paddle版本问题

### Describe the bug 使用3.0以下的话，examples中ernie_csc目录下可以按照项目中执行，但是当paddle3.0以后模型导出和预测需修改如下： export_model.py import argparse import os import paddle from paddle.static import InputSpec from paddlenlp.data import Vocab from paddlenlp.transformers import ErnieModel import sys sys.path.append('../..') from pycorrector.ernie_csc.model import ErnieForCSC parser = argparse.ArgumentParser() parser.add_argument("--params_path", type=str, default='./checkpoints/best_model.pdparams', help="Path to trained model parameters.") parser.add_argument("--output_path", type=str, default='./infer_model/static_graph_params', help="Path to save the static graph model (Paddle 3.x uses .json + .pdiparams).") parser.add_argument("--model_name_or_path", type=str, default="ernie-1.0", choices=["ernie-1.0"]) parser.add_argument("--pinyin_vocab_file_path", type=str, default="pinyin_vocab.txt") args = parser.parse_args() def main(): # 加载拼音词表 pinyin_vocab = Vocab.load_vocabulary( args.pinyin_vocab_file_path, unk_token='[UNK]', pad_token='[PAD]') # 加载预训练 ERNIE ernie = ErnieModel.from_pretrained(args.model_name_or_path) # 构建 CSC 模型 model = ErnieForCSC( ernie, pinyin_vocab_size=len(pinyin_vocab), pad_pinyin_id=pinyin_vocab[pinyin_vocab.pad_token] ) # 加载训练参数 model_dict = paddle.load(args.params_path) model.set_dict(model_dict) model.eval() # 转静态图 model = paddle.jit.to_static( model, input_spec=[ InputSpec([None, None], "int64", "input_ids"), InputSpec([None, None], "int64", "pinyin_ids") ], full_graph=True ) # 创建输出目录 output_dir = os.path.dirname(args.output_path) if not os.path.exists(output_dir): os.makedirs(output_dir) # 保存模型 paddle.jit.save(model, args.output_path) print(f"模型已保存为 {args.output_path}.json + {args.output_path}.pdiparams") predict.py # -*- coding: utf-8 -*- import paddle from paddlenlp.data import Vocab, Pad, Stack, Tuple from paddlenlp.transformers import ErnieTokenizer from functools import partial from pycorrector.ernie_csc.utils import convert_example, parse_decode # Paddle 3.x 支持直接用 CPU 或 GPU # paddle.set_device("gpu") # 如果有可用 GPU 且 cuda/cuDNN 环境正确 paddle.set_device("gpu") # 遇到 CUDA/cuDNN 兼容问题可以临时使用 CPU class Predictor: def __init__(self, model_path, tokenizer, pinyin_vocab, max_seq_length=64): self.max_seq_length = max_seq_length # 加载 Paddle 3.x 导出的静态图模型 self.model = paddle.jit.load(model_path) self.model.eval() self.tokenizer = tokenizer self.pinyin_vocab = pinyin_vocab # batchize 函数，自动 pad self.batchify_fn = lambda samples: [ data for data in Tuple( Pad(axis=0, pad_val=self.tokenizer.pad_token_id, dtype='int64'), # input_ids Pad(axis=0, pad_val=self.tokenizer.pad_token_type_id, dtype='int64'), # token_type_ids Pad(axis=0, pad_val=self.pinyin_vocab.token_to_idx[self.pinyin_vocab.pad_token], dtype='int64'), # pinyin_ids Stack(axis=0, dtype='int64') # seq_len )(samples) ] def predict(self, sentences, batch_size=1): examples = [] texts = [] trans_func = partial( convert_example, tokenizer=self.tokenizer, pinyin_vocab=self.pinyin_vocab, max_seq_length=self.max_seq_length, is_test=True ) # 先将文本转换为模型输入 for text in sentences: example = {"source": text.strip()} input_ids, token_type_ids, pinyin_ids, length = trans_func(example) examples.append((input_ids, token_type_ids, pinyin_ids, length)) texts.append(example["source"]) results = [] # 按 batch_size 分批处理 for i in range(0, len(examples), batch_size): batch = examples[i:i + batch_size] token_ids, token_type_ids, pinyin_ids, length = self.batchify_fn(batch) # 转为 paddle Tensor token_ids = paddle.to_tensor(token_ids, dtype="int64") pinyin_ids = paddle.to_tensor(pinyin_ids, dtype="int64") # 推理 corr_logits, det_error_probs = self.model(token_ids, pinyin_ids) # 取 argmax det_pred = det_error_probs.argmax(axis=-1).numpy() char_preds = corr_logits.argmax(axis=-1).numpy() # decode for j in range(len(length)): pred_result = parse_decode( texts[i + j], char_preds[j], det_pred[j], length[j], self.tokenizer, self.max_seq_length ) results.append(''.join(pred_result)) return results if __name__ == "__main__": tokenizer = ErnieTokenizer.from_pretrained("ernie-1.0") pinyin_vocab = Vocab.load_vocabulary( "pinyin_vocab.txt", unk_token='[UNK]', pad_token='[PAD]' ) predictor = Predictor( model_path="infer_model/static_graph_params", # Paddle 3.x 导出路径 tokenizer=tokenizer, pinyin_vocab=pinyin_vocab, max_seq_length=64 ) samples = [ '遇到逆境时，我们必须勇于面对，而且要愈挫愈勇，这样我们才能朝着成功之路前进。', '人生就是如此，经过磨练才能让自己更加坚强，才能使自己更加乐观。', ] results = predictor.predict(samples, batch_size=2) for source, target in zip(samples, results): print("Source:", source) print("Target:", target)