Skip to content

Software to train/evaluate models to reconstruct missing values in climate data (e.g., HadCRUT4) based on a U-Net with partial convolutions

License

Notifications You must be signed in to change notification settings

FREVA-CLINT/climatereconstructionAI

Repository files navigation

CRAI (Climate Reconstruction AI)

Software to train/evaluate models to reconstruct missing values in climate data (e.g., HadCRUT4) based on a U-Net with partial convolutions.

Dependencies

  • pytorch>=1.11.0
  • tqdm>=4.64.0
  • torchvision>=0.12.0
  • torchmetrics>=0.11.2
  • numpy>=1.21.6
  • matplotlib>=3.5.1
  • tensorboardX>=2.5
  • tensorboard>=2.9.0
  • xarray>=2022.3.0
  • dask>=2022.7.0
  • netcdf4>=1.5.8
  • setuptools==59.5.0
  • xesmf>=0.6.2
  • cartopy>=0.20.2
  • numba>=0.55.1

An Anaconda environment with all the required dependencies can be created using environment.yml:

conda env create -f environment.yml

To activate the environment, use:

conda activate crai

environment-cuda.yml should be used when working with GPUs using CUDA.

Installation

climatereconstructionAI can be installed using pip in the current directory:

pip install .

Usage

The software can be used to:

  • train a model (training)
  • infill climate datasets using a trained model (evaluation)

Input data

The directory containing the climate datasets should have the following sub-directories:

  • data and val for training
  • test for evaluation

The climate datasets should be in netCDF format and placed in the corresponding sub-directories.

The missing values can be defined separately as masks containing zeros (for the missing values) and ones (for the valid values). These masks should be in netCDF format and have the same dimension as the climate dataset. For the training, it is possible to shuffle the sequence of masks by using the "shuffle-masks" option.

A PyTorch model is required for the evaluation.

Execution

Once installed, the package can be used as:

  • a command line interface (CLI):
    • training:
    crai-train
    • evaluation:
    crai-evaluate
  • a Python library:
    • training:
    from climatereconstructionai import train
    train()
    • evaluation:
    from climatereconstructionai import evaluate
    evaluate()

For more information about the arguments:

usage: crai-train [-h] [--data-root-dir DATA_ROOT_DIR] [--mask-dir MASK_DIR] [--log-dir LOG_DIR] [--data-names DATA_NAMES] [--mask-names MASK_NAMES] [--data-types DATA_TYPES]
                  [--n-target-data N_TARGET_DATA] [--device DEVICE] [--shuffle-masks] [--channel-steps CHANNEL_STEPS] [--lstm-steps LSTM_STEPS] [--gru-steps GRU_STEPS]
                  [--encoding-layers ENCODING_LAYERS] [--pooling-layers POOLING_LAYERS] [--conv-factor CONV_FACTOR] [--weights WEIGHTS] [--steady-masks STEADY_MASKS]
                  [--loop-random-seed LOOP_RANDOM_SEED] [--cuda-random-seed CUDA_RANDOM_SEED] [--deterministic] [--attention] [--channel-reduction-rate CHANNEL_REDUCTION_RATE] [--disable-skip-layers]
                  [--disable-first-bn] [--masked-bn] [--lazy-load] [--global-padding] [--normalize-data] [--n-filters N_FILTERS] [--out-channels OUT_CHANNELS] [--dataset-name DATASET_NAME]
                  [--min-bounds MIN_BOUNDS] [--max-bounds MAX_BOUNDS] [--profile] [--val-names VAL_NAMES] [--snapshot-dir SNAPSHOT_DIR] [--resume-iter RESUME_ITER] [--batch-size BATCH_SIZE]
                  [--n-threads N_THREADS] [--multi-gpus] [--finetune] [--lr LR] [--lr-finetune LR_FINETUNE] [--max-iter MAX_ITER] [--log-interval LOG_INTERVAL]
                  [--lr-scheduler-patience LR_SCHEDULER_PATIENCE] [--save-model-interval SAVE_MODEL_INTERVAL] [--n-final-models N_FINAL_MODELS] [--final-models-interval FINAL_MODELS_INTERVAL]
                  [--loss-criterion LOSS_CRITERION] [--eval-timesteps EVAL_TIMESTEPS] [-f LOAD_FROM_FILE] [--vlim VLIM] [--lambda-loss LAMBDA_LOSS] [--val-metrics VAL_METRICS]
                  [--tensor-plots TENSOR_PLOTS] [--early-stopping-delta EARLY_STOPPING_DELTA] [--early-stopping-patience EARLY_STOPPING_PATIENCE] [--n-iters-val N_ITERS_VAL]

