Created by: zhaoyuchen2018

使用下面的test case测试StaticRNN，发现w1的梯度没有正确的传递出去。

import sys import io import numpy as np

import paddle import paddle.fluid as fluid import paddle.fluid.core as core import paddle.fluid.framework as framework from paddle.fluid.executor import Executor

import paddle.fluid.layers as layers from paddle.fluid.layers.control_flow import StaticRNN

batch_size = 1 seq_len = 1 hidden_size = 10

x = layers.data( name='x', shape=[-1, batch_size, hidden_size], dtype='float32') x += 0.0 emb = layers.data( name='emb', shape=[-1, batch_size, hidden_size], dtype='float32')

w1 = layers.create_parameter( shape=[hidden_size, hidden_size], dtype='float32', name='w1')

w2 = layers.create_parameter( shape=[hidden_size, hidden_size], dtype='float32', name='w2')

pre_h = layers.fill_constant( shape=[ seq_len, hidden_size], dtype='float32', value = 0.0)

pre_h.stop_gradient = False

def dot_attention( query, memory ): attn = layers.matmul( query, memory, transpose_y=True) weight = layers.softmax( attn ) weight_memory = layers.matmul( weight, memory)

return weight_memory, weight

rnn = StaticRNN()

emb += 0.0

y = layers.matmul(x, w1)

with rnn.step():

step_in = rnn.step_input(x)

pre_h = rnn.memory( init= pre_h )

new_h = layers.matmul( pre_h, w2 )

new_h, _ = dot_attention( new_h, y)

rnn.update_memory( pre_h, new_h )

rnn.step_output( new_h )

rnn_out = rnn()

loss = layers.reduce_sum( rnn_out )

loss.persistable = True

sgd = fluid.optimizer.SGD(1.0)

sgd.minimize( loss )

fluid.default_startup_program().random_seed = 123

place = core.CPUPlace() exe = Executor(place) exe.run(framework.default_startup_program() )

param_list = fluid.default_main_program().block(0).all_parameters()

np.random.seed( 10 )

w1_np = np.random.uniform( -0.1, 0.1, ( hidden_size, hidden_size)).astype( 'float32') w2_np = np.random.uniform( -0.1, 0.1, ( hidden_size , hidden_size)).astype( "float32")

fluid.global_scope().find_var( 'w1' ).get_tensor().set( w1_np, place) fluid.global_scope().find_var( 'w2' ).get_tensor().set( w2_np, place)

input = np.random.uniform( -0.1, 0.1, ( seq_len, batch_size, hidden_size)).astype( 'float32' ) emb = np.random.uniform( -0.1, 0.1, ( seq_len, batch_size, hidden_size)).astype( 'float32' )

out = exe.run( feed={ 'x' : input, 'emb' : emb }, fetch_list=[loss.name ])

new_w1 = np.asarray( fluid.global_scope().find_var( 'w1' ).get_tensor() ) new_w2 = np.asarray( fluid.global_scope().find_var( 'w2' ).get_tensor() )

print( new_w1 - w1_np )

打印出来的new_w1 - w1_np全为0.