[MLPerf] Add DLRM-DCNv2

examples/ml_perf/configs/datasets/dummy_dataset.py

@@ -0,0 +1,166 @@
+from keras.utils import Config
+
+# === Dataset ===
+dataset_config = Config()
+dataset_config.file_pattern = None
+# Features
+dataset_config.label = "clicked"
+dataset_config.dense = [f"int-feature-{i}" for i in range(1, 14)]
+dataset_config.lookup = [
+    {
+        "name": "categorical-feature-14",
+        "vocabulary_size": 40000000,
+        "feature_list_length": 3,
+        "new_name": "cat_14",
+    },
+    {
+        "name": "categorical-feature-15",
+        "vocabulary_size": 39060,
+        "feature_list_length": 2,
+        "new_name": "cat_15",
+    },
+    {
+        "name": "categorical-feature-16",
+        "vocabulary_size": 17295,
+        "feature_list_length": 1,
+        "new_name": "cat_16",
+    },
+    {
+        "name": "categorical-feature-17",
+        "vocabulary_size": 7424,
+        "feature_list_length": 2,
+        "new_name": "cat_17",
+    },
+    {
+        "name": "categorical-feature-18",
+        "vocabulary_size": 20265,
+        "feature_list_length": 6,
+        "new_name": "cat_18",
+    },
+    {
+        "name": "categorical-feature-19",
+        "vocabulary_size": 3,
+        "feature_list_length": 1,
+        "new_name": "cat_19",
+    },
+    {
+        "name": "categorical-feature-20",
+        "vocabulary_size": 7122,
+        "feature_list_length": 1,
+        "new_name": "cat_20",
+    },
+    {
+        "name": "categorical-feature-21",
+        "vocabulary_size": 1543,
+        "feature_list_length": 1,
+        "new_name": "cat_21",
+    },
+    {
+        "name": "categorical-feature-22",
+        "vocabulary_size": 63,
+        "feature_list_length": 1,
+        "new_name": "cat_22",
+    },
+    {
+        "name": "categorical-feature-23",
+        "vocabulary_size": 40000000,
+        "feature_list_length": 7,
+        "new_name": "cat_23",
+    },
+    {
+        "name": "categorical-feature-24",
+        "vocabulary_size": 3067956,
+        "feature_list_length": 3,
+        "new_name": "cat_24",
+    },
+    {
+        "name": "categorical-feature-25",
+        "vocabulary_size": 405282,
+        "feature_list_length": 8,
+        "new_name": "cat_25",
+    },
+    {
+        "name": "categorical-feature-26",
+        "vocabulary_size": 10,
+        "feature_list_length": 1,
+        "new_name": "cat_26",
+    },
+    {
+        "name": "categorical-feature-27",
+        "vocabulary_size": 2209,
+        "feature_list_length": 6,
+        "new_name": "cat_27",
+    },
+    {
+        "name": "categorical-feature-28",
+        "vocabulary_size": 11938,
+        "feature_list_length": 9,
+        "new_name": "cat_28",
+    },
+    {
+        "name": "categorical-feature-29",
+        "vocabulary_size": 155,
+        "feature_list_length": 5,
+        "new_name": "cat_29",
+    },
+    {
+        "name": "categorical-feature-30",
+        "vocabulary_size": 4,
+        "feature_list_length": 1,
+        "new_name": "cat_30",
+    },
+    {
+        "name": "categorical-feature-31",
+        "vocabulary_size": 976,
+        "feature_list_length": 1,
+        "new_name": "cat_31",
+    },
+    {
+        "name": "categorical-feature-32",
+        "vocabulary_size": 14,
+        "feature_list_length": 1,
+        "new_name": "cat_32",
+    },
+    {
+        "name": "categorical-feature-33",
+        "vocabulary_size": 40000000,
+        "feature_list_length": 12,
+        "new_name": "cat_33",
+    },
+    {
+        "name": "categorical-feature-34",
+        "vocabulary_size": 40000000,
+        "feature_list_length": 100,
+        "new_name": "cat_34",
+    },
+    {
+        "name": "categorical-feature-35",
+        "vocabulary_size": 40000000,
+        "feature_list_length": 27,
+        "new_name": "cat_35",
+    },
+    {
+        "name": "categorical-feature-36",
+        "vocabulary_size": 590152,
+        "feature_list_length": 10,
+        "new_name": "cat_36",
+    },
+    {
+        "name": "categorical-feature-37",
+        "vocabulary_size": 12973,
+        "feature_list_length": 3,
+        "new_name": "cat_37",
+    },
+    {
+        "name": "categorical-feature-38",
+        "vocabulary_size": 108,
+        "feature_list_length": 1,
+        "new_name": "cat_38",
+    },
+    {
+        "name": "categorical-feature-39",
+        "vocabulary_size": 36,
+        "feature_list_length": 1,
+        "new_name": "cat_39",
+    },
+]

