dnn-regression

dnn-regression

This sample project creates and fits a TensorFlow DNNRegressor model based on parquet-formatted input data. Then, the application exports the model to a local file and logs the model using MLflow’s APIs. The arguments to the program are as follows:

• model-dir: (str, required) local path or URI of a directory to which the DNNRegressor’s checkpoints and the final exported SavedModel will be written.
• train: (str, required) local path or URI of a parquet filecontaining training data
• test: (str, required) local path or URI of a parquet file containing test data
• hidden-units: (str, required) size and number of layers for the dnn; string input which has layers delimited by commas (i.e. “10,10” for two layers of 10 nodes each)
• steps: (int) steps to be run while training the regressor; default 100
• batch-size: (int) batch size used for creation of input functions for training and evaluation; default 128
• label-col: (str, required) name of label column in dataset
• feat-cols: (str) names of columns in dataset to be used as features;input is one string with names delimited by commas. If no argument is provided, it is assumed that all columns but the label column are feature columns.

This app currently assumes that the input data is all numerical. Column names must adhere to TensorFlow constraints.

To run the app with default parameters while in the root directory, run the command

mlflow run apps/dnn-regression/ -e main -P model-dir="insert/model/save/path" -P train="insert/data/path/" -P test="insert/data/path/" -P hidden-units="10,10" -P label-col="insert.label.col"