Content

This project provide W&D recommendation model using criteo dataset in CTR prediction.

Available Datasets

You could download criteo dataset from Criteo Display Advertising Challenge, then unzip data and store dataset on local disk as ./data/train.txt

Available Models

Model	Reference
Wide&Deep	HT Cheng, et al. Wide & Deep Learning for Recommender Systems, 2016.

Steps to run

0. env

   pip install -r requirements.txt

1. train and evalute the fp32 model as pth

   bash run_model.sh

2. evalute the fp32 pytorch model and calculate the latency

   bash run_inference.sh

3. produce fp32 onnx using pth and calculate the latency

   bash run_to_onnx.sh

4. test fp32/fp16 onnx's using migraphx and calculate the latency

   bash run_migraphx_onnx.sh

5. test fp32/fp16 onnx's using onnxruntime and calculate the latency

   bash run_inference_onnx.sh
   bash run_inference_fp16_onnx.sh

Performance

We evaluate the model's auc:

dlrm model	auc(%)
Pytorch model	79.09
FP32 onnx model	79.09
FP16 onnx model	79.09