options:
  -h, --help            show this help message and exit
  --data-root-dir DATA_ROOT_DIR
                        Root directory containing the climate datasets
  --mask-dir MASK_DIR   Directory containing the mask datasets
  --log-dir LOG_DIR     Directory where the log files will be stored
  --data-names DATA_NAMES
                        Comma separated list of netCDF files (climate dataset) for training/infilling
  --mask-names MASK_NAMES
                        Comma separated list of netCDF files (mask dataset). If None, it extracts the masks from the climate dataset
  --data-types DATA_TYPES
                        Comma separated list of variable types, in the same order as data-names and mask-names
  --n-target-data N_TARGET_DATA
                        Number of data-names (from last) to be used as target data
  --device DEVICE       Device used by PyTorch (cuda or cpu)
  --shuffle-masks       Select mask indices randomly
  --channel-steps CHANNEL_STEPS
                        Comma separated number of considered sequences for channeled memory:past_steps,future_steps
  --lstm-steps LSTM_STEPS
                        Comma separated number of considered sequences for lstm: past_steps,future_steps
  --gru-steps GRU_STEPS
                        Comma separated number of considered sequences for gru: past_steps,future_steps
  --encoding-layers ENCODING_LAYERS
                        Number of encoding layers in the CNN
  --pooling-layers POOLING_LAYERS
                        Number of pooling layers in the CNN
  --conv-factor CONV_FACTOR
                        Number of channels in the deepest layer
  --weights WEIGHTS     Initialization weight
  --steady-masks STEADY_MASKS
                        Comma separated list of netCDF files containing a single mask to be applied to all timesteps. The number of steady-masks must be the same as out-channels
  --loop-random-seed LOOP_RANDOM_SEED
                        Random seed for iteration loop
  --cuda-random-seed CUDA_RANDOM_SEED
                        Random seed for CUDA
  --deterministic       Disable cudnn backends for reproducibility
  --attention           Enable the attention module
  --channel-reduction-rate CHANNEL_REDUCTION_RATE
                        Channel reduction rate for the attention module
  --disable-skip-layers
                        Disable the skip layers
  --disable-first-bn    Disable the batch normalization on the first layer
  --masked-bn           Use masked batch normalization instead of standard BN
  --lazy-load           Use lazy loading for large datasets
  --global-padding      Use a custom padding for global dataset
  --normalize-data      Normalize the input climate data to 0 mean and 1 std
  --n-filters N_FILTERS
                        Number of filters for the first/last layer
  --out-channels OUT_CHANNELS
                        Number of channels for the output data
  --dataset-name DATASET_NAME
                        Name of the dataset for format checking
  --min-bounds MIN_BOUNDS
                        Comma separated list of values defining the permitted lower-bound of output values
  --max-bounds MAX_BOUNDS
                        Comma separated list of values defining the permitted upper-bound of output values
  --profile             Profile code using tensorboard profiler
  --val-names VAL_NAMES
                        Comma separated list of netCDF files (climate dataset) for validation
  --snapshot-dir SNAPSHOT_DIR
                        Parent directory of the training checkpoints and the snapshot images
  --resume-iter RESUME_ITER
                        Iteration step from which the training will be resumed
  --batch-size BATCH_SIZE
