Fix ppminilm export bug (#1596)

* fix ppminilm export bug

* update export script

* update readme

Author: LiuChiachi

examples/model_compression/pp-minilm/README.md

@@ -58,7 +58,9 @@ PP-MiniLM 压缩方案以面向预训练模型的任务无关知识蒸馏(Task-a
 - batch sizes: 16, 32, 64;
 - learning rates: 3e-5, 5e-5, 1e-4
 
-2.量化后比量化前模型参数量多了 0.1M 是因为保存了 scale 值。
+2.量化后比量化前模型参数量多了 0.1M 是因为保存了 scale 值；
+
+3.性能测试的条件是：batch_size: 32, max_seq_len: 128。
 
 **方案流程**
 
@@ -197,7 +199,7 @@ export MODEL_PATH=ppminilm-6l-768h
 export LR=1e-4
 export BS=32
 
-python export_model.py --task_name ${TASK_NAME} --output_dir ${MODEL_PATH}/models/${TASK_NAME}/${LR}_${BS}/
+python export_model.py --task_name ${TASK_NAME} --model_path ${MODEL_PATH}/models/${TASK_NAME}/${LR}_${BS}/
 ```
 
 静态图（部署）模型路径与动态图模型的路径相同，文件名为 `inference.pdmodel` , `inference.pdiparams` 和 `inference.pdiparams.info` 。

examples/model_compression/pp-minilm/finetuning/export_model.py

@@ -34,7 +34,7 @@ def parse_args():
         help="The name of the task to train selected in the list: " +
         ", ".join(METRIC_CLASSES.keys()), )
     parser.add_argument(
-        "--output_dir",
+        "--model_path",
         default="best_clue_model",
         type=str,
         help="The output directory where the model predictions and checkpoints will be written.",
@@ -50,9 +50,10 @@ def parse_args():
 
 
 def do_export(args):
-    save_path = os.path.join(args.output_dir, "inference")
-    model = PPMiniLMForSequenceClassification.from_pretrained(args.output_dir)
+    save_path = os.path.join(os.path.dirname(args.model_path), "inference")
+    model = PPMiniLMForSequenceClassification.from_pretrained(args.model_path)
     is_text_pair = True
+    args.task_name = args.task_name.lower()
     if args.task_name in ('tnews', 'iflytek', 'cluewsc2020'):
         is_text_pair = False
     model.to_static(