Merge branch 'develop' into deploy

Author: LemonNoel

examples/model_interpretation/README.md

@@ -198,12 +198,17 @@ pip3 install paddlepaddle-gpu
     sample_type： 数据的类性，分为原始数据（ori）和扰动数据（disturb）；
     rel_ids：与原始数据关联的扰动数据的id列表（只有原始数据有）；
 ## 模型运行
-### 证据抽取运行：
+### 模型预测：
 
       model_interpretation/task/{task}/run_inter_all.sh (生成所有结果)
       model_interpretation/task/{task}/run_inter.sh (生成单个配置的结果，配置可以选择不同的评估模型，以及不同的证据抽取方法、语言)
 
 (注：{task}可取值为["senti","similarity","mrc"]，其中senti代表情感分析，similarity代表相似度计算，mrc代表阅读理解)
+
+### 证据抽取:
+      cd model_interpretation/rationale_extraction
+      ./generate.sh
+
 ### 可解释评估：
 #### 合理性（plausibility）：
     model_interpretation/evaluation/plausibility/run_f1.sh

examples/model_interpretation/rationale_extraction/run_2_pred_senti_per.sh

@@ -26,8 +26,7 @@ do
             #CKPT=../../../${TASK}/pretrained_models/saved_model_en/roberta_large_20211207_174631/model_4000/model_state.pdparams
         elif [[ $BASE_MODEL == "lstm" ]]; then
             VOCAB_PATH=../task/${TASK}/rnn/vocab.sst2_train
-            CKPT=../task/${TASK}/rnn/checkpoints_en_ori/final.pdparams
-            #CKPT=../../../${TASK}/rnn/checkpoints_en/final.pdparams
+            CKPT=../task/${TASK}/rnn/checkpoints_en/final.pdparams
         fi
 
     elif [[ $LANGUAGE == "ch" ]]; then
@@ -42,8 +41,7 @@ do
             #CKPT=../../../${TASK}/pretrained_models/saved_model_ch/roberta_large_20211207_143351/model_900/model_state.pdparams
         elif [[ $BASE_MODEL == "lstm" ]]; then
             VOCAB_PATH=../task/${TASK}/rnn
-            CKPT=../task/${TASK}/rnn/checkpoints_ch_ori/final.pdparams
-            #CKPT=../../../${TASK}/rnn/checkpoints_ch/final.pdparams
+            CKPT=../task/${TASK}/rnn/checkpoints_ch/final.pdparams
         fi
     fi
 

examples/model_interpretation/task/mrc/run_train_rc.sh

@@ -6,7 +6,10 @@ export CUDA_VISIBLE_DEVICES=7
 export PYTHONPATH=.:$PYTHONPATH
 
 LANGUAGE=ch                    # LANGUAGE choose in [ch, en]
-BASE_MODEL=roberta_large       # chooices [roberta_base, roberta_large]
+BASE_MODEL=roberta_base       # chooices [roberta_base, roberta_large]
+
+[ -d "logs" ] || mkdir -p "logs"
+set -x
 
 if [[ $LANGUAGE == "ch" ]]; then
     if [[ $BASE_MODEL == "roberta_base" ]]; then

examples/model_interpretation/task/mrc/saliency_map/rc_finetune.py

@@ -298,9 +298,10 @@ def map_fn_DuCheckList_finetune(examples):
                     if step % 1000 == 0:
                         if args.save_dir is not None:
                             paddle.save(model.state_dict(),
-                                        args.save_dir / 'ckpt.bin')
+                                        os.path.join(args.save_dir, 'ckpt.bin'))
                             log.debug('save model!')
 
                 if args.save_dir is not None:
-                    paddle.save(model.state_dict(), args.save_dir / 'ckpt.bin')
+                    paddle.save(model.state_dict(),
+                                os.path.join(args.save_dir, 'ckpt.bin'))
                     log.debug('save model!')

