Fix ernie_gen* bug.

modules/text/text_generation/ernie_gen/module.py

@@ -43,7 +43,7 @@ from .model import StackModel
     type="nlp/text_generation",
 )
 class ErnieGen(hub.Module):
-    def _initialize(self):
+    def __init__(self):
         """
         initialize with the necessary elements
         """
@@ -109,6 +109,7 @@ class ErnieGen(hub.Module):
                     last_ppl(float): last model ppl.
                 }
         """
+        paddle.disable_static()
         paddle.set_device('gpu') if use_gpu else paddle.set_device('cpu')
 
         if init_ckpt_path is not None:
@@ -118,7 +119,8 @@ class ErnieGen(hub.Module):
 
         train_dataset = self._load_dataset(train_path)
         attn_id = self.tokenizer.vocab['[MASK]']
-        trans_func = convert_example(tokenizer=self.tokenizer,
+        trans_func = convert_example(
+            tokenizer=self.tokenizer,
             attn_id=attn_id,
             tgt_type_id=1,
             max_encode_len=max_encode_len,
@@ -137,7 +139,8 @@ class ErnieGen(hub.Module):
             Pad(axis=0, pad_val=self.tokenizer.pad_token_id),  # attn_ids
             Pad(axis=0, pad_val=self.tokenizer.pad_token_id),  # tgt_labels
         ): after_padding(fn(samples))
-        train_data_loader = DataLoader(dataset=train_dataset,
+        train_data_loader = DataLoader(
+            dataset=train_dataset,
             batch_sampler=train_batch_sampler,
             collate_fn=batchify_fn,
             num_workers=0,
@@ -146,11 +149,8 @@ class ErnieGen(hub.Module):
         if dev_path:
             dev_dataset = self._load_dataset(dev_path)
             dev_dataset = dev_dataset.map(trans_func)
-            dev_data_loader = DataLoader(dataset=dev_dataset,
-                                         batch_size=batch_size,
-                                         collate_fn=batchify_fn,
-                                         num_workers=0,
-                                         return_list=True)
+            dev_data_loader = DataLoader(
+                dataset=dev_dataset, batch_size=batch_size, collate_fn=batchify_fn, num_workers=0, return_list=True)
 
         label_num = self.model.word_emb.weight.shape[0]
         train_model = StackModel(self.model)
@@ -158,7 +158,8 @@ class ErnieGen(hub.Module):
         # Generate parameter names needed to perform weight decay.
         # All bias and LayerNorm parameters are excluded.
         decay_params = [p.name for n, p in self.model.named_parameters() if not any(nd in n for nd in ["bias", "norm"])]
-        optimizer = paddle.optimizer.AdamW(learning_rate=lr_scheduler,
+        optimizer = paddle.optimizer.AdamW(
+            learning_rate=lr_scheduler,
             parameters=self.model.parameters(),
             weight_decay=weight_decay,
             grad_clip=nn.ClipGradByGlobalNorm(1.0),
@@ -174,8 +175,8 @@ class ErnieGen(hub.Module):
                 (src_ids, src_tids, src_pids, tgt_ids, tgt_tids, tgt_pids, attn_ids, mask_src_2_src, mask_tgt_2_srctgt,
                  mask_attn_2_srctgtattn, tgt_labels, _) = batch
                 if label_smooth > 0.:
-                    tgt_labels = nn.functional.label_smooth(nn.functional.one_hot(tgt_labels, label_num),
-                                                            epsilon=label_smooth)
+                    tgt_labels = nn.functional.label_smooth(
+                        nn.functional.one_hot(tgt_labels, label_num), epsilon=label_smooth)
 
                 tgt_pos = paddle.nonzero(attn_ids == attn_id)
                 loss = train_model(src_ids, src_tids, src_pids, tgt_ids, tgt_tids, tgt_pids, attn_ids, mask_src_2_src,
@@ -189,8 +190,8 @@ class ErnieGen(hub.Module):
                 if global_step % log_interval == 0 and paddle.distributed.get_rank() == 0:
                     loss_np = loss.numpy()
                     ppl = np.exp(loss_np)
-                    logger.info('[step %d / %d]train loss %.5f, ppl %.5f, elr %.3e' %
-                                (global_step, max_steps, loss_np, ppl, lr_scheduler.get_lr()))
+                    logger.info('[step %d / %d]train loss %.5f, ppl %.5f, elr %.3e' % (global_step, max_steps, loss_np,
+                                                                                       ppl, lr_scheduler.get_lr()))
                 if save_dir and global_step % save_interval == 0 and global_step > 0:
                     loss_np = loss.numpy()
                     ppl = np.exp(loss_np)
@@ -213,8 +214,8 @@ class ErnieGen(hub.Module):
         if global_step % save_interval != 0:
             loss_np = loss.numpy()
             ppl = np.exp(loss_np)
-            logger.info('[final step %d]train loss %.5f, ppl %.5f, elr %.3e' %
-                        (global_step, loss_np, ppl, lr_scheduler.get_lr()))
+            logger.info('[final step %d]train loss %.5f, ppl %.5f, elr %.3e' % (global_step, loss_np, ppl,
+                                                                                lr_scheduler.get_lr()))
             if save_dir:
                 save_name = "step_%s_ppl_%.5f.pdparams" % (global_step, ppl)
                 save_path = os.path.join(save_dir, save_name)
@@ -304,7 +305,8 @@ class ErnieGen(hub.Module):
         for data in data_loader:
             (src_ids, src_tids, src_pids, _, _, _, _, _, _, _, _, raw_tgt_labels) = data  # never use target when infer
             # Use greedy_search_infilling or beam_search_infilling to get predictions
-            output_ids = beam_search_infilling(model,
+            output_ids = beam_search_infilling(
+                model,
                 src_ids,
                 src_tids,
                 eos_id=eos_id,
@@ -359,7 +361,8 @@ class ErnieGen(hub.Module):
 
