[WIP] generating ranking data and using transform

src/common/metrics.py

@@ -48,7 +48,7 @@ def open(self):
     def close(self):
         """Close the MLFlow session."""
         if MetricsLogger._initialized:
-            self._logger.info(f"Finalizing MLFLOW [session='{self._session_name}']")
+            self._logger.info(f"Finalizing MLFLOW [session='{self._session_name}', session_id='{mlflow.active_run().info.run_id}'']")
             mlflow.end_run()
             MetricsLogger._initialized = False
         else:

src/scripts/data_processing/generate_data/generate.py

@@ -52,7 +52,7 @@ def get_arg_parser(cls, parser=None):
         # add arguments that are specific to the script
         group_params = parser.add_argument_group("Synthesis params")
         group_params.add_argument(
-            "--type", required=True, type=str, choices=["classification", "regression"]
+            "--type", required=True, type=str, choices=["classification", "regression", "lambdarank"]
         )
         group_params.add_argument("--train_samples", required=True, type=int)
         group_params.add_argument("--test_samples", required=True, type=int)
@@ -61,6 +61,8 @@ def get_arg_parser(cls, parser=None):
         group_params.add_argument("--n_informative", required=True, type=int)
         group_params.add_argument("--n_redundant", required=False, type=int)
         group_params.add_argument("--random_state", required=False, default=None, type=int)
+        group_params.add_argument("--docs_per_query", required=False, default=20, type=int)
+        group_params.add_argument("--delimiter", required=False, default=',', type=str)
 
         group_o = parser.add_argument_group("Outputs")
         group_o.add_argument(
@@ -81,6 +83,12 @@ def get_arg_parser(cls, parser=None):
             type=str,
             help="Output data location (directory)",
         )
+        group_o.add_argument(
+            "--output_header",
+            required=True,
+            type=str,
+            help="Output header location (directory)",
+        )
 
         return parser
 
@@ -98,6 +106,7 @@ def run(self, args, logger, metrics_logger, unknown_args):
         os.makedirs(args.output_train, exist_ok=True)
         os.makedirs(args.output_test, exist_ok=True)
         os.makedirs(args.output_inference, exist_ok=True)
+        os.makedirs(args.output_header, exist_ok=True)
 
         metrics_logger.log_parameters(
             type=args.type,
@@ -131,6 +140,19 @@ def run(self, args, logger, metrics_logger, unknown_args):
                     n_informative=args.n_informative,
                     random_state=args.random_state,
                 )
+            elif args.type == "lambdarank":
+                X, y = make_regression(
+                    n_samples=total_samples,
+                    n_features=args.n_features,
+                    n_informative=args.n_informative,
+                    random_state=args.random_state,
+                )
+                # add query column
+                query_col = [[i // args.docs_per_query] for i in range(total_samples)]
+                X = numpy.hstack((query_col, X))
+                # create 30 ranking labels
+                y = ((y - min(y))/(max(y)-min(y))*30).astype(int)
+
             else:
                 raise NotImplementedError(f"--type {args.type} is not implemented.")
 
@@ -150,30 +172,36 @@ def run(self, args, logger, metrics_logger, unknown_args):
             inference_data = X[args.train_samples + args.test_samples :]
             logger.info(f"Inference data shape: {inference_data.shape}")
 
+        # create a header 
+        header = [f'Column_{i}' for i in range(train_data.shape[1])]
+        if args.delimiter == 'tab':
+            args.delimiter = "\t"
         # save as CSV
         logger.info(f"Saving data...")
         with metrics_logger.log_time_block("time_data_saving"):
             numpy.savetxt(
                 os.path.join(args.output_train, "train.txt"),
                 train_data,
-                delimiter=",",
+                delimiter=args.delimiter,
                 newline="\n",
                 fmt="%1.3f",
             )
             numpy.savetxt(
                 os.path.join(args.output_test, "test.txt"),
                 test_data,
-                delimiter=",",
+                delimiter=args.delimiter,
                 newline="\n",
                 fmt="%1.3f",
             )
             numpy.savetxt(
                 os.path.join(args.output_inference, "inference.txt"),
                 inference_data,
-                delimiter=",",
+                delimiter=args.delimiter,
                 newline="\n",
                 fmt="%1.3f",
             )
+            with open(os.path.join(args.output_header, "header.txt"), 'w') as hf:
+                hf.writelines(args.delimiter.join(header))
 
 
 def get_arg_parser(parser=None):

src/scripts/data_processing/generate_data/spec.yaml

@@ -13,6 +13,7 @@ inputs:
     enum:
       - regression
       - classification
+      - lambdarank
   train_samples:
     type: Integer
     description: Number of training samples
@@ -41,6 +42,15 @@ inputs:
     type: Integer
     description: random seed
     optional: true
+  delimiter:
+    type: String
+    description: selimiter
+    optional: true
+    default: ","
+  docs_per_query:
+    type: Integer
+    description: random seed
+    optional: true
   verbose:
     type: Boolean
     default: False
@@ -56,6 +66,8 @@ outputs:
     type: AnyDirectory
   output_inference:
     type: AnyDirectory
+  output_header:
+    type: AnyDirectory
 
