FACTS is an efficient and fully automated CTS forecasting framework that can find an ST-block that offers optimal accuracy on arbitrary unseen tasks and can make forecasts in minutes.
We provide the source code of FACTS in this repository.
We are further optimizing our code and welcome any suggestions for modifications.
You can obtained the well pre-processed datasets from DropBox.Then place the downloaded data under the folder ./dataset.Click on the link below to access the experimental datasets in the paper.
We introduce some main files in our project as follows:
-
NAS_Net/The main code of FACTS.-
ArchPredictorThe main structure of TAP. -
operations.pyThe operators used in FACTS' search space. -
st_net.pyThe main backbone of FACTS. -
operations_inherit.pyst_net_1.pyThe main methodology of weights inheritance.
-
-
exps/Experimental scripts.-
collect_seeds.pyThe process of iteratively pruning the search space, sampling the ST-blocks and training them. -
generate_seeds.pyThe script for training the optimal ST-blocks. -
dataset_slice.pyThe script for creating CTS forecasting tasks. -
Generate_task_feature.pyGenerating the semantic features for tasks. -
Generate_statistical_feature.pyGenerating the statistical features for tasks.
-
-
scripts/automated testing scripts -
utils.pySome utility functions such like dataloader, normalization.
Given a python environment (note: this project is fully tested under python 3.8), install the dependencies with the following command:
pip install -r requirements.txtWe provide all the scripts to repoduce the main experimental results in our paper:
As illustrated in FACTS, we first iteratively prune the large search space through a GBM-based pruner, while the pruning process is coupled with the collection of seeds. We provide the correspongding scripts as belows:
First, we create 100 tasks from 11 datasets equipped with two forecasting settings for pretraining, this process is implemented by a heuristic slicing method. With the random seed provided in the code, we can reproduce the process and generate subsets in a specified dir ./subsets , we also show the task settings in exps/Task_config.py to ensure the reproducibility of pretraining tasks:
# Create tasks for pretraining
python ./exps/dataset_slice.pyThen we generate statistical features for all subsets with the tsfresh package:
# pip install tsfresh
# Generate statistical features for all subsets
bash ./scripts/generate_statistical_feature_for_subsets.shThe semantic features are generated by representation learning model TS2Vec, we inregrate it into our code:
# Generate semantic features for all subsets
bash ./scripts/generate_task_feature_for_subsets.shAfter running the scripts above, we create all the static information for tasks, we then iteratively prune the search space and collect the seeds in each phase. Specifically, we can prune the search space with the script below, and there are some hyperparameters needed to be changed in the script to keep a half-cut:
# iteratively prune the search space
python ./exps/iteratively_search_space_pruning.py --mode iterativelyThen we collect the seeds in the pruned search space by parallel running on 32 A800 gpus. We provide the scripts used in our machine. One can edit their own gpu number in the scripts:
# Collect seeds
bash ./scripts/pretrain_0.sh &
bash ./scripts/pretrain_1.sh &
bash ./scripts/pretrain_2.sh &
bash ./scripts/pretrain_3.sh &
bash ./scripts/pretrain_4.sh &
bash ./scripts/pretrain_5.sh &
bash ./scripts/pretrain_6.sh &By looping above two processes 7 times, one can prune the search space to a minimal one as illustrated in our paper. Meanwhile, a large number of seeds are collected in the whole process and stored in the dir ./seeds. With the huge amounts of seeds and the static information (statistical features and semantic features) of each task, we can pretrain the TAP with the scripts below:
# Pretrain the TAP
python ./exps/rank_the_optimal_subspace.py --mode trainThen we can use the pretrained TAP to rank the optimal subspace pruned above and find an optimal architecture for each task:
# Search the optimal ST-block (template)
python ./exps/rank_the_optimal_subspace.py --mode search --dataset <dataset_name> --seq_len <the windwow size> --pred_len <the horizon size>
# Specific
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMS-BAY --seq_len 12 --pred_len 12
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMS-BAY --seq_len 24 --pred_len 24
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMS-BAY --seq_len 48 --pred_len 48
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMS-BAY --seq_len 168 --pred_len 1
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Electricity --seq_len 12 --pred_len 12
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Electricity --seq_len 24 --pred_len 24
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Electricity --seq_len 48 --pred_len 48
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Electricity --seq_len 168 --pred_len 1
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMSD7M --seq_len 12 --pred_len 12
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMSD7M --seq_len 24 --pred_len 24
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMSD7M --seq_len 48 --pred_len 48
python ./exps/rank_the_optimal_subspace.py --mode search --dataset PEMSD7M --seq_len 168 --pred_len 1
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-TAXI --seq_len 12 --pred_len 12
