{ "id": 21996, "title": "XGB newbie stuck...", "url": "/competitions/grupo-bimbo-inventory-demand/discussion/21996", "author_name": "", "post_date": "2016-07-01T17:08:10.677Z", "votes": null, "comment_count": 2, "views": 933, "content": "

I am new to using XGBoost and trying to use a regression objective (then want to do poisson if I can ever get it working).

\n\n

I've attempted to use the following code and at the point of PREDICT I get the error:\nError in xgb.DMatrix(newdata) : \n REAL() can only be applied to a 'numeric', not a 'integer'

\n\n

I've tried using lapply to the matrix, I've tried setting a variable to numeric, the training seems to move through, it's the prediction step that bombs.

\n\n

train$id <- NULL;\ntest$id <- NULL;

\n\n

cat("Treating features...\\n");

\n\n

impute train and test here

\n\n

one hot encoding

\n\n

sparse_matrix <- sparse.model.matrix(Demanda_uni_equil ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = train)\nsparse_matrix <- sparse.model.matrix( ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = test)\ntrain$Cliente_ID<-as.numeric(train$Cliente_ID)\ntrain.y = train$Demanda_uni_equil;

\n\n

cat("Dropping features...\\n");\ntrain$Demanda_uni_equil <- NULL;

\n\n

cat("Creating data.matrix...\\n");\npreds <- rep(0,nrow(test));\ntrainM<-data.matrix(train, rownames.force = NA);\ncat("Creating DMatrix for xgboost...\\n");\ndtrain <- xgb.DMatrix(data=trainM, label=train.y, missing = NaN);\nwatchlist <- list(trainM=dtrain);

\n\n

set.seed(1969);

\n\n

param <- list( objective = "reg:linear", \n booster = "gblinear",\n #eval_metric ="mae",\n eta = 0.035,\n max_depth = 6,\n subsample = 1.00,\n colsample_bytree = 0.40\n)

\n\n

clf <- xgb.cv( params = param, \n data = dtrain, \n nrounds = 200, \n verbose = 1,\n watchlist = watchlist,\n maximize = FALSE,\n nfold = 3,\n early.stop.round = 10,\n print.every.n = 1\n);

\n\n

bestRound <- which.min( as.matrix(clf)[,3] );\ncat("Best round:", bestRound,"\\n");\ncat("Best result:",min(as.matrix(clf)[,3]),"\\n");

\n\n

clf <- xgb.train( params = param, \n data = dtrain, \n nrounds = bestRound, \n verbose = 1,\n watchlist = watchlist,\n maximize = FALSE\n)\ntestM <-data.matrix(test, rownames.force = NA);

\n\n

preds <- predict(clf, testM);

", "messages": [ { "id": "125701", "postDate": "07/01/2016 17:08:10", "content": "

I am new to using XGBoost and trying to use a regression objective (then want to do poisson if I can ever get it working).

\n\n

I've attempted to use the following code and at the point of PREDICT I get the error:\nError in xgb.DMatrix(newdata) : \n REAL() can only be applied to a 'numeric', not a 'integer'

\n\n

I've tried using lapply to the matrix, I've tried setting a variable to numeric, the training seems to move through, it's the prediction step that bombs.

\n\n

train$id <- NULL;\ntest$id <- NULL;

\n\n

cat("Treating features...\\n");

\n\n

impute train and test here

\n\n

one hot encoding

\n\n

sparse_matrix <- sparse.model.matrix(Demanda_uni_equil ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = train)\nsparse_matrix <- sparse.model.matrix( ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = test)\ntrain$Cliente_ID<-as.numeric(train$Cliente_ID)\ntrain.y = train$Demanda_uni_equil;

\n\n

cat("Dropping features...\\n");\ntrain$Demanda_uni_equil <- NULL;

\n\n

cat("Creating data.matrix...\\n");\npreds <- rep(0,nrow(test));\ntrainM<-data.matrix(train, rownames.force = NA);\ncat("Creating DMatrix for xgboost...\\n");\ndtrain <- xgb.DMatrix(data=trainM, label=train.y, missing = NaN);\nwatchlist <- list(trainM=dtrain);