 if __name__ == "__main__":
     module = ErnieGen()
-    result = module.finetune(train_path='test_data/train.txt',
+    result = module.finetune(
+        train_path='test_data/train.txt',
         dev_path='test_data/dev.txt',
         max_steps=30,
         batch_size=2,

modules/text/text_generation/ernie_gen_acrostic_poetry/module.py

@@ -39,7 +39,7 @@ from ernie_gen_acrostic_poetry.decode import beam_search_infilling
     type="nlp/text_generation",
 )
 class ErnieGen(hub.NLPPredictionModule):
-    def _initialize(self, line=4, word=7):
+    def __init__(self, line=4, word=7):
         """
         initialize with the necessary elements
         """
@@ -73,14 +73,16 @@ class ErnieGen(hub.NLPPredictionModule):
         Returns:
              results(list): the poetry continuations.
         """
+        paddle.disable_static()
+
         if texts and isinstance(texts, list) and all(texts) and all([isinstance(text, str) for text in texts]):
             predicted_data = texts
         else:
             raise ValueError("The input texts should be a list with nonempty string elements.")
         for i, text in enumerate(texts):
             if len(text) > self.line:
-                logger.warning('The input text: %s, contains more than %i characters, which will be cut off' %
-                               (text, self.line))
+                logger.warning(
+                    'The input text: %s, contains more than %i characters, which will be cut off' % (text, self.line))
                 texts[i] = text[:self.line]
 
             for char in text:
@@ -104,7 +106,8 @@ class ErnieGen(hub.NLPPredictionModule):
             encode_text = self.tokenizer.encode(text)
             src_ids = paddle.to_tensor(encode_text['input_ids']).unsqueeze(0)
             src_sids = paddle.to_tensor(encode_text['token_type_ids']).unsqueeze(0)
-            output_ids = beam_search_infilling(self.model,
+            output_ids = beam_search_infilling(
+                self.model,
                 src_ids,
                 src_sids,
                 eos_id=self.tokenizer.vocab['[SEP]'],
@@ -130,10 +133,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Add the command config options
         """
-        self.arg_config_group.add_argument('--use_gpu',
-                                           type=ast.literal_eval,
-                                           default=False,
-                                           help="whether use GPU for prediction")
+        self.arg_config_group.add_argument(
+            '--use_gpu', type=ast.literal_eval, default=False, help="whether use GPU for prediction")
 
         self.arg_config_group.add_argument('--beam_width', type=int, default=5, help="the beam search width")
 
@@ -142,7 +143,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Run as a command
         """
-        self.parser = argparse.ArgumentParser(description='Run the %s module.' % self.name,
+        self.parser = argparse.ArgumentParser(
+            description='Run the %s module.' % self.name,
             prog='hub run %s' % self.name,
             usage='%(prog)s',
             add_help=True)

modules/text/text_generation/ernie_gen_couplet/module.py

@@ -39,7 +39,7 @@ from ernie_gen_couplet.decode import beam_search_infilling
     type="nlp/text_generation",
 )
 class ErnieGen(hub.NLPPredictionModule):
-    def _initialize(self):
+    def __init__(self):
         """
         initialize with the necessary elements
         """
@@ -67,6 +67,8 @@ class ErnieGen(hub.NLPPredictionModule):
         Returns:
              results(list): the right rolls.
         """
+        paddle.disable_static()
+
         if texts and isinstance(texts, list) and all(texts) and all([isinstance(text, str) for text in texts]):
             predicted_data = texts
         else:
@@ -93,7 +95,8 @@ class ErnieGen(hub.NLPPredictionModule):
             encode_text = self.tokenizer.encode(text)
             src_ids = paddle.to_tensor(encode_text['input_ids']).unsqueeze(0)
             src_sids = paddle.to_tensor(encode_text['token_type_ids']).unsqueeze(0)
-            output_ids = beam_search_infilling(self.model,
+            output_ids = beam_search_infilling(
+                self.model,
                 src_ids,
                 src_sids,
                 eos_id=self.tokenizer.vocab['[SEP]'],
@@ -119,10 +122,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Add the command config options
         """
-        self.arg_config_group.add_argument('--use_gpu',
-                                           type=ast.literal_eval,
-                                           default=False,
-                                           help="whether use GPU for prediction")
+        self.arg_config_group.add_argument(
+            '--use_gpu', type=ast.literal_eval, default=False, help="whether use GPU for prediction")
 
         self.arg_config_group.add_argument('--beam_width', type=int, default=5, help="the beam search width")
 