                        Batch size
  --n-threads N_THREADS
                        Number of workers used in the data loader
  --multi-gpus          Use multiple GPUs, if any
  --finetune            Enable the fine tuning mode (use fine tuning parameterization and disable batch normalization
  --lr LR               Learning rate
  --lr-finetune LR_FINETUNE
                        Learning rate for fine tuning
  --max-iter MAX_ITER   Maximum number of iterations
  --log-interval LOG_INTERVAL
                        Iteration step interval at which a tensorboard summary log should be written
  --lr-scheduler-patience LR_SCHEDULER_PATIENCE
                        Patience for the lr scheduler
  --save-model-interval SAVE_MODEL_INTERVAL
                        Iteration step interval at which the model should be saved
  --n-final-models N_FINAL_MODELS
                        Number of final models to be saved
  --final-models-interval FINAL_MODELS_INTERVAL
                        Iteration step interval at which the final models should be saved
  --loss-criterion LOSS_CRITERION
                        Index defining the loss function (0=original from Liu et al., 1=MAE of the hole region)
  --eval-timesteps EVAL_TIMESTEPS
                        Sample indices for which a snapshot is created at each iter defined by log-interval
  -f LOAD_FROM_FILE, --load-from-file LOAD_FROM_FILE
                        Load all the arguments from a text file
  --vlim VLIM           Comma separated list of vmin,vmax values for the color scale of the snapshot images
  --lambda-loss LAMBDA_LOSS
                        Comma separated list of lambda factors (key) followed by their corresponding values.Overrides the loss_criterion pre-setting
  --val-metrics VAL_METRICS
                        Comma separated list of metrics that are evaluated on the val dataset at log-interval
  --tensor-plots TENSOR_PLOTS
                        Comma separated list of 2D plots to be added to tensorboard (error, distribution, correlation)
  --early-stopping-delta EARLY_STOPPING_DELTA
                        Mean relative delta of the val loss used for the termination criterion
  --early-stopping-patience EARLY_STOPPING_PATIENCE
                        Number of log-interval iterations used for the termination criterion
  --n-iters-val N_ITERS_VAL
                        Number of batch iterations used to average the validation loss
usage: crai-evaluate [-h] [--data-root-dir DATA_ROOT_DIR] [--mask-dir MASK_DIR] [--log-dir LOG_DIR] [--data-names DATA_NAMES] [--mask-names MASK_NAMES] [--data-types DATA_TYPES]
                     [--n-target-data N_TARGET_DATA] [--device DEVICE] [--shuffle-masks] [--channel-steps CHANNEL_STEPS] [--lstm-steps LSTM_STEPS] [--gru-steps GRU_STEPS]
                     [--encoding-layers ENCODING_LAYERS] [--pooling-layers POOLING_LAYERS] [--conv-factor CONV_FACTOR] [--weights WEIGHTS] [--steady-masks STEADY_MASKS]
                     [--loop-random-seed LOOP_RANDOM_SEED] [--cuda-random-seed CUDA_RANDOM_SEED] [--deterministic] [--attention] [--channel-reduction-rate CHANNEL_REDUCTION_RATE] [--disable-skip-layers]
                     [--disable-first-bn] [--masked-bn] [--lazy-load] [--global-padding] [--normalize-data] [--n-filters N_FILTERS] [--out-channels OUT_CHANNELS] [--dataset-name DATASET_NAME]
                     [--min-bounds MIN_BOUNDS] [--max-bounds MAX_BOUNDS] [--profile] [--model-dir MODEL_DIR] [--model-names MODEL_NAMES] [--evaluation-dirs EVALUATION_DIRS] [--eval-names EVAL_NAMES]
                     [--use-train-stats] [--create-graph] [--plot-results PLOT_RESULTS] [--partitions PARTITIONS] [--maxmem MAXMEM] [--split-outputs] [-f LOAD_FROM_FILE]

