火炬 nngraph/init.lua:31: 输入 [1] 为零（错字/错误索引？）

Question

我是 Torch7 的新手。我使用 nngraph 模型定义了一个火炬网。当 nngraph init.lua 初始化这个网络时，结果是错误 “ckage/torch/torch/install/share/lua/5.1/nngraph/init.lua:31：输入[1] 为零（错字/错误索引？）”。我给出了净输入，为什么它说输入[1]是零？以下是源代码。

function attention.recursive_atten_revise(input_size, embedding_size,      last_embed_size, output_size)
local inputs = {}
local outputs = {}
table.insert(inputs, nn.Identity()())
table.insert(inputs, nn.Identity()())
table.insert(inputs, nn.Identity()())
table.insert(inputs, nn.Identity()())
table.insert(inputs, nn.Identity()())
table.insert(inputs, nn.Identity()())
table.insert(inputs, nn.Identity()())

local img_feat = inputs[1]
local embed_ques = inputs[2]
local embed_img = inputs[3]
local conv_ques = inputs[4]
local conv_img = inputs[5]
local lstm_ques = inputs[6]
local lstm_img = inputs[7]
print('initialize inputs.')
-- word level image attention split and join
local img1_feat = nn.Narrow(2,1,1)(img_feat)
local img2_feat = nn.Narrow(2,2,1)(img_feat)
local img3_feat = nn.Narrow(2,3,1)(img_feat)

local w_att1_img = nn.Narrow(2,1,1)(embed_img)
local w_att2_img = nn.Narrow(2,2,1)(embed_img)
local w_att3_img = nn.Narrow(2,3,1)(embed_img)-- need to check if it is n_batch*1

local rep_img_atten1 = nn.Replicate(input_size,2)(w_att1_img)
local rep_img_atten1_dim = nn.View(-1,input_size)(rep_img_atten1)
local rep_img_atten2 = nn.Replicate(input_size,2)(w_att2_img)
local rep_img_atten2_dim = nn.View(-1,input_size)(rep_img_atten2)
local rep_img_atten3 = nn.Replicate(input_size,2)(w_att3_img)
local rep_img_atten3_dim = nn.View(-1,input_size)(rep_img_atten3)

local w_att_feat1 = nn.CMulTable()({img1_feat,rep_img_atten1_dim})
local w_att_feat2 = nn.CMulTable()({img2_feat,rep_img_atten2_dim})
local w_att_feat3 = nn.CMulTable()({img3_feat,rep_img_atten3_dim})
local w_join_feat = nn.JoinTable(2)({w_att_feat1,w_att_feat2,w_att_feat3,embed_ques})
print('word level recursive')
-- phrase level image attention split and join
local p_att1_img = nn.Narrow(2,1,1)(conv_img)
local p_att2_img = nn.Narrow(2,2,1)(conv_img)
local p_att3_img = nn.Narrow(2,3,1)(conv_img)

local rep_p_att1 = nn.Replicate(input_size,2)(p_att1_img)
local rep_p_att1_dim = nn.View(-1,input_size)(rep_p_att1)
local rep_p_att2 = nn.Replicate(input_size,2)(p_att2_img)
local rep_p_att2_dim = nn.View(-1,input_size)(rep_p_att2)
local rep_p_att3 = nn.Replicate(input_size,2)(p_att3_img)
local rep_p_att3_dim = nn.View(-1,input_size)(rep_p_att3)

local p_att_feat1 = nn.CMulTable()({img1_feat,rep_p_att1_dim})
local p_att_feat2 = nn.CMulTable()({img2_feat,rep_p_att2_dim})
local p_att_feat3 = nn.CMulTable()({img3_feat,rep_p_att3_dim})
local p_join_feat = nn.JoinTable(2)({p_att_feat1,p_att_feat2,p_att_feat3,conv_ques})
print('phrase level recursive')

-- question level image attention split and join
local q_att1_img = nn.Narrow(2,1,1)(lstm_img)
local q_att2_img = nn.Narrow(2,2,1)(lstm_img)
local q_att3_img = nn.Narrow(2,3,1)(lstm_img)

local rep_q_att1 = nn.Replicate(input_size,2)(q_att1_img)
local rep_q_att1_dim = nn.View(-1,input_size)(rep_q_att1)
local rep_q_att2 = nn.Replicate(input_size,2)(q_att2_img)
local rep_q_att2_dim = nn.View(-1,input_size)(rep_q_att2)
local rep_q_att3 = nn.Replicate(input_size,2)(q_att3_img)
local rep_q_att3_dim = nn.View(-1,input_size)(rep_q_att3)

local q_att_feat1 = nn.CMulTable()({img1_feat,rep_q_att1_dim})
local q_att_feat2 = nn.CMulTable()({img2_feat,rep_q_att2_dim})
local q_att_feat3 = nn.CMulTable()({img3_feat,rep_q_att3_dim})
local q_join_feat = nn.JoinTable(2)({q_att_feat1,q_att_feat2,q_att_feat3,lstm_ques})
print('question level recursive')

local w_join_feat_dim = nn.View(-1,input_size*4,1)(w_join_feat_dim)
local p_join_feat_dim = nn.View(-1,input_size*4,1)(p_join_feat_dim)
local q_join_feat_dim = nn.View(-1,input_size*4,1)(q_join_feat_dim)
local combine_feat = nn.JoinTable(3)({w_join_feat_dim,p_join_feat_dim,q_join_feat_dim})
local max_feat = nn.Max(3)(combine_feat)

local hidden1 = nn.Tanh()(nn.Linear(input_size*4,embedding_size)(max_feat))
local hidden2 = nn.Tanh()(nn.Linear(embedding_size, last_embed_size)(nn.Dropout(0.5)(hidden1)))
local out_feat = nn.Linear(last_embed_size, output_size)(nn.Dropout(0.5)(hidden2))

table.insert(outputs,out_feat)
return nn.gModule(inputs,outputs)

end

有人可以帮我吗？非常感谢。

Answer 1

您可能已经解决了这个问题，但无论如何。

您的错误指向L31，它清楚地表明您在某处有一个空输入：

if nArgs == 1 and input == nil then
   error(utils.expectingNodeErrorMessage(input, 'inputs', 1))
end

如果我正确理解您的代码，这就是您传递空输入的地方：

local w_join_feat_dim = nn.View(-1,input_size*4,1)(w_join_feat_dim)
local p_join_feat_dim = nn.View(-1,input_size*4,1)(p_join_feat_dim)
local q_join_feat_dim = nn.View(-1,input_size*4,1)(q_join_feat_dim)

你的意思可能是：

local w_join_feat_dim = nn.View(-1,input_size*4,1)(w_join_feat)
local p_join_feat_dim = nn.View(-1,input_size*4,1)(p_join_feat)
local q_join_feat_dim = nn.View(-1,input_size*4,1)(q_join_feat)

*_join_feat_dim 在这些行之前没有被初始化。