In [1]:
--miscellaneous imports
require 'torch'
require 'nn'
require 'nngraph'
require 'image'
require 'cutorch'
require 'cunn'
require 'lfs'
require 'image'
modelPath = paths.concat(lfs.currentdir(), 'bestmodels/filt=5-hidden1=5-hidden2=15-hidden3=15-fc1=150-fc2=150/model#180.net')
--'bestmodels/filt=5-hidden1=5-hidden2=15-hidden3=5-fc1=300-fc2=150/model#150.net')
In [2]:
--load the model
model = torch.load(modelPath)
In [26]:
--visualize the first layer kernels
conv1 = model:get(1)
weights = conv1.weight
images = {}
for index = 1, weights:size(1) do
kernel = weights[index]
kernel:resize(3,5,5)
kernel = kernel:double()
images[index] = image.scale(kernel, 75, 75, 'simple')
end
itorch.image(images, {nrow = 6, padding = 20})
In [25]:
--visualize the first layer kernels
conv2 = model:get(3)
--print(conv2)
weights = conv2.weight
for index = 1, weights:size(1) do
kernel = weights[index]
kernel:resize(5, 5, 5)
images = {}
for weightIndex = 1, kernel:size(1) do
individualKernel = kernel[weightIndex]
individualKernel = individualKernel:double()
images[weightIndex] = image.scale(individualKernel, 75, 75)
--print(images[weightIndex]:size())
end
itorch.image(images, {nrow = 6, padding = 20})
end
In [4]:
--visualize the output of layers
dofile '1_data.lua'
Out[4]:
Out[4]:
Out[4]:
In [5]:
classes = {
[1] = 'cow',
[2] = 'boat',
[3] = 'sheep',
[4] = 'bottle',
[5] = 'bird',
}
function getClassName(number)
return classes[number]
end
In [6]:
dofile 'utilities.lua'
function getSampleImagesForClasses(samples, labels)
sampleImages = {}
rawImages = {}
for index = 1, samples:size(1) do
--assign emptz table if key doesn't exists
name = getClassName(labels[index])
if(sampleImages[name] == nil) then
sampleImages[name] = {}
rawImages[name] = {}
end
allKeys = table.getAllKeys(sampleImages[name])
if(#allKeys <= 20) then
table.insert(sampleImages[name], image.scale(samples[index], 60, 60))
table.insert(rawImages[name], samples[index])
end
end
return sampleImages, rawImages
end
In [7]:
first50Samples = trainData.data[{ {1,150},{} }]
labels = trainData.labels[{ {1,150}}]
sampleImages, rawImages = getSampleImagesForClasses(first50Samples, labels)
for k, v in pairs(sampleImages) do
print(k)
itorch.image(sampleImages[k] , {nrow=5, padding = 5, zoom=5, scaleeach=5})
end
Out[7]:
Out[7]:
Out[7]:
Out[7]:
Out[7]:
Out[7]:
In [8]:
--forward the input for each of the classes and print the layer outputs
function saveLayerOutputs(className, samples, index)
imageInstance = samples[className][index]
imageInstance = imageInstance:cuda()
pred = model:forward(imageInstance)
print(className)
itorch.image(imageInstance)
convLayers = {1, 3, 5}
for k, v in ipairs(convLayers) do
print('convolution layer ' .. k)
conv = model:get(v)
outputBuffer = {}
for index = 1, conv.output:size(1) do
currOutput = conv.output[index]:double()
currOutput = image.scale(currOutput, 50, 50)
currOutput:resize(1, currOutput:size(1), currOutput:size(2))
currOutput = (currOutput - currOutput:min()) * 100
table.insert(outputBuffer, image.y2jet(currOutput + 1))
end
itorch.image(outputBuffer, {nrow = 10, padding = 5})
end
end
In [9]:
saveLayerOutputs('cow', rawImages, 8)
Out[9]:
Out[9]:
Out[9]:
Out[9]:
Out[9]:
In [10]:
saveLayerOutputs('boat', rawImages, 8)
Out[10]:
Out[10]:
Out[10]:
Out[10]:
Out[10]:
In [11]:
saveLayerOutputs('sheep', rawImages, 13)
Out[11]:
Out[11]:
Out[11]:
Out[11]:
Out[11]:
In [12]:
saveLayerOutputs('bottle', rawImages, 3)
Out[12]:
Out[12]:
Out[12]:
Out[12]:
Out[12]:
In [17]:
saveLayerOutputs('bird', rawImages, 3)
Out[17]:
Out[17]:
Out[17]:
Out[17]:
Out[17]:
In [14]:
function collectMisclassifiedImages(actualLabels, predictedLabels, imgs)
misclassifiedImgs = {}
count = 1;
for index = 1, imgs:size(1) do
if(actualLabels[index] ~= predictedLabels[index]) then
actualClass = getClassName(actualLabels[index])
predictedClass = getClassName(predictedLabels[index])
if(misclassifiedImgs[actualClass] == nil) then
misclassifiedImgs[actualClass] = {}
end
if(misclassifiedImgs[actualClass][predictedClass] == nil) then
misclassifiedImgs[actualClass][predictedClass] = {}
end
table.insert(misclassifiedImgs[actualClass][predictedClass], imgs[index])
end
end
return misclassifiedImgs
end
In [15]:
--get the wrong-predicted images
dofile '6_test.lua'
--dofile '5_validate.lua'
require 'optim'
dofile 'utilities.lua'
confusion = optim.ConfusionMatrix(classes)
--validationData.data = validationData.data:cuda()
--validate()
testData.data = testData.data:cuda()
predictedLabels = test()
misclassifiedImgs = collectMisclassifiedImages(testData.labels, predictedLabels, testData.data)
Out[15]:
Out[15]:
Out[15]:
In [16]:
picsFolder = '/media/arul/envision/jan-may-2016/kernelmethods/Assignment-2/report/pics/'
--display misclassified images
for actualClass,v in pairs(misclassifiedImgs) do
--k is actual class name, v is a table of images with key as 'predicted class'
for predictedClass, imgs in pairs(v) do
fileName = 'miscallssify_' .. actualClass .. '_to_' .. predictedClass .. '.png'
currImage = image.toDisplayTensor({input=imgs, nrow = 6, padding = 5, zoom=100})
image.save(picsFolder .. fileName, currImage)
end
end