Update seq2seq using api2.0

seq2seq/README.md

@@ -1,4 +1,4 @@
-运行本目录下的范例模型需要安装PaddlePaddle Fluid 1.7版。如果您的 PaddlePaddle 安装版本低于此要求，请按照[安装文档](https://www.paddlepaddle.org.cn/#quick-start)中的说明更新 PaddlePaddle 安装版本。
+运行本目录下的范例模型需要安装PaddlePaddle 2.0版。如果您的 PaddlePaddle 安装版本低于此要求，请按照[安装文档](https://www.paddlepaddle.org.cn/#quick-start)中的说明更新 PaddlePaddle 安装版本。
 
 # Sequence to Sequence (Seq2Seq)
 
@@ -12,8 +12,7 @@
 ├── download.py            # 数据下载程序
 ├── train.py               # 训练主程序
 ├── predict.py             # 预测主程序
-├── seq2seq_attn.py        # 带注意力机制的翻译模型程序
-└── seq2seq_base.py        # 无注意力机制的翻译模型程序
+└── seq2seq_base.py        # 翻译模型程序
 ```
 
 ## 简介
@@ -35,7 +34,7 @@ Sequence to Sequence (Seq2Seq)，使用编码器-解码器（Encoder-Decoder）
 git clone https://github.com/PaddlePaddle/hapi
 cd hapi
 export PYTHONPATH=$PYTHONPATH:`pwd`
-cd examples/seq2seq
+cd seq2seq
 ```
 
 ## 数据介绍
@@ -157,21 +156,19 @@ python predict.py \
 使用 [*multi-bleu.perl*](https://github.com/moses-smt/mosesdecoder.git) 工具来评价模型预测的翻译质量，使用方法如下：
 
 ```sh
-mosesdecoder/scripts/generic/multi-bleu.perl tst2013.vi < infer_output.txt
+perl mosesdecoder/scripts/generic/multi-bleu.perl tst2013.vi < infer_output.txt
 ```
 