examples/model_interpretation/task/senti/pretrained_models/run_train.sh

@@ -16,6 +16,9 @@ elif [[ $LANGUAGE == "en" ]]; then
     MAX_SEQ_LENGTH=512
 fi
 
+[ -d "logs" ] || mkdir -p "logs"
+set -x
+
 python3 ./train.py  \
     --learning_rate ${LEARNING_RATE} \
     --max_seq_length ${MAX_SEQ_LENGTH} \

examples/model_interpretation/task/senti/pretrained_models/train.py

@@ -31,7 +31,9 @@
 from paddlenlp.transformers import LinearDecayWithWarmup
 from paddlenlp.transformers.roberta.tokenizer import RobertaTokenizer, RobertaBPETokenizer
 sys.path.append('..')
+sys.path.append('../../..')
 from roberta.modeling import RobertaForSequenceClassification
+sys.path.remove('../../..')
 sys.path.remove('..')
 from utils import convert_example
 

examples/model_interpretation/task/senti/rnn/lstm_train.sh

@@ -6,7 +6,7 @@ unset CUDA_VISIBLE_DEVICES
 LANGUAGE=en
 
 if [[ $LANGUAGE == 'ch' ]]; then
-    VOCAB_PATH='./vocab.txt'
+    VOCAB_PATH='./'
 else
     VOCAB_PATH='vocab.sst2_train'
 fi

examples/model_interpretation/task/senti/rnn/train.py

@@ -25,7 +25,7 @@
 
 from model import LSTMModel, SelfInteractiveAttention, BiLSTMAttentionModel
 from utils import convert_example, CharTokenizer
-from paddlenlp.transformers.ernie.tokenizer import ErnieTokenizer
+from ernie.tokenizing_ernie import ErnieTokenizer
 
 # yapf: disable
 parser = argparse.ArgumentParser(__doc__)

examples/model_interpretation/task/senti/run_inter.sh

@@ -27,8 +27,7 @@ if [[ $LANGUAGE == "en" ]]; then
         #CKPT=pretrained_models/saved_model_en/roberta_large_20211207_174631/model_4000/model_state.pdparams
     elif [[ $BASE_MODEL == "lstm" ]]; then
         VOCAB_PATH='rnn/vocab.sst2_train'
-        CKPT=rnn/checkpoints_en_ori/final.pdparams
-        #CKPT=rnn/checkpoints_en/final.pdparams
+        CKPT=rnn/checkpoints_en/final.pdparams
     fi
 
 elif [[ $LANGUAGE == "ch" ]]; then
@@ -43,8 +42,7 @@ elif [[ $LANGUAGE == "ch" ]]; then
         #CKPT=pretrained_models/saved_model_ch/roberta_large_20211229_105019/model_900/model_state.pdparams
     elif [[ $BASE_MODEL == "lstm" ]]; then
         VOCAB_PATH='rnn'
-        CKPT=rnn/checkpoints_ch_ori/final.pdparams
-        #CKPT=rnn/checkpoints_ch/final.pdparams
+        CKPT=rnn/checkpoints_ch/final.pdparams
     fi
 fi
 

examples/model_interpretation/task/senti/run_inter_all.sh

@@ -30,7 +30,7 @@ do
                     #CKPT=pretrained_models/saved_model_en/roberta_large_20211207_174631/model_4000/model_state.pdparams
                 elif [[ $BASE_MODEL == "lstm" ]]; then
                     VOCAB_PATH='rnn/vocab.sst2_train'
-                    CKPT=rnn/checkpoints_en_ori/final.pdparams
+                    CKPT=rnn/checkpoints_en/final.pdparams
                     #CKPT=rnn/checkpoints_en/final.pdparams
                 fi
 
@@ -46,7 +46,7 @@ do
                     #CKPT=pretrained_models/saved_model_ch/roberta_large_20211229_105019/model_900/model_state.pdparams
                 elif [[ $BASE_MODEL == "lstm" ]]; then
                     VOCAB_PATH='rnn'
-                    CKPT=rnn/checkpoints_ch_ori/final.pdparams
+                    CKPT=rnn/checkpoints_ch/final.pdparams
                     #CKPT=rnn/checkpoints_ch/final.pdparams
                 fi
             fi