Patch (#499)

* upgrade to the new lac interface

* remove jieba

* + init_checkppoint for seq_cls demo

* backto jieba

demo/finetune_classifier.py

@@ -25,7 +25,6 @@
 import numpy as np
 import multiprocessing
 import tempfile
-import jieba
 import re
 
 import paddle

demo/finetune_classifier_dygraph.py

@@ -58,6 +58,7 @@
     parser.add_argument('--save_dir', type=str, default=None, help='model output directory')
     parser.add_argument('--max_steps', type=int, default=None, help='max_train_steps, set this to EPOCH * NUM_SAMPLES / BATCH_SIZE')
     parser.add_argument('--wd', type=float, default=0.01, help='weight decay, aka L2 regularizer')
+    parser.add_argument('--init_checkpoint', type=str, default=None, help='checkpoint to warm start from')
 
 
     args = parser.parse_args()
@@ -103,6 +104,11 @@ def map_fn(seg_a, seg_b, label):
     with FD.guard(place):
         model = ErnieModelForSequenceClassification.from_pretrained(args.from_pretrained, num_labels=3, name='')
 
+        if args.init_checkpoint is not None:
+            log.info('loading checkpoint from %s' % args.init_checkpoint)
+            sd, _ = FD.load_dygraph(args.init_checkpoint)
+            model.set_dict(sd)
+
         g_clip = F.clip.GradientClipByGlobalNorm(1.0) #experimental
         if args.use_lr_decay:
             opt = AdamW(learning_rate=LinearDecay(args.lr, int(args.warmup_proportion * args.max_steps), args.max_steps), parameter_list=model.parameters(), weight_decay=args.wd, grad_clip=g_clip)

demo/pretrain/make_pretrain_data.py

@@ -1,7 +1,6 @@
 import sys
 import argparse
 import struct
-#import jieba
 import random as r
 import re
 import gzip

demo/pretrain/pretrain.py

@@ -44,7 +44,6 @@
 import propeller.paddle as propeller
 from propeller.paddle.data import Dataset
 
-#import jieba
 from propeller import log
 
 log.setLevel(logging.DEBUG)

demo/pretrain/pretrain_dygraph.py

@@ -45,7 +45,6 @@
 import propeller.paddle as propeller
 from propeller.paddle.data import Dataset
 
-#import jieba
 from propeller import log
 
 log.setLevel(logging.DEBUG)

ernie/tokenizing_ernie.py

@@ -222,14 +222,14 @@ def from_pretrained(cls, pretrain_dir_or_url, force_download=False, **kwargs):
     def __init__(self, vocab, sp_model_path, **kwargs):
         super(ErnieTinyTokenizer, self).__init__(vocab, **kwargs)
         import sentencepiece as spm
+        import jieba as jb
         self.sp_model = spm.SentencePieceProcessor()
         self.window_size = 5
         self.sp_model.Load(sp_model_path)
-        from LAC import LAC
-        self.lac = LAC()
+        self.jb = jb
 
     def cut(self, sentence):
-        return self.lac.lexer(sentence)
+        return self.jb.cut(sentence)
 
     def tokenize(self, text):
         if len(text) == 0:

requirements.txt

@@ -3,4 +3,4 @@ pyzmq==18.0.2
 six==1.11.0
 sklearn==0.0
 sentencepiece==0.1.8
-LAC
+jieba==0.39