fluid.layers.lstm接口参数is_bidirec不能发挥双向的作用

[remarkableJ]

fluid.layers.lstm这个接口的参数is_bidirec貌似不能发挥双向的作用，输出的结果维度跟描述也不一致

[Aurelius84]

请问您可否提供下发现问题的代码？

[3wGTA]

我也遇到同样的问题，如下代码确实没有发挥双向作用

import numpy as np
import paddle.fluid as fluid

weight = np.random.randint(0,2,(32,5))
# weight = np.ones((32,5))
weight[0] = 0
print(weight[:2])

batch_size = 5
def yield_data():
    for i in range(100):
        max_len = 7
        s_x = np.random.randint(2,5)
        s_y = np.random.randint(2,8)
        label = np.random.randint(2)
        x = np.random.randint(0,10,(1,s_x)).tolist()
        y = np.random.randint(0,10,(1,s_y)).tolist()
        yield x[0],y[0] ,label
#         yield np.random.randint(0,10,(1,max_len)),np.random.randint(0,10,(1,max_len)),np.array(label)
#         yield np.random.randint(0,10,(1,s_x))[0],np.random.randint(0,10,(1,s_y))[0],np.array(label)

train_reader = fluid.io.batch(reader=yield_data,batch_size=batch_size)

for i,data in enumerate(train_reader()):
    print(data)
#     print(type(data[0][0][0]))
    break

V = 32
h = 3
emb_size=5
max_len = 7 #最大为5，每个句子最长为5
hidden_size = 2
num_layers = 1

label = fluid.data(name='label', shape=[None, 1], dtype='int64')
x = fluid.data(name='t', shape=[None], dtype='int64',lod_level=1)
y = fluid.data(name='h', shape=[None], dtype='int64',lod_level=1)


w = fluid.ParamAttr(name='emb_vec', initializer=fluid.initializer.NumpyArrayInitializer(weight), trainable=False)

emb_x = fluid.embedding(input=x, size=[32,5], param_attr=w)
emb_y = fluid.embedding(input=y, size=[32,5], param_attr=w)

pad_value = fluid.layers.assign(input=np.array([0.0], dtype=np.float32))
pad_x,info_x = fluid.layers.sequence_pad(emb_x,pad_value)
pad_y,info_y = fluid.layers.sequence_pad(emb_y,pad_value)

init_h = fluid.layers.fill_constant([num_layers, batch_size, hidden_size], 'float32', 0)
init_c = fluid.layers.fill_constant([num_layers, batch_size, hidden_size], 'float32', 0)
# lstm 网络

# 返回的形状是 batch_size, seq_len, hiddensize
lstm_x, x_last_h, x_last_c = fluid.layers.lstm(pad_x, init_h, init_c, max_len, hidden_size, num_layers,is_bidirec=True)


temp_x = lstm_x[:,:1]
# print("lstm_shape:", lstm_x.shape)
# lstm_y, y_last_h, y_last_c = fluid.layers.lstm(emb_y, init_h, init_c, max_len, hidden_size, num_layers)

# # 输出是 senq, batch_size, hiddensize
# lstm_bi_x, bi_x_last_h, bi_x_last_c = fluid.layers.lstm(
#                 input=pad_x, init_h=init_h, init_c=init_c, 
#                 max_len=max_len, hidden_size=hidden_size, num_layers=num_layers, 
#                 dropout_prob=0.2,
#                 is_bidirec=True,
#                 is_test=False,
#                 name='bidirec')

# print(lstm_x.shape)
# print(lstm_y.shape)

use_gpu = True
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
main_program = fluid.default_main_program()
feeder = fluid.DataFeeder(feed_list=['t', 'h','label'], place=place)
exe.run(fluid.default_startup_program())

fetch_var = [x, y, emb_x, emb_y, pad_x, pad_y, lstm_x, x_last_h]

for i, data in enumerate(train_reader()):
    print(data)
    result = exe.run(
        main_program,
        feeder.feed(data),
        fetch_list=fetch_var,
        return_numpy=False
        )
    break

for i, data in enumerate(result):
    print(fetch_var[i].name)
    temp = np.array(data)
    print(data.lod())
    print(temp)
    print(temp.shape)
    print("_____"*10)

毫无双向作用，经过 许多次测试，没有发挥双向作用，而且输出和官方文档描述的不符合！！！
输出形状不符合
版本 1.72-gpu

[guoshengCS]

原fluid.layers.lstm问题较多，已经在高优重新实现了 #25502 ，还望见谅

还请先行使用paddle.fluid.layers.rnn接口 https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/layers_cn/rnn_cn.html#rnn 完成，双向的话暂时需要is_reverse=True和False的两个RNN。

[paddle-bot-old]

Since you haven't replied for more than a year, we have closed this issue/pr.
If the problem is not solved or there is a follow-up one, please reopen it at any time and we will continue to follow up.
由于您超过一年未回复，我们将关闭这个issue/pr。
若问题未解决或有后续问题，请随时重新打开，我们会继续跟进。