By using the Liu dataset, that contains SHARP parameters (not images), we train a DNN to predict the likelihood of a flare erupting from a sunspot.
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
git clone https://github.com/Dewald928/DeepFlarePred.git
Install conda and install environment:
conda env create -f environment.yml
Activate environment:
conda activate DeepFlarePred
If you have errors with the predefined validation set do get dev branch 0.9:
pip install git+https://github.com/skorch-dev/skorch.git
The Liu data should be downloaded from Liu dataset.
For the main_LSTM_Liu.py and main_TCN_Liu.py the Liu dataset needs to be downloaded and extracted to the
"/Data/Liu" folder such that:
./Data/
├── Liu
│ ├── C
│ ├── M
│ └── M5
To fix the incorrect column names used, run the following (Not needed if
repo is cloned, dataset already the fixed version):
python liu_data_fix
To create the normalized dataset run:
- for z_train:
python normalize_z_train.py - for z_minmax_train:
python normalize_z_minmax_train.py - fot z_minmax_all:
cp Data/Liu/M5/ Data/Liu/z_minmax_all -rf
To create the power transformed dataset run:
python normality_check.py
To run the script, you can either do a Weight & Biases sweep, or just simply run:
python main_TCN_Liu.py
The default configurations can be changed in config-defaults.yaml.
Follow instructions here
- Install docker
- Install GPU drivers
- Install nvidia docker Quickstart
- Then build dockerfile and push image to hub (useful Link1 Link2)
dvd928/deep_flare_pred:latest
paperspace jobs create --container dvd928/deep_flare_pred:latest --machineType P4000 --command wandb agent 5ks0xbql --ports 5000:5000 --project Liu_pytorch
| Script | Description |
|---|---|
cme_svm_updated_for_pyastro.ipynb |
Example notebook of Bobra's CME SVM |
data_aquisition_pipeline.ipynb |
Notebook for generating Liu et al. data (WIP) |
feature_selection.py |
For Univariate Feature selection and RFE |
inspectdata.py |
Basic data analysis & Pair plot generation |
nested_crossval.py |
Example script for nested crossval |
plot_classifier_comparison.py |
Sklearn script |
plot_cv_indices.py |
Sklearn script |
regression.py |
Synthetic LSTM regression testing |
roc_test.py |
ROC vs. PR for imbalanced dataset |
skorchCV.py |
Used for generating Toy Unbalanced Classification |
test_tcn.py |
For analysis of TCN and 1D convolution using sequences |
Titanic_Basic_Interpret.py |
Captum Example |
moving_std_protocol.py |
Protocol for downloading wandb runs and model selection, based on smooth training |
WNBtestscript.py |
wandb setup script |
workers_test.py |
Pytorch optimal workers test |
- Copy Liu's code to pytorch somewhat.
- Copy Liu architecture completely
- Cross-validation: Skorch library
- Regularization: L2 + Dropout
- Shuffled vs. Unshuffled. Shuffling is not very advers.
- GPU integration
- GPU optimization, just use larger batch sizes
- Implement Weight and Biases
- W&B sweeps check if it work
- W&B multiple gpu sweep. How to by using
tmux - Pytorch bottleneck test: Inconclusive, revisit
- Attention models Tested, but not sweeped
- Understand LSTM + TCN better
- GRU, LSTM and RNN switchable between
- hdf5 test script: Chen data uses hdf5, but unable to read the data
- MLP skorch test: RNN and custom logs not well supported
- TCN networks : better so far, slightly
- Early stopping and checkpointing on best validation TSS (LSTM only, so far)
- Test data? best wat to test network?
- LR scheduler
-
Create MLP that is equivalent to Nishzuka et al paper - Establish a baseline MLP
- Understand TCN operation
- Synthetic dataset
- Change sequence length with TCN
- LSTM vs. TCN?
- TCN baseline (Liu dataset (20/40 features?))
- ROC + Precision Recall curves, with AUC (train, val & test set)
- Find best features out of the 40. (captum)
- Occlusions method compare.
- What do these best features mean? (fits with other literature?)
- SHARP only TCN
- Case studies
- What does TSS mean in this context?
- How to interpret W&B parameters?
- SHARP query and infer model pipeline (un-normalize data)
- Incorporate SHARP magnetogram images like Chen article
- Use GAN for detecting anomalies
- MLP/LSTM attention models
- See a regression problem? LSTM regression is possible.