Demonstrates very basic example of how to use a DLRM for making predictions using TorchRec capabilities. (pytorch#3043)

Summary:
Pull Request resolved: pytorch#3043

examples demonstrating how to use TorchRec for recommendation model predictions:

*   **DLRM (Deep Learning Recommendation Model)**: Implementation using TorchRec's EmbeddingBagCollection and KeyedJaggedTensor
*   **Two Tower Model**: Alternative architecture for recommendation systems

Both examples include complete implementations with training, evaluation, and prediction capabilities.

Differential Revision: D75989524

Author: yupadhyay

examples/prediction/__init__.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
+"""
+TorchRec recommendation model examples.
+
+This package contains examples of different recommendation models implemented using TorchRec:
+- DLRM (Deep Learning Recommendation Model)
+- Two-Tower Model
+
+Each model is organized in its own subdirectory with complete implementation, tests, and documentation.
+"""
+
+# Import main components from DLRM
+from torchrec.github.examples.prediction.dlrm.predict_using_dlrm import (
+    create_kjt_from_batch,
+    DLRMRatingWrapper,
+    RecommendationDataset as DLRMDataset,
+    TorchRecDLRM,
+)
+
+# Import main components from Two-Tower
+from torchrec.github.examples.prediction.twoTower.predict_using_twotower import (
+    create_kjt_from_ids,
+    RecommendationDataset as TwoTowerDataset,
+    TwoTowerModel,
+    TwoTowerRatingWrapper,
+)
+
+__all__ = [
+    # DLRM components
+    "DLRMRatingWrapper",
+    "DLRMDataset",
+    "TorchRecDLRM",
+    "create_kjt_from_batch",
+    # Two-Tower components
+    "TwoTowerModel",
+    "TwoTowerRatingWrapper",
+    "TwoTowerDataset",
+    "create_kjt_from_ids",
+]

examples/prediction/dlrm/README.md

@@ -0,0 +1,125 @@
+# DLRM Prediction Example
+
+This example demonstrates how to use a Deep Learning Recommendation Model (DLRM) for making predictions using TorchRec capabilities. The code includes:
+
+1. A DLRM implementation using TorchRec's EmbeddingBagCollection and KeyedJaggedTensor
+2. Training with random data
+3. Evaluation
+4. Making sample predictions
+
+## DLRM Architecture
+
+The Deep Learning Recommendation Model (DLRM) is a state-of-the-art architecture for recommendation systems that combines:
+
+- **Bottom MLP**: Processes dense features
+- **Embedding Tables**: Process sparse features
+- **Feature Interaction**: Computes dot products between embeddings and dense features
+- **Top MLP**: Processes the combined features to produce predictions
+
+This architecture is particularly effective for CTR prediction and ranking tasks in recommendation systems.
+
+## TorchRec Integration
+
+This implementation uses TorchRec's capabilities:
+- Uses `KeyedJaggedTensor` for sparse features
+- Uses `EmbeddingBagCollection` for embedding tables
+- Follows the DLRM architecture as described in the paper: https://arxiv.org/abs/1906.00091
+
+The example demonstrates how to leverage TorchRec's efficient sparse feature handling for recommendation models.
+
+## Dependencies
+
+Install the required dependencies:
+
+```bash
+# Install PyTorch
+pip install torch torchvision
+
+# Install NumPy
+pip install numpy
+
+# Install TorchRec
+pip install torchrec
+```
+
+**Important**: This implementation requires torchrec to run, as it uses TorchRec's specialized modules for recommendation systems.
+
+## Running the Example Locally
+
+1. Download the `predict_using_dlrm.py` file to your local machine.
+
+2. Run the example:
+
+```bash
+python3 predict_using_dlrm.py
+```
+
+3. If you're using a different Python environment:
+
+```bash
+# For conda environments
+conda activate your_environment_name
+python predict_using_dlrm.py
+
+# For virtual environments
+source your_venv/bin/activate
+python predict_using_dlrm.py
+```
+
+## What to Expect
+
+When you run the example, you'll see:
+
+1. Training progress for 10 epochs with loss and learning rate information
+2. Evaluation results showing MSE and RMSE metrics
+3. Sample predictions for a specific user on multiple items
+
+## Implementation Details
+
+This example uses TorchRec's capabilities to implement a DLRM model that:
+
+- Takes dense features and sparse features (as KeyedJaggedTensor) as input
+- Processes dense features through a bottom MLP
+- Processes sparse features through EmbeddingBagCollection
+- Computes feature interactions using dot products
+- Processes the interactions through a top MLP
+- Outputs rating predictions on a 0-5 scale
+
+The implementation demonstrates how to use TorchRec's specialized modules for recommendation systems, making it more efficient and scalable than a custom implementation.
+
+## Key TorchRec Components Used
+
+1. **KeyedJaggedTensor**: Efficiently represents sparse features with variable lengths
+2. **EmbeddingBagConfig**: Configures embedding tables with parameters like dimensions and feature names
+3. **EmbeddingBagCollection**: Manages multiple embedding tables for different categorical features
+
+## Comparison with Two-Tower Model
+
+While both DLRM and Two-Tower models are used for recommendation systems, they have different architectures and use cases:
+
+- **DLRM**: Combines multiple categorical features and dense features with feature interactions, suitable for CTR prediction and ranking tasks.
+- **Two-Tower**: Separates user and item processing into distinct towers, making it more suitable for retrieval tasks and large-scale recommendation systems.
+
+## Troubleshooting
+
+If you encounter any issues:
+
+1. **Python version**: This code has been tested with Python 3.8+. Make sure you're using a compatible version.
+
+2. **PyTorch and TorchRec installation**: If you have issues with PyTorch or TorchRec, try installing specific versions:
+   ```bash
+   pip install torch==2.0.0 torchvision==0.15.0
+   pip install torchrec==0.5.0
+   ```
+
+3. **Memory issues**: If you run out of memory, try reducing the batch size by modifying this line in the code:
+   ```python
+   batch_size = 256  # Try a smaller value like 64 or 32
+   ```
+
+4. **CPU vs GPU**: The code automatically uses CUDA if available. To force CPU usage, modify:
+   ```python
+   device = torch.device("cpu")
+   ```
+
+5. **TorchRec compatibility**: If you encounter compatibility issues with TorchRec, make sure you're using compatible versions of PyTorch and TorchRec.

examples/prediction/dlrm/__init__.py

@@ -0,0 +1,8 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict