This repository contains the deep learning model which achieved the 3rd position in the M5 Accuracy competition. This is the same model that submitted the results to the competition (Last model update: 2020-05-31).
This model is developed by mofl
Python 3.7.4
CUDA 10.1
CUDNN 7.6.5
nvidia drivers 435.21
PyTorch 1.4
* Initialize docker setup for the GPU (pytorch/pytorch:1.4-cuda10.1-cudnn7-devel)
* run sh ./install.sh
1. make data directory to [DATA_PATH], i.e. /data/m5
2. copy 4 files (calendar.csv, sales_train_evaluation.csv, sample_submission.csv, sell_prices.csv) to [DATA_PATH]
The train/predict code will also call this script if it has not already been run on the relevant data. This takes about 10 minutes.
python ./examples/m5/preprocessing.py --data_path [DATA_PATH]
a) train 8 models (can be excuted parallelly using multiple gpus). One model takes about 6 hour to train
python examples/m5/training.py --trial t1 --data_path [DATA_PATH]
python examples/m5/training.py --trial t2 --data_path [DATA_PATH]
python examples/m5/training.py --trial t3 --data_path [DATA_PATH]
python examples/m5/training.py --trial t4 --data_path [DATA_PATH]
python examples/m5/training.py --trial t5 --data_path [DATA_PATH]
python examples/m5/training.py --trial t6 --data_path [DATA_PATH]
python examples/m5/training.py --trial t7 --data_path [DATA_PATH]
python examples/m5/training.py --trial t8 --data_path [DATA_PATH]
b) make 14 predictions of 10 epochs per each trained models (can be excuted parallelly using multiple gpus). One model takes about 1 hour to generate predictions. If gpu memory is not enough, adjust batch size with --bs option. e.g. --bs 640 (default: 6400)
python examples/m5/make_predictions.py --trial t1 --data_path [DATA_PATH]
python examples/m5/make_predictions.py --trial t2 --data_path [DATA_PATH]
python examples/m5/make_predictions.py --trial t3 --data_path [DATA_PATH]
python examples/m5/make_predictions.py --trial t4 --data_path [DATA_PATH]
python examples/m5/make_predictions.py --trial t5 --data_path [DATA_PATH]
python examples/m5/make_predictions.py --trial t6 --data_path [DATA_PATH]
python examples/m5/make_predictions.py --trial t7 --data_path [DATA_PATH]
python examples/m5/make_predictions.py --trial t8 --data_path [DATA_PATH]
c) make final ensemble model (this takes less than 5 minutes). Final csv file is ./logs/m5_submission/rolled_deepAR/drop1/submission_final.csv
python examples/m5/ensemble.py --data_path [DATA_PATH]
| File | Description | Entry points |
|---|---|---|
| ./examples/m5/accuracy_evaluator.py | caulating wrmsse | |
| ./examples/m5/ensemble.py | making final prediction | v |
| ./examples/m5/load_dataset.py | generating features and dataset | |
| ./examples/m5/make_predictions.py | make predictions from saved model | v |
| ./examples/m5/preprocessing.py | preprocessing raw data | v |
| ./examples/m5/training.py | training model from scratch | v |
| ./examples/m5/utils.py | auxiliary code | |
| ./pts | code for backbone engine | |
| ./test | test script for backbone engine |
| File | Description |
|---|---|
| ./logs/m5_submission | root path of logs |
| ./logs/m5_submission/rolled_deepAR/drop1 | root path of rolled deepAR model without dropout |
| ./logs/m5_submission/rolled_deepAR/drop1/t1~t8 | logs of each trials |
| ./logs/m5_submission/rolled_deepAR/drop1/tx/CV/train_net_xx | 14 predictions of xx epoch |
| ./logs/m5_submission/rolled_deepAR/drop1/tx/predictor | saved predictor configurations |
| ./logs/m5_submission/rolled_deepAR/drop1/tx/trained_model | saved model parameter per each epoch |
| ./logs/m5_submission/rolled_deepAR/drop1/tx/events.out... | event file for tensorboard |
| ./logs/m5_submission/rolled_deepAR/drop1/tx/log.txt | txt log of training |
| ./logs/m5_submission/rolled_deepAR/drop1/ensemble.log | txt log of ensemble |
| ./logs/m5_submission/rolled_deepAR/drop1/submission_final.csv | final prediction result |
This model is developed based on PyTorchTS.