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-TAXI --seq_len 24 --pred_len 24
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-TAXI --seq_len 48 --pred_len 48
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-TAXI --seq_len 168 --pred_len 1
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-BIKE --seq_len 12 --pred_len 12
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-BIKE --seq_len 24 --pred_len 24
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-BIKE --seq_len 48 --pred_len 48
python ./exps/rank_the_optimal_subspace.py --mode search --dataset NYC-BIKE --seq_len 168 --pred_len 1
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Los-Loop --seq_len 12 --pred_len 12
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Los-Loop --seq_len 24 --pred_len 24
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Los-Loop --seq_len 48 --pred_len 48
python ./exps/rank_the_optimal_subspace.py --mode search --dataset Los-Loop --seq_len 168 --pred_len 1
python ./exps/rank_the_optimal_subspace.py --mode search --dataset SZ-TAXI --seq_len 12 --pred_len 12
python ./exps/rank_the_optimal_subspace.py --mode search --dataset SZ-TAXI --seq_len 24 --pred_len 24
python ./exps/rank_the_optimal_subspace.py --mode search --dataset SZ-TAXI --seq_len 48 --pred_len 48
python ./exps/rank_the_optimal_subspace.py --mode search --dataset SZ-TAXI --seq_len 168 --pred_len 1After finding the optimal architectures for all the tasks, we can reproduce the SOTA results in Table 6 and Table 7 of our paper by training them with our fast parameter adaptation strategy:
# Train the optimal ST-blocks with fast parameter adaptation (template)
python ./exps/generate_seeds.py --mode train --dataset <dataset_name> --seq_len <the windwow size> --pred_len <the horizon size>
# Specific
python ./exps/generate_seeds.py --mode train --dataset PEMS-BAY --seq_len 12 --pred_len 12
python ./exps/generate_seeds.py --mode train --dataset PEMS-BAY --seq_len 24 --pred_len 24
python ./exps/generate_seeds.py --mode train --dataset PEMS-BAY --seq_len 48 --pred_len 48
python ./exps/generate_seeds.py --mode train --dataset PEMS-BAY --seq_len 168 --pred_len 1
python ./exps/generate_seeds.py --mode train --dataset Electricity --seq_len 12 --pred_len 12
python ./exps/generate_seeds.py --mode train --dataset Electricity --seq_len 24 --pred_len 24
python ./exps/generate_seeds.py --mode train --dataset Electricity --seq_len 48 --pred_len 48
python ./exps/generate_seeds.py --mode train --dataset Electricity --seq_len 168 --pred_len 1
python ./exps/generate_seeds.py --mode train --dataset PEMSD7M --seq_len 12 --pred_len 12
python ./exps/generate_seeds.py --mode train --dataset PEMSD7M --seq_len 24 --pred_len 24
python ./exps/generate_seeds.py --mode train --dataset PEMSD7M --seq_len 48 --pred_len 48
python ./exps/generate_seeds.py --mode train --dataset PEMSD7M --seq_len 168 --pred_len 1
python ./exps/generate_seeds.py --mode train --dataset NYC-TAXI --seq_len 12 --pred_len 12
python ./exps/generate_seeds.py --mode train --dataset NYC-TAXI --seq_len 24 --pred_len 24
python ./exps/generate_seeds.py --mode train --dataset NYC-TAXI --seq_len 48 --pred_len 48
python ./exps/generate_seeds.py --mode train --dataset NYC-TAXI --seq_len 168 --pred_len 1
python ./exps/generate_seeds.py --mode train --dataset NYC-BIKE --seq_len 12 --pred_len 12
python ./exps/generate_seeds.py --mode train --dataset NYC-BIKE --seq_len 24 --pred_len 24
python ./exps/generate_seeds.py --mode train --dataset NYC-BIKE --seq_len 48 --pred_len 48
python ./exps/generate_seeds.py --mode train --dataset NYC-BIKE --seq_len 168 --pred_len 1
python ./exps/generate_seeds.py --mode train --dataset Los-Loop --seq_len 12 --pred_len 12
python ./exps/generate_seeds.py --mode train --dataset Los-Loop --seq_len 24 --pred_len 24
python ./exps/generate_seeds.py --mode train --dataset Los-Loop --seq_len 48 --pred_len 48
python ./exps/generate_seeds.py --mode train --dataset Los-Loop --seq_len 168 --pred_len 1
python ./exps/generate_seeds.py --mode train --dataset SZ-TAXI --seq_len 12 --pred_len 12
python ./exps/generate_seeds.py --mode train --dataset SZ-TAXI --seq_len 24 --pred_len 24
python ./exps/generate_seeds.py --mode train --dataset SZ-TAXI --seq_len 48 --pred_len 48
python ./exps/generate_seeds.py --mode train --dataset SZ-TAXI --seq_len 168 --pred_len 1We compare our iteratively pruning strategy with the one-time pruning baseline in our paper, so that we also provide the scripts to reproduce the results of the one-time pruning strategy:
# prune the search space in one time
python ./exps/iteratively_search_space_pruning.py --mode one-shotWith the one-time strategy, we obtain a pruned search space and we also collect seeds from them to pretrain the TAP:
# Collect seeds
bash ./scripts/pretrain_0.sh &
bash ./scripts/pretrain_1.sh &
bash ./scripts/pretrain_2.sh &
bash ./scripts/pretrain_3.sh &
bash ./scripts/pretrain_4.sh &
bash ./scripts/pretrain_5.sh &
bash ./scripts/pretrain_6.sh &# Pretrain the TAP
python ./exps/rank_the_optimal_subspace.py --mode trainAnd we can also reproduce the results by running the scripts as below:
# Search the optimal ST-block (template)
python ./exps/rank_the_optimal_subspace.py --mode search --dataset <dataset_name> --seq_len <the windwow size> --pred_len <the horizon size>
# Train the optimal ST-blocks (template)
python ./exps/generate_seeds.py --mode train --dataset <dataset_name> --seq_len <the windwow size> --pred_len <the horizon size>
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent