Hi, Is there development happening on fixing issue with SGD that generates models which are as good as random prediction?
I am not sure why such issue is not noticed and raised by others ? May be this specific algo is not used in practical applications. Thanks, Rajesh >> >> On Tue, Oct 16, 2012 at 10:23 PM, Ted Dunning <[email protected]>wrote: >> >>> Rajesh, >>> >>> In the testing that I did, I ran 100, 1000 and 10,000 passes through the >>> data. All produced identical results. Thus it isn't an issue of SGD >>> converging. >>> >>> I also did a parameter scan of lambda and saw no effect. >>> >>> I also did the standard thing in R with glm and got the expected >>> (correct) >>> results. >>> >>> I haven't looked yet in detail, but I really suspect that the reading of >>> the data is horked. This is exactly how that behaves. >>> >>> On Tue, Oct 16, 2012 at 4:49 AM, Rajesh Nikam <[email protected]> >>> wrote: >>> >>> > Hi Ted, >>> > >>> > I was thinking, this might be due to having only 100 instances for >>> > training. >>> > >>> > So I have created test set with two classes having ~49K instances, >>> included >>> > all features as predictors. >>> > PFA sgd.grps.zip with test file. >>> > >>> > mahout trainlogistic --input /usr/local/mahout/trainme/sgd-grps.csv >>> > --output /usr/local/mahout/trainme/sgd-grps.model --target class >>> > --categories 2 --features 128 --types n --predictors a1 a2 a3 a4 a5 a6 >>> a7 >>> > a8 a9 a10 a11 a12 a13 a14 a15 a16 a17 a18 a19 a20 a21 a22 a23 a24 a25 >>> a26 >>> > a27 a28 a29 a30 a31 a32 a33 a34 a35 a36 a37 a38 a39 a40 a41 a42 a43 >>> a44 a45 >>> > a46 a47 a48 a49 a50 a51 a52 a53 a54 a55 a56 a57 a58 a59 a60 a61 a62 >>> a63 a64 >>> > a65 a66 a67 a68 a69 a70 a71 a72 a73 a74 a75 a76 a77 a78 a79 a80 a81 >>> a82 a83 >>> > a84 a85 a86 a87 a88 a89 a90 a91 a92 a93 a94 a95 a96 a97 a98 a99 a100 >>> a101 >>> > a102 a103 a104 a105 a106 a107 a108 a109 a110 a111 a112 a113 a114 a115 >>> a116 >>> > a117 a118 a119 a120 a121 a122 a123 a124 a125 a126 a127 >>> > >>> > >>> > mahout runlogistic --input /usr/local/mahout/trainme/sgd-grps.csv >>> --model >>> > /usr/local/mahout/trainme/sgd-grps.model --auc --confusion >>> > >>> > Still the results are similar, it classifies everything as class_1. >>> > >>> > AUC = 0.50 >>> > confusion: [[*26563.0, 23006.0*], [0.0, 0.0]] >>> > entropy: [[-0.0, -0.0], [-46.1, -21.4]] >>> > >>> > I am not sure why this is failing all the time. >>> > >>> > Looking forward for your reply. >>> > >>> > Thanks >>> > Rajesh >>> > >>> > >>> > >>> > On Tue, Oct 16, 2012 at 3:57 AM, Ted Dunning <[email protected]> >>> > wrote: >>> > >>> > > I would love to help and will before long. Just can't do it in the >>> first >>> > > part of this week. >>> > > >>> > > On Mon, Oct 15, 2012 at 6:28 AM, Rajesh Nikam <[email protected] >>> > >>> > > wrote: >>> > > >>> > > > Hello, >>> > > > >>> > > > I have asked below question on issue with using sgd on mahout >>> forum. >>> > > > >>> > > > Similar issue with sgd is reported by >>> > > > >>> > > > >>> > > >>> > >>> http://stackoverflow.com/questions/11221436/using-sgd-classifier-in-mahout >>> > > > >>> > > > Even