Question

首先，我有数据（hex_glrm），其中包含适合模型的4938列（要素）。在拟合之后，我下载并以MOJO格式将模型保存到文件夹tmp。

script.R

...
model_glrm <- tryCatch({
  h2o.glrm(hex_glrm,
           cols = names(hex_glrm),
           model_id = "model_glrm",
           transform = "STANDARDIZE",
           loss = "Huber",
           regularization_x = "NonNegative",
           regularization_y = "UnitOneSparse",
           k = 100,
           seed = 2017,
           gamma_x = 0.2,
           gamma_y = 0.2,
           init = "Random",
           svd_method = "GramSVD",
           impute_original = TRUE,
           max_runtime_secs = 1200)
}, error = function(err) {
  flog.error(err)
  NULL
})

if (!is.null(model_glrm)) {
  flog.info(" - GLRM model is fitted!")

  # Download generated MOJO
  h2o.download_mojo(model_glrm, path = "/tmp", get_genmodel_jar = TRUE)
}
...

然后我创建EasyPredictModelWrapper并加载model_glrm.zip。之后，我获得了用于预测的新用户数据，并将其填入RowData。然后我得到预测（drmp）并创建响应的json。

PredictionServlet.java

...
// Get the MOJO model and the prediction
EasyPredictModelWrapper model = new EasyPredictModelWrapper(MojoModel.load("tmp/model_glrm.zip"));
RowData row = getRowDataFromJson(json); // New data for prediction (4938 features)
DimReductionModelPrediction drmp = model.predictDimReduction(row);

// Create the json response 
JsonObject json = new JsonObject();       
StringBuilder builder = new StringBuilder();
for (double i : drmp.dimensions) {
      builder.append(i);
      builder.append(", ");
}
json.addProperty("recommendations", builder.toString());
json.addProperty("size", drmp.dimensions.length);    

// Emit the prediction to the servlet response
response.getWriter().write(json.toString());
response.setContentType("application/json; charset=UTF-8");
response.setHeader("Access-Control-Allow-Origin", "*");
response.setStatus(HttpServletResponse.SC_OK);
...

response.json

{ ＆＃34;推荐＆＃34;：＆＃34; 1.2323,2.3343，...＆＃34;，＆＃34;大小＆＃34;：100 }

但是当我得到预测时drmp.dimensions只包含100个值。如何使用所有重建特征进行预测（4938）？

如何使用MOJO（GLRM）的所有重建特征进行预测？

0 个答案: