2016-08-15 112 views
1

我有一个Lusty(OpenResty的框架)API,它包装了一个Torch分类器。到目前为止,我已经能够得到一个单一的请求工作,但是到API每个后续请求触发以下错误,没有详细的堆栈跟踪:Lua Torch7&OpenResty:试图索引一个零值

attempt to index a nil value 

出现的错误,当我打电话给抛出:

net:add(SpatialConvolution(3, 96, 7, 7, 2, 2)) 

成功完成第一个请求而每个附加请求失败的行为是解决问题的线索。

我已将以下完整代码粘贴到app/requests/classify.lua。这似乎是某种变量缓存/初始化问题,虽然我对Lua的有限知识不能帮助我调试问题。我已经尝试过做很多事情,包括将我的导入改为local torch = require('torch')等本地化变量,并将这些导入移到classifyImage()函数中。

torch = require 'torch' 
nn = require 'nn' 
image = require 'image' 
ParamBank = require 'ParamBank' 
label  = require 'classifier_label' 
torch.setdefaulttensortype('torch.FloatTensor') 

function classifyImage() 

    local opt = { 
    inplace = false, 
    network = "big", 
    backend = "nn", 
    save = "model.t7", 
    img = context.input.image, 
    spatial = false, 
    threads = 4 
    } 
    torch.setnumthreads(opt.threads) 

    require(opt.backend) 
    local SpatialConvolution = nn.SpatialConvolutionMM 
    local SpatialMaxPooling = nn.SpatialMaxPooling 
    local ReLU = nn.ReLU 
    local SpatialSoftMax = nn.SpatialSoftMax 

    local net = nn.Sequential() 

    print('==> init a big overfeat network') 
    net:add(SpatialConvolution(3, 96, 7, 7, 2, 2)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialMaxPooling(3, 3, 3, 3)) 
    net:add(SpatialConvolution(96, 256, 7, 7, 1, 1)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialMaxPooling(2, 2, 2, 2)) 
    net:add(SpatialConvolution(256, 512, 3, 3, 1, 1, 1, 1)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialConvolution(512, 512, 3, 3, 1, 1, 1, 1)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialConvolution(512, 1024, 3, 3, 1, 1, 1, 1)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialConvolution(1024, 1024, 3, 3, 1, 1, 1, 1)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialMaxPooling(3, 3, 3, 3)) 
    net:add(SpatialConvolution(1024, 4096, 5, 5, 1, 1)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialConvolution(4096, 4096, 1, 1, 1, 1)) 
    net:add(ReLU(opt.inplace)) 
    net:add(SpatialConvolution(4096, 1000, 1, 1, 1, 1)) 
    net:add(nn.View(1000)) 
    net:add(SpatialSoftMax()) 
    -- print(net) 

    -- init file pointer 
    print('==> overwrite network parameters with pre-trained weigts') 
    ParamBank:init("net_weight_1") 
    ParamBank:read(  0, {96,3,7,7},  net:get(1).weight) 
    ParamBank:read( 14112, {96},   net:get(1).bias) 
    ParamBank:read( 14208, {256,96,7,7}, net:get(4).weight) 
    ParamBank:read( 1218432, {256},   net:get(4).bias) 
    ParamBank:read( 1218688, {512,256,3,3}, net:get(7).weight) 
    ParamBank:read( 2398336, {512},   net:get(7).bias) 
    ParamBank:read( 2398848, {512,512,3,3}, net:get(9).weight) 
    ParamBank:read( 4758144, {512},   net:get(9).bias) 
    ParamBank:read( 4758656, {1024,512,3,3}, net:get(11).weight) 
    ParamBank:read( 9477248, {1024},   net:get(11).bias) 
    ParamBank:read( 9478272, {1024,1024,3,3}, net:get(13).weight) 
    ParamBank:read(18915456, {1024},   net:get(13).bias) 
    ParamBank:read(18916480, {4096,1024,5,5}, net:get(16).weight) 
    ParamBank:read(123774080, {4096},   net:get(16).bias) 
    ParamBank:read(123778176, {4096,4096,1,1}, net:get(18).weight) 
    ParamBank:read(140555392, {4096},   net:get(18).bias) 
    ParamBank:read(140559488, {1000,4096,1,1}, net:get(20).weight) 
    ParamBank:read(144655488, {1000},   net:get(20).bias) 