options:
  -h, --help            show this help message and exit
  --data-root-dir DATA_ROOT_DIR
                        Root directory containing the climate datasets
  --mask-dir MASK_DIR   Directory containing the mask datasets
  --log-dir LOG_DIR     Directory where the log files will be stored
  --data-names DATA_NAMES
                        Comma separated list of netCDF files (climate dataset) for training/infilling
  --mask-names MASK_NAMES
                        Comma separated list of netCDF files (mask dataset). If None, it extracts the masks from the climate dataset
  --data-types DATA_TYPES
                        Comma separated list of variable types, in the same order as data-names and mask-names
  --n-target-data N_TARGET_DATA
                        Number of data-names (from last) to be used as target data
  --device DEVICE       Device used by PyTorch (cuda or cpu)
  --shuffle-masks       Select mask indices randomly
  --channel-steps CHANNEL_STEPS
                        Comma separated number of considered sequences for channeled memory:past_steps,future_steps
  --lstm-steps LSTM_STEPS
                        Comma separated number of considered sequences for lstm: past_steps,future_steps
  --gru-steps GRU_STEPS
                        Comma separated number of considered sequences for gru: past_steps,future_steps
  --encoding-layers ENCODING_LAYERS
                        Number of encoding layers in the CNN
  --pooling-layers POOLING_LAYERS
                        Number of pooling layers in the CNN
  --conv-factor CONV_FACTOR
                        Number of channels in the deepest layer
  --weights WEIGHTS     Initialization weight
  --steady-masks STEADY_MASKS
                        Comma separated list of netCDF files containing a single mask to be applied to all timesteps. The number of steady-masks must be the same as out-channels
  --loop-random-seed LOOP_RANDOM_SEED
                        Random seed for iteration loop
  --cuda-random-seed CUDA_RANDOM_SEED
                        Random seed for CUDA
  --deterministic       Disable cudnn backends for reproducibility
  --attention           Enable the attention module
  --channel-reduction-rate CHANNEL_REDUCTION_RATE
                        Channel reduction rate for the attention module
  --disable-skip-layers
                        Disable the skip layers
  --disable-first-bn    Disable the batch normalization on the first layer
  --masked-bn           Use masked batch normalization instead of standard BN
  --lazy-load           Use lazy loading for large datasets
  --global-padding      Use a custom padding for global dataset
  --normalize-data      Normalize the input climate data to 0 mean and 1 std
  --n-filters N_FILTERS
                        Number of filters for the first/last layer
  --out-channels OUT_CHANNELS
                        Number of channels for the output data
  --dataset-name DATASET_NAME
                        Name of the dataset for format checking
  --min-bounds MIN_BOUNDS
                        Comma separated list of values defining the permitted lower-bound of output values
  --max-bounds MAX_BOUNDS
                        Comma separated list of values defining the permitted upper-bound of output values
  --profile             Profile code using tensorboard profiler
  --model-dir MODEL_DIR
                        Directory of the trained models
  --model-names MODEL_NAMES
                        Model names
  --evaluation-dirs EVALUATION_DIRS
                        Directory where the output files will be stored
  --eval-names EVAL_NAMES
                        Prefix used for the output filenames
  --use-train-stats     Use mean and std from training data for normalization
  --create-graph        Create a Tensorboard graph of the NN
  --plot-results PLOT_RESULTS
                        Create plot images of the results for the comma separated list of time indices
  --partitions PARTITIONS
                        Split the climate dataset into several partitions along the time coordinate
  --maxmem MAXMEM       Maximum available memory in MB (overwrite partitions parameter)
  --split-outputs       Do not merge the outputs when using multiple models and/or partitions
  -f LOAD_FROM_FILE, --load-from-file LOAD_FROM_FILE
                        Load all the arguments from a text file

Example

An example can be found in the directory demo. The instructions to run the example are given in the README.md file.

License

CRAI is licensed under the terms of the BSD 3-Clause license.

Contributions

CRAI is maintained by the Climate Informatics and Technology group at DKRZ (Deutsches Klimarechenzentrum).

  • Previous contributing authors: Naoto Inoue, Christopher Kadow, Stephan Seitz
  • Current contributing authors: Johannes Meuer, Maximilian Witte, Étienne Plésiat.

About

Software to train/evaluate models to reconstruct missing values in climate data (e.g., HadCRUT4) based on a U-Net with partial convolutions

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages