You are on page 1of 2

Assignment1:Devnagricharacterrecognition

ThisassignmentistobedoneonTensorflowviaitsPythoninterface.Thelinksheremightbe
helpfulinbuildingthebackground.

Thedatasetcontainslabeledimagesofhandwrittendevnagricharacters.Theimagesareinpng
format.Total10,000imagesarethereintrainset.Yourtaskistopredictcorrectlabelfortest
imagesthatwewillprovide.

TheimagescanbereadusingthepythonAPIscikitimage.Thefollowingexamplecodeisthere
toreadpngimageintonumpyndarray(matrix)

fromskimageimportio
image=io.imread(<filename.png>)

Now,imagewillcontainthenumpyndarray(320*320inourcase).

YourgoalistodesignafeedforwardneuralnetworkforthetaskofDevnagricharacter
recognition.Yourreportshouldincludethefollowing:
1. Theaccuracyonthevalidationdatasetforincreasingnumberoflevels
2. Accuracywithincreasingwidthforafixednumberoflevels
3. EffectofchangingthehiddenunittypefromRelUtoSigmoidtoTanh
4. Accuracywithdifferentlearningrates
5. Effectofdifferentkindsofregularizers:L2regularizervsdropout.

Downloadinformationforthedataset

Thetrainingdatasetcanbedownloadedfromthislink
https://drive.google.com/file/d/0BzIqj5JgNb5RRlo1aUwyTDNEdzg/view

Thevalidationdataset(totestyourmodel)canbedownloadedfromthislink
https://drive.google.com/open?id=0BzIqj5JgNb5Ra0duNW95UDNHN0U

Submissionformatandinstruction

Thereportshouldcontainalinktofoldercontainingallthesourcecodesusedforbothtraining
andpreprocessing(ifany).Also,thefoldershouldcontainthetrainedmodelinabinaryfile,
whichcanbeloadedinTensorflowlateron.Themodelwhengivena320x320pixelimage
shouldoutputtheprobabilitiesofalltheclasses.Theoutputshouldonlycontainprobabilities
seriallyseparatedbyspaceandnothingelse.Yourmodelwillbeevaluatedontestdataonly.


Note:Ifyoupreprocesstheinputbeforeforwardpropagation,provideascriptwhichtakesana
320x320imagefromthelocationprovidedasanargumentandprovidetheabovementioned
output.Thescriptshoulddoallthepreprocessing.