Getting Started

Basic usage

TorchCTR has a set of built-in datasets and models.

Quick example

From file examples/basic_usage.py

#!/usr/bin/env python
# encoding: utf-8

from torchctr.datasets import Titanic
from torchctr.models import FactorizationMachine
from torchctr.trainer import Trainer

# Load the well known Kaggle Titanic dataset
dataset = Titanic()
dataset.build_data()

# Now we build the famous Factorization Machine model
model = FactorizationMachine(feature_dims=dataset.feature_dims)

# Also we need to build a trainer for our model and dasaset
trainer = Trainer(model, dataset)

# At last we train it
trainer.train()

# save the model to local disk
trainer.save_model("checkpoints/test.pt")

The results should be:

| Warning | Didn't specify the func for dense field, so we will use default log
| Warning | Didn't specify the func for target column, so we will use raw data
| WARNING | embed_dim should be specified, otherwise we'll use default value 4
| building parameters ...
| building trainer ...
| Didn't find dashboard
| Start training ...
| Training 1/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 67.87it/s, auc=0.567, loss=0.946]
| Validating: 100%|██████████████████████████████████████████████████| 1/1 [00:00<00:00, 27.05it/s, auc=0.410, loss=1.53]
| Training 2/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 84.69it/s, auc=0.725, loss=0.703]
| Validating: 100%|██████████████████████████████████████████████████| 1/1 [00:00<00:00, 22.27it/s, auc=0.490, loss=1.11]
| Training 3/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 76.40it/s, auc=0.813, loss=0.521]
| Validating: 100%|█████████████████████████████████████████████████| 1/1 [00:00<00:00, 28.83it/s, auc=0.602, loss=0.873]
| Training 4/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 78.02it/s, auc=0.875, loss=0.412]
| Validating: 100%|█████████████████████████████████████████████████| 1/1 [00:00<00:00, 27.46it/s, auc=0.685, loss=0.766]
| Training 5/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 83.62it/s, auc=0.875, loss=0.388]
| Validating: 100%|██████████████████████████████████████████████████| 1/1 [00:00<00:00, 26.10it/s, auc=0.721, loss=0.73]
| Training 6/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 88.59it/s, auc=0.867, loss=0.397]
| Validating: 100%|█████████████████████████████████████████████████| 1/1 [00:00<00:00, 28.79it/s, auc=0.750, loss=0.694]
| Training 7/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 81.10it/s, auc=0.875, loss=0.386]
| Validating: 100%|█████████████████████████████████████████████████| 1/1 [00:00<00:00, 25.55it/s, auc=0.772, loss=0.638]
| Training 8/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 77.11it/s, auc=0.871, loss=0.371]
| Validating: 100%|██████████████████████████████████████████████████| 1/1 [00:00<00:00, 24.10it/s, auc=0.782, loss=0.59]
| Training 9/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 79.48it/s, auc=0.879, loss=0.368]
| Validating: 100%|█████████████████████████████████████████████████| 1/1 [00:00<00:00, 27.65it/s, auc=0.791, loss=0.562]
| Training 10/10: 100%|██████████████████████████████████████████████| 7/7 [00:00<00:00, 83.20it/s, auc=0.883, loss=0.37]
| Validating: 100%|█████████████████████████████████████████████████| 1/1 [00:00<00:00, 27.10it/s, auc=0.789, loss=0.552]
| Didn't find dir, so we will create it

the method will offer to download the Titanic data if it has not already been downloaded, and it’ll be saved in .torchctr folder in your home directory.

Build model with custom trainer hyper-parameters

Basicly there are two hyper-parameters setting we can customirize, the first one is model hyper-parameter, such as the vector dimension in Factorization Machine model, the neural network’s hidden dimension in Google Wide & Deep model, which shoule be set with model’s introduction. The second one is the trainer hyper-parameter, which could be set with your own environment.

hyper_parameters = {
   "batch_size": 128,
   "device": "cpu",
   "learning_rate": 0.01,
   "weight_decay": 1e-6,
   "epochs": 30,
   "metrics": ["auc", "acc"]
   }