examples/ml_perf/configs/models/default_model.py

@@ -0,0 +1,19 @@
+from keras.utils import Config
+
+# === Model ===
+model_config = Config()
+# Embedding
+model_config.embedding_dim = 128
+model_config.allow_id_dropping = True
+model_config.embedding_threshold = 21000
+model_config.max_ids_per_partition = 4096
+model_config.max_unique_ids_per_partition = 2048
+model_config.learning_rate = 0.005
+
+# MLP
+model_config.bottom_mlp_dims = [512, 256, 128]
+model_config.top_mlp_dims = [1024, 1024, 512, 256, 1]
+
+# DCN
+model_config.num_dcn_layers = 3
+model_config.dcn_projection_dim = 512

examples/ml_perf/configs/training/default_training.py

@@ -0,0 +1,7 @@
+from keras.utils import Config
+
+# === Training Hyperparameters ===
+training_config = Config()
+training_config.learning_rate = 0.005
+training_config.global_batch_size = 128
+training_config.num_epochs = 1

examples/ml_perf/configs/v6e_16.py

@@ -0,0 +1,16 @@
+from keras.utils import Config
+
+from .datasets.dummy_dataset import dataset_config
+from .models.default_model import model_config
+from .training.default_training import training_config
+
+config = Config()
+
+config.experiment_name = "v6e_16"
+config.model_dir = "./v6e_16"
+
+config.dataset = dataset_config
+config.model = model_config
+config.training = training_config
+
+config.freeze()

examples/ml_perf/configs/v6e_8.py

@@ -0,0 +1,16 @@
+from keras.utils import Config
+
+from .datasets.dummy_dataset import dataset_config
+from .models.default_model import model_config
+from .training.default_training import training_config
+
+config = Config()
+
+config.experiment_name = "v6e_8"
+config.model_dir = "./v6e_8"
+
+config.dataset = dataset_config
+config.model = model_config
+config.training = training_config
+
+config.freeze()

examples/ml_perf/configs/v6e_8_full_dataset.py

@@ -0,0 +1,27 @@
+from keras.utils import Config
+
+from .datasets.dummy_dataset import dataset_config
+from .models.default_model import model_config
+from .training.default_training import training_config
+
+config = Config()
+
+config.experiment_name = "v6e_8_full_dataset"
+config.model_dir = "./v6e_8_full_dataset"
+
+config.dataset = dataset_config
+config.dataset.file_pattern = (
+    "gs://qinyiyan-vm/mlperf-dataset/criteo_merge_balanced_4224/"
+    "train-00000-of-01024tfrecord"
+)
+config.dataset.val_file_pattern = (
+    "gs://qinyiyan-vm/mlperf-dataset/criteo_merge_balanced_4224/"
+    "train-00000-of-01024tfrecord"
+)
+# The path which we are reading from already has the batched dataset.
+config.dataset.file_batch_size = 4224
+config.model = model_config
+config.training = training_config
+config.training.batch_size = 256
+
+config.freeze()