 每个模型分别训练了10次，单次取第10个epoch保存的模型进行预测，取beam_size=10。效果如下（为了便于观察，对10次结果按照升序进行了排序）：
 
 ```
 > no attention
 tst2012 BLEU:
-[10.75 10.85 10.9  10.94 10.97 11.01 11.01 11.04 11.13 11.4]
+
 tst2013 BLEU:
-[10.71 10.71 10.74 10.76 10.91 10.94 11.02 11.16 11.21 11.44]
 
 > with attention
 tst2012 BLEU:
-[21.14 22.34 22.54 22.65 22.71 22.71 23.08 23.15 23.3  23.4]
+
 tst2013 BLEU:
-[23.41 24.79 25.11 25.12 25.19 25.24 25.39 25.61 25.61 25.63]
-```
+24.94

seq2seq/args.py

@@ -36,8 +36,8 @@ def parse_args():
     parser.add_argument(
         "--attention",
         type=eval,
-        default=False,
-        help="Whether use attention model")
+        default=True,
+        help="Whether use attention in model")
 
     parser.add_argument(
         "--optimizer",
@@ -56,11 +56,13 @@ def parse_args():
         type=int,
         default=1,
         help="layers number of encoder and decoder")
+
     parser.add_argument(
         "--hidden_size",
         type=int,
         default=100,
         help="hidden size of encoder and decoder")
+
     parser.add_argument("--src_vocab_size", type=int, help="source vocab size")
     parser.add_argument("--tar_vocab_size", type=int, help="target vocab size")
 
@@ -105,6 +107,7 @@ def parse_args():
 
     parser.add_argument(
         "--infer_file", type=str, help="file name for inference")
+
     parser.add_argument(
         "--infer_output_file",
         type=str,
@@ -120,7 +123,7 @@ def parse_args():
         help='Whether using gpu [True|False]')
 
     parser.add_argument(
-        '--eager_run', type=eval, default=False, help='Whether to use dygraph')
+        '--eager_run', type=eval, default=True, help='Whether to use dygraph')
 
     parser.add_argument(
         "--enable_ce",

seq2seq/download.py

@@ -47,7 +47,7 @@ def main(arguments):
         url = remote_path + '/' + filename
         tar_file = os.path.join(tar_path, filename)
         URLLIB.urlretrieve(url, tar_file)
-    print("Downloaded sucess......")
+    print("Downloaded success......")
 
 
 if __name__ == '__main__':

seq2seq/infer.sh

@@ -0,0 +1,19 @@
+export CUDA_VISIBLE_DEVICES="4"
+python predict.py \
+    --attention True \
+    --src_lang en --tar_lang vi \
+    --num_layers 2 \
+    --hidden_size 512 \
+    --src_vocab_size 17191 \
+    --tar_vocab_size 7709 \
+    --batch_size 128 \
+    --dropout 0.2 \
+    --init_scale  0.1 \
+    --max_grad_norm 5.0 \
+    --vocab_prefix data/en-vi/vocab \
+    --infer_file data/en-vi/tst2013.en \
+    --reload_model attention_models/10 \
+    --infer_output_file infer_output.txt \
+    --beam_size 10 \
+    --use_gpu True \
+    --eager_run True

seq2seq/predict.py

@@ -20,14 +20,11 @@
 
 import numpy as np
 import paddle
-import paddle.fluid as fluid
-from paddle.fluid.layers.utils import flatten
-from paddle.fluid.io import DataLoader
+from paddle.io import DataLoader
 from paddle.static import InputSpec as Input
 
 from args import parse_args
-from seq2seq_base import BaseInferModel
-from seq2seq_attn import AttentionInferModel
+from seq2seq import Seq2SeqInfer
 from reader import Seq2SeqDataset, Seq2SeqBatchSampler, SortType, prepare_infer_input
 
 
@@ -50,7 +47,7 @@ def post_process_seq(seq, bos_idx, eos_idx, output_bos=False,
 
 def do_predict(args):
     device = paddle.set_device("gpu" if args.use_gpu else "cpu")
-    fluid.enable_dygraph(device) if args.eager_run else None
+    paddle.enable_static() if not args.eager_run else None
 
     # define model
     inputs = [
@@ -84,14 +81,14 @@ def do_predict(args):
         num_workers=0,
         return_list=True)
 
-    model_maker = AttentionInferModel if args.attention else BaseInferModel
     model = paddle.Model(
-        model_maker(
+        Seq2SeqInfer(
             args.src_vocab_size,
             args.tar_vocab_size,
             args.hidden_size,
             args.hidden_size,
             args.num_layers,
+            args.attention,
             args.dropout,
             bos_id=bos_id,
             eos_id=eos_id,
@@ -109,7 +106,7 @@ def do_predict(args):
     # TODO(guosheng): use model.predict when support variant length
     with io.open(args.infer_output_file, 'w', encoding='utf-8') as f:
         for data in data_loader():
-            finished_seq = model.test_batch(inputs=flatten(data))[0]
+            finished_seq = model.test_batch(inputs=list(data))[0]
             finished_seq = finished_seq[:, :, np.newaxis] if len(
                 finished_seq.shape) == 2 else finished_seq
             finished_seq = np.transpose(finished_seq, [0, 2, 1])

seq2seq/reader.py

@@ -24,9 +24,8 @@
 from functools import partial
 
 import numpy as np
-import paddle.fluid as fluid
-from paddle.fluid.dygraph.parallel import ParallelEnv
-from paddle.fluid.io import BatchSampler, DataLoader, Dataset
+import paddle
+from paddle.io import BatchSampler, DataLoader, Dataset
 
 
 def create_data_loader(args, device, for_train=True):
@@ -68,16 +67,15 @@ def create_data_loader(args, device, for_train=True):
             num_workers=0,
             return_list=True)
         data_loaders[i] = data_loader
-    return data_loaders
+    return data_loaders, eos_id
 
 
 def prepare_train_input(insts, bos_id, eos_id, pad_id):
     src, src_length = pad_batch_data(
         [inst[0] for inst in insts], pad_id=pad_id)
     trg, trg_length = pad_batch_data(
         [inst[1] for inst in insts], pad_id=pad_id)
-    trg_length = trg_length - 1
-    return src, src_length, trg[:, :-1], trg_length, trg[:, 1:, np.newaxis]
+    return src, src_length, trg[:, :-1], trg[:, 1:, np.newaxis]
 
 
 def prepare_infer_input(insts, bos_id, eos_id, pad_id):
@@ -359,9 +357,9 @@ def __init__(self,
         self._random.seed(seed)
         # for multi-devices
         self._distribute_mode = distribute_mode
-        self._nranks = ParallelEnv().nranks
-        self._local_rank = ParallelEnv().local_rank
-        self._device_id = ParallelEnv().dev_id
+        self._nranks = paddle.distributed.ParallelEnv().nranks
+        self._local_rank = paddle.distributed.ParallelEnv().local_rank
+        self._device_id = paddle.distributed.ParallelEnv().dev_id
 
     def __iter__(self):
         # global sort or global shuffle