)
 
 ## 环境准备
-
-paddlenlp >= 2.5.0
-
+使用有SimpleServing功能的PaddleNLP版本
 
 ## Server服务启动
 ### 文本分类任务启动

model_zoo/ernie-m/configs/test.yaml

@@ -28,8 +28,8 @@ pip install -r requirements_gpu.txt
 
 ### 2.2 CPU端推理样例
 在 CPU 端，请使用如下命令进行部署
-```
-python infer_cpu.py --task_name seq_cls --model_path ../../finetuned_models/export/model
+```sh
+python inference.py --device cpu --task_name seq_cls --model_path ../../finetuned_models/export/model
 ```
 输出打印如下:
 ```
@@ -58,8 +58,8 @@ infer_cpu.py 脚本中的参数说明：
 
 ### 2.3 GPU端推理样例
 在 GPU 端，请使用如下命令进行部署
-```
-python infer_gpu.py --task_name seq_cls --model_path ../../finetuned_models/export/model
+```sh
+python inference.py --device gpu --task_name seq_cls --model_path ../../finetuned_models/export/model
 ```
 输出打印如下:
 ```
@@ -77,8 +77,8 @@ label: neutral   confidence: 0.9880901575088501
 -----------------------------
 ```
 如果需要 FP16 进行加速，可以设置 precision_mode 为 fp16，具体命令为
-```
-python infer_gpu.py --task_name seq_cls --model_path ../../finetuned_models/export/model --precision_mode fp16
+```sh
+python inference.py --device gpu --task_name seq_cls --model_path ../../finetuned_models/export/model --precision_mode fp16
 ```
 infer_gpu.py 脚本中的参数说明：
 | 参数 |参数说明 |

model_zoo/ernie-m/deploy/predictor/README.md

@@ -14,12 +14,17 @@
 
 import argparse
 
-from ernie_m_predictor import ErnieMPredictor
+from psutil import cpu_count
+
+from .ernie_m_predictor import ErnieMPredictor
 
 
 def parse_args():
     parser = argparse.ArgumentParser()
     # Required parameters
+    parser.add_argument(
+        "--device", default="cpu", type=str, choices=["gpu", "cpu"], help="Device selected for inference."
+    )
     parser.add_argument(
         "--task_name",
         default="seq_cls",
@@ -69,14 +74,19 @@ def parse_args():
         choices=["fp32", "fp16", "int8"],
         help="Inference precision.",
     )
+    parser.add_argument(
+        "--num_threads",
+        default=cpu_count(logical=False),
+        type=int,
+        help="num_threads for cpu.",
+    )
     args = parser.parse_args()
     return args
 
 
 def main():
     args = parse_args()
     args.task_name = args.task_name.lower()
-    args.device = "gpu"
     predictor = ErnieMPredictor(args)
 
     if args.task_name == "seq_cls":

model_zoo/ernie-m/deploy/predictor/infer_cpu.py

@@ -0,0 +1,51 @@
+classifier:
+  default:
+    do_train: True
+    do_eval: True
+    do_export: True
+    task_type: cross-lingual-transfer
+    model_name_or_path: __internal_testing__/ernie-m
+    output_dir: ./output_dir
+    export_model_dir: ./output_dir
+    per_device_train_batch_size: 16
+    per_device_eval_batch_size: 16
+    max_seq_length: 256
+    learning_rate: 5e-5
+    classifier_dropout: 0.1
+    weight_decay: 0.0
+    layerwise_decay: 0.8
+    warmup_ratio: 0.1
+    max_steps: 2
+    logging_steps: 1
+    save_steps: 1
+    eval_steps: 1
+    device: cpu
+    fp16: False
+    load_best_model_at_end: True
+    metric_for_best_model: eval_accuracy
+    overwrite_output_dir: True
+  
+  slow:
+    do_train: True
+    do_eval: True
+    do_export: True
+    task_type: cross-lingual-transfer
+    model_name_or_path: ernie-m-base
+    output_dir: ./output_dir
+    export_model_dir: ./output_dir
+    per_device_train_batch_size: 16
+    per_device_eval_batch_size: 16
+    max_seq_length: 256
+    learning_rate: 5e-5
+    classifier_dropout: 0.1
+    weight_decay: 0.0
+    layerwise_decay: 0.8
+    num_train_epochs: 5
+    logging_steps: 767
+    save_steps: 12272
+    eval_steps: 767
+    device: gpu
+    fp16: False
+    load_best_model_at_end: Trues
+    metric_for_best_model: eval_accuracy
+    overwrite_output_dir: True