Fix bugs

README.md

@@ -1,12 +1,12 @@
-##XNOR-Net: ImageNet Classification Using Binary Convolutional Neurl Networks.
+##XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks.
 
-This is the Torch 7.0 implementation of XNOR-Net: ImageNet Classification Using Binary Convolutional Neurl Networks.
+This is the Torch 7.0 implementation of XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks.
 
 ### Citation 
 ```bash
 @inproceedings{rastegariECCV16,
     Author = {Mohammad Rastegari and Vicente Ordonez and Joseph Redmon and Ali Farhadi},
-    Title = {XNOR-Net: ImageNet Classification Using Binary Convolutional Neurl Networks},
+    Title = {XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks},
     Booktitle = {ECCV},
     Year = {2016}
 }
@@ -15,19 +15,36 @@ This is the Torch 7.0 implementation of XNOR-Net: ImageNet Classification Using
 ### Requirements
 This software is implemented on top of the implementation of [ImageNet-multiGPU](https://github.com/soumith/imagenet-multiGPU.torch) and has all the same requirements.
 
+Download the cache files from [here](https://s3-us-west-2.amazonaws.com/ai2-vision/xnornet/cache.tar) and run:
+```bash
+tar -xvf ./cache.tar
+```
+
 ### Training Binary Weight Network
-You just need to run this:
+
 ```bash
 th main.lua -data [path to ImageNet dataset] -nGPU 1 -batchSize 128 -netType alexnet -binaryWeight -dropout 0.1
 ``` 
 ### Training XNOR-Networks
 ```bash
-th main.lua -data /mnt/raid00/data/imagenet2012/ -nGPU 4 -batchSize 800 -netType alexnetxnor -binaryWeight -optimType adam -epochSize 1500
+th main.lua -data [path to ImageNet dataset] -nGPU 4 -batchSize 800 -netType alexnetxnor -binaryWeight -optimType adam -epochSize 1500
 ```
 ### Trained Models
-To use the trained models use the option `-retrain`
+To use the trained models use the option `-retrain [path to the trained model file]` and `-testOnly`
+
 [Binary-Weight-Network(BWN)](https://s3-us-west-2.amazonaws.com/ai2-vision/xnornet/alexnet_BWN.t7)
+
 [XNOR-Network](https://s3-us-west-2.amazonaws.com/ai2-vision/xnornet/alexnet_XNOR.t7)
 
+If you use the same image pre-processing as [here](https://github.com/soumith/imagenet-multiGPU.torch) by
+```bash
+find . -name "*.JPEG" | xargs -I {} convert {} -resize "256^>" {}
+``` 
+then the accuracies for top-1 should be: 
+
+alexnet_BWN  %56.8
+
+alexnet_XNOR %43.3 
+
 ### License
-By downloading this software you acknowledged that you agreed on the terms and conditions in the `SOFTWARE-LICENSE-AGREEMENT.lic`
+By downloading this software you acknowledged that you agreed on the terms and conditions in the `SOFTWARE-LICENSE-AGREEMENT.lic`

donkey.lua

@@ -56,10 +56,14 @@ local mean,std
 
 -- function to load the image, jitter it appropriately (random crops etc.)
 local trainHook = function(self, path)
+   if string.find(path,"/mnt/raid00/data/imagenet2012") then
+     path = string.gsub(path ,"/mnt/raid00/data/imagenet2012",opt.data)
+   end
    collectgarbage()
    opt.testMode = false
    local excep, input = pcall(loadImage,path);
    if not excep then
+      print('ERROR: image can not be loaded!!!')
       input = torch.Tensor(3,opt.imageSize,opt.imageSize):fill(mean[1]);
    end
    local iW = input:size(3)
@@ -87,9 +91,13 @@ if paths.filep(trainCache) then
    print('Loading train metadata from cache')
    trainLoader = torch.load(trainCache)
    trainLoader.sampleHookTrain = trainHook
-   assert(trainLoader.paths[1] == paths.concat(opt.data, 'train'),
-          'cached files dont have the same path as opt.data. Remove your cached files at: '
-             .. trainCache .. ' and rerun the program')
+   --assert(trainLoader.paths[1] == paths.concat(opt.data, 'train'),
+   --       'cached files dont have the same path as opt.data. Remove your cached files at: '
+   --          .. trainCache .. ' and rerun the program')