\n\n

set.seed(1969);

\n\n

param <- list( objective = "reg:linear", \n booster = "gblinear",\n #eval_metric ="mae",\n eta = 0.035,\n max_depth = 6,\n subsample = 1.00,\n colsample_bytree = 0.40\n)

\n\n

clf <- xgb.cv( params = param, \n data = dtrain, \n nrounds = 200, \n verbose = 1,\n watchlist = watchlist,\n maximize = FALSE,\n nfold = 3,\n early.stop.round = 10,\n print.every.n = 1\n);

\n\n

bestRound <- which.min( as.matrix(clf)[,3] );\ncat("Best round:", bestRound,"\\n");\ncat("Best result:",min(as.matrix(clf)[,3]),"\\n");

\n\n

clf <- xgb.train( params = param, \n data = dtrain, \n nrounds = bestRound, \n verbose = 1,\n watchlist = watchlist,\n maximize = FALSE\n)\ntestM <-data.matrix(test, rownames.force = NA);

\n\n

preds <- predict(clf, testM);

", "rawMarkdown": "I am new to using XGBoost and trying to use a regression objective (then want to do poisson if I can ever get it working). \r\n\r\nI've attempted to use the following code and at the point of PREDICT I get the error:\r\n**Error in xgb.DMatrix(newdata) : \r\n REAL() can only be applied to a 'numeric', not a 'integer'**\r\n\r\nI've tried using lapply to the matrix, I've tried setting a variable to numeric, the training seems to move through, it's the prediction step that bombs. \r\n\r\n\r\n\r\n\r\ntrain$id <- NULL;\r\ntest$id <- NULL;\r\n\r\ncat(\"Treating features...\\n\");\r\n## impute train and test here\r\n#one hot encoding\r\nsparse_matrix <- sparse.model.matrix(Demanda_uni_equil ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = train)\r\nsparse_matrix <- sparse.model.matrix( ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = test)\r\ntrain$Cliente_ID<-as.numeric(train$Cliente_ID)\r\ntrain.y = train$Demanda_uni_equil;\r\n\r\ncat(\"Dropping features...\\n\");\r\ntrain$Demanda_uni_equil <- NULL;\r\n\r\n\r\ncat(\"Creating data.matrix...\\n\");\r\npreds <- rep(0,nrow(test));\r\ntrainM<-data.matrix(train, rownames.force = NA);\r\ncat(\"Creating DMatrix for xgboost...\\n\");\r\ndtrain <- xgb.DMatrix(data=trainM, label=train.y, missing = NaN);\r\nwatchlist <- list(trainM=dtrain);\r\n\r\nset.seed(1969);\r\n\r\nparam <- list( objective = \"reg:linear\", \r\n booster = \"gblinear\",\r\n #eval_metric =\"mae\",\r\n eta = 0.035,\r\n max_depth = 6,\r\n subsample = 1.00,\r\n colsample_bytree = 0.40\r\n)\r\n\r\nclf <- xgb.cv( params = param, \r\n data = dtrain, \r\n nrounds = 200, \r\n verbose = 1,\r\n watchlist = watchlist,\r\n maximize = FALSE,\r\n nfold = 3,\r\n early.stop.round = 10,\r\n print.every.n = 1\r\n);\r\n\r\nbestRound <- which.min( as.matrix(clf)[,3] );\r\ncat(\"Best round:\", bestRound,\"\\n\");\r\ncat(\"Best result:\",min(as.matrix(clf)[,3]),\"\\n\");\r\n\r\nclf <- xgb.train( params = param, \r\n data = dtrain, \r\n nrounds = bestRound, \r\n verbose = 1,\r\n watchlist = watchlist,\r\n maximize = FALSE\r\n)\r\ntestM <-data.matrix(test, rownames.force = NA);\r\n\r\n**preds <- predict(clf, testM);**", "votes": null }, { "id": "125702", "postDate": "07/01/2016 17:14:08", "content": "

I see that you've changed a column to numeric in your train set, but it doesn't look like you did it to your test set. Try that first and see if that helps.