below link has similar output: >>> > > > >>> > > > AUC = 0.57*confusion: [[27.0, 13.0], [0.0, 0.0]]* >>> > > > entropy: [[-0.4, -0.3], [-1.2, -0.7]] >>> > > > >>> > > > >>> > > > >>> > >>> http://sujitpal.blogspot.in/2012/09/learning-mahout-classification.html >>> > > > >>> > > > I am still wannder confusion how then this model works and used by >>> > many ? >>> > > > Not able to get any points on how to use SGD that generates >>> effective >>> > > > model. >>> > > > >>> > > > Could someone point out what is missing in input file or provided >>> > > > parameters. >>> > > > >>> > > > I appreciate your help. >>> > > > >>> > > > Below is description of steps that I followed. >>> > > > >>> > > > PF Attached uses input files for experiment. >>> > > > >>> > > > I am using Iris Plants Database from Michael Marshall. PFA >>> iris.arff. >>> > > > Converted this to csv file just by updating header: >>> iris-3-classes.csv >>> > > > >>> > > > mahout org.apache.mahout.classifier. >>> > > > sgd.TrainLogistic --input >>> > > /usr/local/mahout/trunk/*iris-3-classes.csv*--features 4 --output >>> > > /usr/local/mahout/trunk/ >>> > > > *iris-3-classes.model* --target class *--categories 3* --predictors >>> > > > sepallength sepalwidth petallength petalwidth --types n >>> > > > >>> > > > >> it gave following error. >>> > > > Exception in thread "main" java.lang.IllegalArgumentException: Can >>> only >>> > > > call classifyScalar with two categories >>> > > > >>> > > > Now created csv with only 2 classes. PFA iris-2-classes.csv >>> > > > >>> > > > >> trained iris-2-classes.csv with sgd >>> > > > >>> > > > mahout org.apache.mahout.classifier.sgd.TrainLogistic --input >>> > > > /usr/local/mahout/trunk/*iris-2-classes.csv* --features 4 --output >>> > > > /usr/local/mahout/trunk/*iris-2-classes.mode*l --target class >>> > > *--categories >>> > > > 2* --predictors sepallength sepalwidth petallength petalwidth >>> --types n >>> > > > >>> > > > mahout runlogistic --input >>> /usr/local/mahout/trunk/iris-2-classes.csv >>> > > > --model /usr/local/mahout/trunk/iris-2-classes.model --auc >>> --confusion >>> > > > >>> > > > AUC = 0.14 >>> > > > confusion: [[50.0, 50.0], [0.0, 0.0]] >>> > > > entropy: [[-0.6, -0.3], [-0.8, -0.4]] >>> > > > >>> > > > >> AUC seems to poor. Now changed --predictors >>> > > > >>> > > > mahout org.apache.mahout.classifier.sgd.TrainLogistic --input >>> > > > /usr/local/mahout/trunk/*iris-2-classes.csv* --features 4 --output >>> > > > /usr/local/mahout/trunk/*iris-2-classes.mode*l --target class >>> > > *--categories >>> > > > 2* --predictors sepalwidth petallength --types n >>> > > > >>> > > > mahout runlogistic --input >>> /usr/local/mahout/trunk/iris-2-classes.csv >>> > > > --model /usr/local/mahout/trunk/iris-2-classes.model --auc >>> --confusion >>> > > > --scores >>> > > > >>> > > > AUC = 0.80 >>> > > > *confusion: [[50.0, 50.0], [0.0, 0.0]]* >>> > > > entropy: [[-0.7, -0.3], [-0.7, -0.4]] >>> > > > >>> > > > This model classifies everything as category 1 which of no use. >>> > > > >>> > > > Thanks >>> > > > Rajesh >>> > > > >>> > > > >>> > > > >>> > > > >>> > > >>> > >>> >> >> >
