Document adding cuML and fix the gnn_fraud_detection_pipeline example

docker/conda/environments/cuda11.8_dev.yml

@@ -57,6 +57,10 @@ dependencies:
     - gxx_linux-64=11.2
     - include-what-you-use=0.18
     - isort
+    # If cuml is installed, we need to constrain these two libraries. See:
+    # https://docs.rapids.ai/notices/rsn0026/ for libcusolver & libcusparse
+    - libcusolver<=11.4.1.48
+    - libcusparse<12
     - librdkafka=1.7.0
     - mlflow>1.29,<2
     - mrc=23.03
@@ -71,7 +75,7 @@ dependencies:
     - pip
     - pkg-config # for mrc cmake
     - pluggy=1.0
-    - protobuf=3.20.2
+    - protobuf=3.20.3
     - pybind11-stubgen=0.10.5
     - pydot
     - pytest

docker/conda/environments/requirements.txt

@@ -9,7 +9,7 @@ jupyter-core>=4.11.2,<5.0
 jupyterlab
 nvidia-pyindex
 # Duplicated in conda dev to ensure parity with libprotobuf
-protobuf==3.20.2
+protobuf==3.20.3
 pytest-benchmark>=4.0
 torch==1.13.1+cu116
 tritonclient[all]==2.17.*

examples/gnn_fraud_detection_pipeline/README.md

@@ -34,7 +34,10 @@ export MORPHEUS_ROOT=$(pwd)
 Use Morpheus to run the GNN fraud detection Pipeline with the transaction data. A pipeline has been configured in `run.py` with several command line options:
 
 ```bash
+cd ${MORPHEUS_ROOT}/examples/gnn_fraud_detection_pipeline
 python run.py --help
+```
+```
 Usage: run.py [OPTIONS]
 
 Options:
@@ -61,8 +64,10 @@ Options:
 To launch the configured Morpheus pipeline with the sample data that is provided at `$MORPHEUS_ROOT/models/dataset`, run the following:
 
 ```bash
-
+cd ${MORPHEUS_ROOT}/examples/gnn_fraud_detection_pipeline
 python run.py
+```
+```
 ====Building Pipeline====
 Added source: <from-file-0; FileSourceStage(filename=validation.csv, iterative=None, file_type=auto, repeat=1, filter_null=False)>
   └─> morpheus.MessageMeta

examples/gnn_fraud_detection_pipeline/requirements.yml

@@ -16,11 +16,16 @@
 channels:
     - rapidsai
     - nvidia
-    - nvidia/label/dev # For pre-releases of MRC. Should still default to full releases if available
-    - stellargraph
     - conda-forge
 dependencies:
     - chardet=5.0.0
-    - cuml=22.08
-    - stellargraph=1.2.1
-    - tensorflow=2.9.1
+    - cuml=22.10
+    # https://docs.rapids.ai/notices/rsn0026/ for libcusolver & libcusparse
+    - libcusolver<=11.4.1.48
+    - libcusparse<12
+    - cuda-python<=11.7.0 # Remove when Issue #251 is closed
+    - pip:
+        - stellargraph==1.2.1
+        - tensorflow==2.12.0
+        # Duplicated in conda dev to ensure parity with libprotobuf
+        - protobuf==3.20.3

examples/gnn_fraud_detection_pipeline/stages/graph_sage_stage.py

@@ -17,6 +17,7 @@
 import typing
 
 import mrc
+import numpy as np
 
 import cudf
 
@@ -95,7 +96,7 @@ def _inductive_step_hinsage(
         generator = HinSAGENodeGenerator(graph, self._batch_size, self._sample_size, head_node_type=self._target_node)
         test_gen_not_shuffled = generator.flow(node_identifiers, shuffle=False)
 
-        inductive_emb = trained_model.predict(test_gen_not_shuffled)
+        inductive_emb = np.concatenate([trained_model.predict(row[0]) for row in test_gen_not_shuffled])
         inductive_emb = cudf.DataFrame(inductive_emb, index=node_identifiers)
 
         return inductive_emb