", "rawMarkdown": "I see that you've changed a column to numeric in your train set, but it doesn't look like you did it to your test set. Try that first and see if that helps.", "votes": null }, { "id": "125727", "postDate": "07/01/2016 22:42:51", "content": "

Thanks much Branden. That allowed me to move through. Doesn't seem to matter which column I select as long as there is one.

", "rawMarkdown": "Thanks much Branden. That allowed me to move through. Doesn't seem to matter which column I select as long as there is one.", "votes": null } ], "comments": [ { "id": 125702, "author_name": "brandenkmurray", "author_url": "", "post_date": "07/01/2016 17:14:08", "content": "

I see that you've changed a column to numeric in your train set, but it doesn't look like you did it to your test set. Try that first and see if that helps.

", "votes": null, "replies": [] }, { "id": 125727, "author_name": "curuos", "author_url": "", "post_date": "07/01/2016 22:42:51", "content": "

Thanks much Branden. That allowed me to move through. Doesn't seem to matter which column I select as long as there is one.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "125701": "I am new to using XGBoost and trying to use a regression objective (then want to do poisson if I can ever get it working). \r\n\r\nI've attempted to use the following code and at the point of PREDICT I get the error:\r\n**Error in xgb.DMatrix(newdata) : \r\n REAL() can only be applied to a 'numeric', not a 'integer'**\r\n\r\nI've tried using lapply to the matrix, I've tried setting a variable to numeric, the training seems to move through, it's the prediction step that bombs. \r\n\r\n\r\n\r\n\r\ntrain$id <- NULL;\r\ntest$id <- NULL;\r\n\r\ncat(\"Treating features...\\n\");\r\n## impute train and test here\r\n#one hot encoding\r\nsparse_matrix <- sparse.model.matrix(Demanda_uni_equil ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = train)\r\nsparse_matrix <- sparse.model.matrix( ~ Agencia_ID+Canal_ID +Ruta_SAK +Producto_ID +Cliente_ID -1, data = test)\r\ntrain$Cliente_ID<-as.numeric(train$Cliente_ID)\r\ntrain.y = train$Demanda_uni_equil;\r\n\r\ncat(\"Dropping features...\\n\");\r\ntrain$Demanda_uni_equil <- NULL;\r\n\r\n\r\ncat(\"Creating data.matrix...\\n\");\r\npreds <- rep(0,nrow(test));\r\ntrainM<-data.matrix(train, rownames.force = NA);\r\ncat(\"Creating DMatrix for xgboost...\\n\");\r\ndtrain <- xgb.DMatrix(data=trainM, label=train.y, missing = NaN);\r\nwatchlist <- list(trainM=dtrain);\r\n\r\nset.seed(1969);\r\n\r\nparam <- list( objective = \"reg:linear\", \r\n booster = \"gblinear\",\r\n #eval_metric =\"mae\",\r\n eta = 0.035,\r\n max_depth = 6,\r\n subsample = 1.00,\r\n colsample_bytree = 0.40\r\n)\r\n\r\nclf <- xgb.cv( params = param, \r\n data = dtrain, \r\n nrounds = 200, \r\n verbose = 1,\r\n watchlist = watchlist,\r\n maximize = FALSE,\r\n nfold = 3,\r\n early.stop.round = 10,\r\n print.every.n = 1\r\n);\r\n\r\nbestRound <- which.min( as.matrix(clf)[,3] );\r\ncat(\"Best round:\", bestRound,\"\\n\");\r\ncat(\"Best result:\",min(as.matrix(clf)[,3]),\"\\n\");\r\n\r\nclf <- xgb.train( params = param, \r\n data = dtrain, \r\n nrounds = bestRound, \r\n verbose = 1,\r\n watchlist = watchlist,\r\n maximize = FALSE\r\n)\r\ntestM <-data.matrix(test, rownames.force = NA);\r\n\r\n**preds <- predict(clf, testM);**", "125702": "I see that you've changed a column to numeric in your train set, but it doesn't look like you did it to your test set. Try that first and see if that helps.", "125727": "Thanks much Branden. That allowed me to move through. Doesn't seem to matter which column I select as long as there is one." }, "source": "meta" }