@@ -131,7 +132,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Run as a command
         """
-        self.parser = argparse.ArgumentParser(description='Run the %s module.' % self.name,
+        self.parser = argparse.ArgumentParser(
+            description='Run the %s module.' % self.name,
             prog='hub run %s' % self.name,
             usage='%(prog)s',
             add_help=True)

modules/text/text_generation/ernie_gen_lover_words/module.py

@@ -39,7 +39,7 @@ from ernie_gen_lover_words.decode import beam_search_infilling
     type="nlp/text_generation",
 )
 class ErnieGen(hub.NLPPredictionModule):
-    def _initialize(self):
+    def __init__(self):
         """
         initialize with the necessary elements
         """
@@ -67,6 +67,8 @@ class ErnieGen(hub.NLPPredictionModule):
         Returns:
              results(list): the poetry continuations.
         """
+        paddle.disable_static()
+
         if texts and isinstance(texts, list) and all(texts) and all([isinstance(text, str) for text in texts]):
             predicted_data = texts
         else:
@@ -85,7 +87,8 @@ class ErnieGen(hub.NLPPredictionModule):
             encode_text = self.tokenizer.encode(text)
             src_ids = paddle.to_tensor(encode_text['input_ids']).unsqueeze(0)
             src_sids = paddle.to_tensor(encode_text['token_type_ids']).unsqueeze(0)
-            output_ids = beam_search_infilling(self.model,
+            output_ids = beam_search_infilling(
+                self.model,
                 src_ids,
                 src_sids,
                 eos_id=self.tokenizer.vocab['[SEP]'],
@@ -111,10 +114,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Add the command config options
         """
-        self.arg_config_group.add_argument('--use_gpu',
-                                           type=ast.literal_eval,
-                                           default=False,
-                                           help="whether use GPU for prediction")
+        self.arg_config_group.add_argument(
+            '--use_gpu', type=ast.literal_eval, default=False, help="whether use GPU for prediction")
 
         self.arg_config_group.add_argument('--beam_width', type=int, default=5, help="the beam search width")
 
@@ -123,7 +124,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Run as a command
         """
-        self.parser = argparse.ArgumentParser(description='Run the %s module.' % self.name,
+        self.parser = argparse.ArgumentParser(
+            description='Run the %s module.' % self.name,
             prog='hub run %s' % self.name,
             usage='%(prog)s',
             add_help=True)

modules/text/text_generation/ernie_gen_poetry/module.py

@@ -39,7 +39,7 @@ from ernie_gen_poetry.decode import beam_search_infilling
     type="nlp/text_generation",
 )
 class ErnieGen(hub.NLPPredictionModule):
-    def _initialize(self):
+    def __init__(self):
         """
         initialize with the necessary elements
         """
@@ -67,6 +67,8 @@ class ErnieGen(hub.NLPPredictionModule):
         Returns:
              results(list): the poetry continuations.
         """
+        paddle.disable_static()
+
         if texts and isinstance(texts, list) and all(texts) and all([isinstance(text, str) for text in texts]):
             predicted_data = texts
         else:
@@ -102,7 +104,8 @@ class ErnieGen(hub.NLPPredictionModule):
             encode_text = self.tokenizer.encode(text)
             src_ids = paddle.to_tensor(encode_text['input_ids']).unsqueeze(0)
             src_sids = paddle.to_tensor(encode_text['token_type_ids']).unsqueeze(0)
-            output_ids = beam_search_infilling(self.model,
+            output_ids = beam_search_infilling(
+                self.model,
                 src_ids,
                 src_sids,
                 eos_id=self.tokenizer.vocab['[SEP]'],
@@ -128,10 +131,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Add the command config options
         """
-        self.arg_config_group.add_argument('--use_gpu',
-                                           type=ast.literal_eval,
-                                           default=False,
-                                           help="whether use GPU for prediction")
+        self.arg_config_group.add_argument(
+            '--use_gpu', type=ast.literal_eval, default=False, help="whether use GPU for prediction")
 
         self.arg_config_group.add_argument('--beam_width', type=int, default=5, help="the beam search width")
 
@@ -140,7 +141,8 @@ class ErnieGen(hub.NLPPredictionModule):
         """
         Run as a command
         """
-        self.parser = argparse.ArgumentParser(description='Run the %s module.' % self.name,
+        self.parser = argparse.ArgumentParser(
+            description='Run the %s module.' % self.name,
             prog='hub run %s' % self.name,
             usage='%(prog)s',
             add_help=True)