    ParamBank:close() 

    -- load and preprocess image 
    print('==> prepare an input image') 
    local img = image.load(opt.img):mul(255) 

    -- use image larger than the eye size in spatial mode 
    if not opt.spatial then 
    local dim = (opt.network == 'small') and 231 or 221 
    local img_scale = image.scale(img, '^'..dim) 
    local h = math.ceil((img_scale:size(2) - dim)/2) 
    local w = math.ceil((img_scale:size(3) - dim)/2) 
    img = image.crop(img_scale, w, h, w + dim, h + dim):floor() 
    end 

    -- memcpy from system RAM to GPU RAM if cuda enabled 
    if opt.backend == 'cunn' or opt.backend == 'cudnn' then 
    net:cuda() 
    img = img:cuda() 
    end 

    -- save bare network (before its buffer filled with temp results) 
    print('==> save model to:', opt.save) 
    torch.save(opt.save, net) 

    -- feedforward network 
    print('==> feed the input image') 
    timer = torch.Timer() 
    img:add(-118.380948):div(61.896913) 
    local out = net:forward(img) 

    -- find output class name in non-spatial mode 
    local results = {} 
    local topN = 10 
    local probs, idxs = torch.topk(out, topN, 1, true) 

    for i=1,topN do 
    print(label[idxs[i]], probs[i]) 
    local r = {} 
    r.label = label[idxs[i]] 
    r.prob = probs[i] 
    results[i] = r 
    end 

    return results 
end 

function errorHandler(err) 
    return tostring(err) 
end 

local success, result = xpcall(classifyImage, errorHandler) 


context.template = { 
    type = "mustache", 
    name = "app/templates/layout", 

    partials = { 
    content = "app/templates/classify", 
    } 
} 


context.output = { 
    success = success, 
    result = result, 
    request = context.input 
} 

context.response.status = 200 

感谢您的帮助!

更新1

新增print(net)之前和之后local net而且之后我打电话net:add。每次在local net初始化之前,它显示的值为nil。正如预期的那样,在初始化net之后,它显示了一个火炬对象作为值。看样子:add调用里面的东西被创造的错误,所以我加了声明我classifyImage功能后立即以下几点:

print(tostring(torch)) 
print(tostring(nn)) 
print(tostring(net)) 

添加这些新的打印报表后,我得到的一号请求以下

nil 
nil 
nil 

,然后在第二个请求:

table: 0x41448a08 
table: 0x413bdb10 
nil 

和位于3要求:

table: 0x41448a08 
table: 0x413bdb10 
nil 

那些看起来像指向内存中的对象的指针,所以在这里假设Torch正在创建它自己的全局对象是安全的吗?

+0

尝试在调用之前和之后放置'print(net)'。 – hjpotter92

+0

很快完成并在问题中添加详细信息。本质上,在我在第一/第二次调用中删除“本地网络”之前,我成功地获得了'nil'。初始化'net'后,我也得到一个新的对象。只有当我调用add时,它才会失败。你认为这与'torch'或'nn'本身有关? – crockpotveggies

+0

@ hjpotter92增加了一些更多的信息,看起来像'torch'本身正在创建内存中的干扰代码的全局对象? – crockpotveggies

回答

0

当需要torch及其模块时,它最终创建一个自身的全局实例,该实例在整个过程的整个过程中保留在内存中。为我工作的修复是引用火炬在主app.lua文件中精力充沛并粘贴以下顶部:

require 'torch' 
require 'nn' 

image = require 'image' 
ParamBank = require 'ParamBank' 
label  = require 'classifier_label' 
torch.setdefaulttensortype('torch.FloatTensor') 
torch.setnumthreads(4) 

SpatialConvolution = nn.SpatialConvolutionMM 
SpatialMaxPooling = nn.SpatialMaxPooling 
ReLU = nn.ReLU 
SpatialSoftMax = nn.SpatialSoftMax 

的变量范围为classifyImage,现在它的每个请求成功。这是一个肮脏的问题,但由于Torch正在维护它自己的全局对象,所以我无法看到它的解决方法。

相关问题