 command: >-
   python generate.py
@@ -67,11 +79,14 @@ command: >-
   --n_informative {inputs.n_informative}
   [--n_redundant {inputs.n_redundant}]
   [--random_state {inputs.random_state}]
+  [--docs_per_query {inputs.docs_per_query}]
   --output_train {outputs.output_train}
   --output_test {outputs.output_test}
   --output_inference {outputs.output_inference}
+  --output_header {outputs.output_header}
   --verbose {inputs.verbose}
   [--custom_properties {inputs.custom_properties}]
+  [--delimiter {inputs.delimiter}]
 
 environment:
   conda:

src/scripts/training/lightgbm_python/default.dockerfile

@@ -1,10 +1,29 @@
-FROM mcr.microsoft.com/azureml/openmpi3.1.2-ubuntu18.04:20210615.v1
-LABEL lightgbmbenchmark.linux.cpu.mpi.pip.version="3.3.0/20211111.1"
+FROM mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:latest
+LABEL lightgbmbenchmark.linux.cpu.mpi.pip.version="3.3.1/20211210.1"
+
+# https://github.com/microsoft/lightgbm-transform/blob/main/docs/Installation-Guide.rst
+# Install CMake, gcc, g++, boost.
+RUN apt-get update && apt-get -y upgrade && DEBIAN_FRONTEND="noninteractive" apt-get install -y libboost-all-dev gcc g++ wget cmake git curl libtinfo5
+
+# Install LLVM with RTTI feature.
+WORKDIR /root
+RUN wget https://github.com/llvm/llvm-project/archive/refs/tags/llvmorg-3.5.1.tar.gz && tar zxvf llvmorg-3.5.1.tar.gz
+WORKDIR /root/llvm-project-llvmorg-3.5.1/llvm
+RUN mkdir build && cd build && cmake -DLLVM_REQUIRES_RTTI=1 .. && make -j4 && make install
+
+# Install bond.
+WORKDIR /root
+RUN git clone --recursive https://github.com/microsoft/bond.git
+RUN DEBIAN_FRONTEND="noninteractive" apt-get install -y clang zlib1g-dev
+RUN curl -sSL https://get.haskellstack.org/ | sh
+WORKDIR /root/bond
+RUN mkdir build && cd build && cmake -DBOND_ENABLE_GRPC=FALSE .. && make -j4 && make install
+
 
 # Those arguments will NOT be used by AzureML
 # they are here just to allow for lightgbm-benchmark build to actually check
 # dockerfiles in a PR against their actual branch
-ARG lightgbm_version="3.3.0"
+ARG lightgbm_version="3.3.1"
 
 ENV AZUREML_CONDA_ENVIRONMENT_PATH /azureml-envs/lightgbm
 
@@ -26,12 +45,14 @@ RUN HOROVOD_WITH_TENSORFLOW=1 \
                 'azureml-defaults==1.35.0' \
                 'azureml-mlflow==1.35.0' \
                 'azureml-telemetry==1.35.0' \
-                'mpi4py==3.1.1'
+                'mpi4py==3.1.1' \
+                'omegaconf'
 
 # install lightgbm with mpi
 RUN pip install --upgrade pip setuptools wheel && \
     pip install 'cmake==3.21.0' && \
-    pip install lightgbm==${lightgbm_version} --install-option=--mpi
+    pip install lightgbm==${lightgbm_version} --install-option=--mpi &&\
+    pip install lightgbm-transform
 
 # This is needed for mpi to locate libpython
 ENV LD_LIBRARY_PATH $AZUREML_CONDA_ENVIRONMENT_PATH/lib:$LD_LIBRARY_PATH

src/scripts/training/lightgbm_python/spec.yaml

@@ -16,6 +16,10 @@ inputs:
     type: AnyDirectory
     description: directory to the testing data
     optional: false
+  parser_config_file:
+    type: AnyDirectory
+    description: directory to the transform parser config
+    optional: true
   construct:
     type: Boolean
     description: "Use lazy intialization during data loading phase, see https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.Dataset.html#lightgbm.Dataset.construct"
@@ -144,6 +148,7 @@ launcher:
     python train.py
     --train {inputs.train}
     --test {inputs.test}
+    [--parser_config_file {inputs.parser_config_file}]
     --construct {inputs.construct}
     --header {inputs.header}
     [--label_column {inputs.label_column}]

src/scripts/training/lightgbm_python/train.py

@@ -62,6 +62,7 @@ def get_arg_parser(cls, parser=None):
         group_i.add_argument("--header", required=False, default=False, type=strtobool)
         group_i.add_argument("--label_column", required=False, default="0", type=str)
         group_i.add_argument("--group_column", required=False, default=None, type=str)
+        group_i.add_argument("--parser_config_file", required=False, default=None, type=str, help="transfomr parser config location (file path)")
 
         group_o = parser.add_argument_group("Outputs")
         group_o.add_argument("--export_model",