+   if trainLoader.paths[1] ~= paths.concat(opt.data, 'train') then
+      trainLoader.paths[1] = paths.concat(opt.data, 'train');
+   end
+
 else
    print('Creating train metadata')
    trainLoader = dataLoader{
@@ -122,11 +130,15 @@ end
 
 -- function to load the image
 testHook = function(self, path)
+   if string.find(path,"/mnt/raid00/data/imagenet2012") then
+     path = string.gsub(path ,"/mnt/raid00/data/imagenet2012",opt.data)
+   end
    collectgarbage()
    opt.testMode = true
    local excep, input = pcall(loadImage,path);
    if not excep then
-      input = torch.CudaTensor(3,opt.imageSize,opt.imageSize):fill(mean[1]);
+      print('ERROR: image can not be loaded!!!')
+      input = torch.Tensor(3,opt.imageSize,opt.imageSize):fill(mean[1]);
    end
    local oH = sampleSize[2]
    local oW = sampleSize[3]
@@ -147,9 +159,12 @@ if paths.filep(testCache) then
    print('Loading test metadata from cache')
    testLoader = torch.load(testCache)
    testLoader.sampleHookTest = testHook
-   assert(testLoader.paths[1] == paths.concat(opt.data, 'val'),
-          'cached files dont have the same path as opt.data. Remove your cached files at: '
-             .. testCache .. ' and rerun the program')
+   --assert(testLoader.paths[1] == paths.concat(opt.data, 'val'),
+   --       'cached files dont have the same path as opt.data. Remove your cached files at: '
+   --          .. testCache .. ' and rerun the program')
+   if testLoader.paths[1] ~= paths.concat(opt.data, 'val') then
+      testLoader.paths[1] = paths.concat(opt.data, 'val');
+   end
 else
    print('Creating test metadata')
    testLoader = dataLoader{

main.lua

@@ -43,9 +43,12 @@ paths.dofile('test.lua')
 
 
 epoch = opt.epochNumber
-
-for i=1,opt.nEpochs do
+if opt.testOnly then
+	test()
+else
+  for i=1,opt.nEpochs do
    train()
    test()
    epoch = epoch + 1
-end
+  end
+end

model.lua

@@ -36,10 +36,7 @@ else
       loadParams(model,saved_model);
    end
 end
-
-debugger.enter()
-
-
+--debugger.enter()
 -- This is useful for fitting ResNet-50 on 4 GPUs, but requires that all
 -- containers override backwards to call backwards recursively on submodules
 if opt.shareGradInput then

opts.lua

@@ -16,7 +16,7 @@ function M.parse(arg)
     cmd:text('Options:')
     ------------ General options --------------------
 
-    cmd:option('-cache', '/mnt/raid00/' .. os.getenv('USER') ..'/NNExpLog/imagenet1k/', 'subdirectory in which to save/log experiments')
+    cmd:option('-cache', './cache/', 'subdirectory in which to save/log experiments')
     cmd:option('-data', './imagenet/imagenet_raw_images/256', 'Home of ImageNet dataset')
     cmd:option('-dataset',  'imagenet', 'Dataset Name: imagenet |cifar')
     cmd:option('-manualSeed',         2, 'Manually set RNG seed')
@@ -41,6 +41,7 @@ function M.parse(arg)
     cmd:option('-weightDecay',     0, 'weight decay')
     cmd:option('-shareGradInput',  true, 'Sharing the gradient memory')
     cmd:option('-binaryWeight',    false, 'Sharing the gradient memory')
+    cmd:option('-testOnly',    false, 'Sharing the gradient memory')
     ---------- Model options ----------------------------------
     cmd:option('-netType',     'alexnet', 'Options: alexnet | overfeat | alexnetowtbn | vgg | googlenet | resnet')
     cmd:option('-optimType',     'sgd', 'Options: sgd | adam')

test.lua

@@ -28,7 +28,7 @@ function test()
    if opt.binaryWeight then
       binarizeConvParms(convNodes)
    end
-
+   
    top1Sum = 0
    top5Sum = 0
    loss = 0

util.lua

@@ -82,7 +82,7 @@ function updateBinaryGradWeight(convNodes)
     local m = convNodes[i].weight:norm(1,4):sum(3):sum(2):div(n):expand(s);
     m[convNodes[i].weight:le(-1)]=0;
     m[convNodes[i].weight:ge(1)]=0;
-    m:add(1/(n)):mul(1-1/s[2]);
+    m:add(1/(n)):mul(1-1/s[2]):mul(n);
     convNodes[i].gradWeight:cmul(m)--:cmul(mg)
    end
    if opt.nGPU >1 then