├── data
├── 2019
│ └── .gitignore
└── 2020
│ └── .gitignore
├── submission
├── 2019
│ └── .gitignore
└── 2020
│ └── .gitignore
├── utils
├── cached_sample_ids
│ └── .gitignore
├── evaluate.py
├── w4c_dataloader.py
└── data_utils.py
├── images
├── weather4cast_v1000-26.png
└── opera_satelite_context_explained.png
├── mk_baseline.sh
├── mk_env.sh
├── LICENSE
├── .gitignore
├── mk_pred_core_3boxi2019.sh
├── mk_pred_core.sh
├── mk_heldout_core.sh
├── mk_pred_transfer.sh
├── mk_heldout_transfer.sh
├── w4cNew.yml
├── models
├── configurations
│ ├── config_baseline_stage1.yaml
│ ├── config_baseline_stage2-pred.yaml
│ └── config_baseline_stage2.yaml
├── unet.patch
└── unet_lightning.py
├── train.py
├── w4c.yml
├── README.md
└── COPYING
/data/2019/.gitignore:
--------------------------------------------------------------------------------
1 |
--------------------------------------------------------------------------------
/data/2020/.gitignore:
--------------------------------------------------------------------------------
1 |
--------------------------------------------------------------------------------
/submission/2019/.gitignore:
--------------------------------------------------------------------------------
1 |
--------------------------------------------------------------------------------
/submission/2020/.gitignore:
--------------------------------------------------------------------------------
1 |
--------------------------------------------------------------------------------
/utils/cached_sample_ids/.gitignore:
--------------------------------------------------------------------------------
1 |
--------------------------------------------------------------------------------
/images/weather4cast_v1000-26.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/iarai/weather4cast-2022/HEAD/images/weather4cast_v1000-26.png
--------------------------------------------------------------------------------
/images/opera_satelite_context_explained.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/iarai/weather4cast-2022/HEAD/images/opera_satelite_context_explained.png
--------------------------------------------------------------------------------
/mk_baseline.sh:
--------------------------------------------------------------------------------
1 | cd models
2 | rm unet.py 2>/dev/null
3 | wget https://raw.githubusercontent.com/ELEKTRONN/elektronn3/f754796d861f1cfe1c19dfc7819087972573ce40/elektronn3/models/unet.py
4 | patch .
20 |
--------------------------------------------------------------------------------
/.gitignore:
--------------------------------------------------------------------------------
1 | # Emacs backup files
2 | *~
3 |
4 | #Local files to ignore
5 | *test.py
6 | *initial_eda.ipynb
7 | *old_data_utils.py
8 | *old_visualise.py
9 | *.pdf
10 | *.pkl
11 | split_generation_logs/
12 | ./plots
13 | *.h5
14 |
15 | # large files
16 | tmp_w4c/
17 | utils/*lightning_logs/
18 | *lightning_logs/
19 |
20 | # Byte-compiled / optimized / DLL files
21 | __pycache__
22 | *.py[cod]
23 | *$py.class
24 |
25 | # C extensions
26 | *.so
27 |
28 | # Distribution / packaging
29 | .Python
30 | build/
31 | develop-eggs/
32 | dist/
33 | downloads/
34 | eggs/
35 | .eggs/
36 | lib/
37 | lib64/
38 | parts/
39 | sdist/
40 | var/
41 | wheels/
42 | pip-wheel-metadata/
43 | share/python-wheels/
44 | *.egg-info/
45 | .installed.cfg
46 | *.egg
47 | MANIFEST
48 |
49 | # PyInstaller
50 | # Usually these files are written by a python script from a template
51 | # before PyInstaller builds the exe, so as to inject date/other infos into it.
52 | *.manifest
53 | *.spec
54 |
55 | # Installer logs
56 | pip-log.txt
57 | pip-delete-this-directory.txt
58 |
59 | # Unit test / coverage reports
60 | htmlcov/
61 | .tox/
62 | .nox/
63 | .coverage
64 | .coverage.*
65 | .cache
66 | nosetests.xml
67 | coverage.xml
68 | *.cover
69 | *.py,cover
70 | .hypothesis/
71 | .pytest_cache/
72 |
73 | # Translations
74 | *.mo
75 | *.pot
76 |
77 | # Django stuff:
78 | *.log
79 | local_settings.py
80 | db.sqlite3
81 | db.sqlite3-journal
82 |
83 | # Flask stuff:
84 | instance/
85 | .webassets-cache
86 |
87 | # Scrapy stuff:
88 | .scrapy
89 |
90 | # Sphinx documentation
91 | docs/_build/
92 |
93 | # PyBuilder
94 | target/
95 |
96 | # Jupyter Notebook
97 | .ipynb_checkpoints
98 |
99 | # IPython
100 | profile_default/
101 | ipython_config.py
102 |
103 | # pyenv
104 | .python-version
105 |
106 | # pipenv
107 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
108 | # However, in case of collaboration, if having platform-specific dependencies or dependencies
109 | # having no cross-platform support, pipenv may install dependencies that don't work, or not
110 | # install all needed dependencies.
111 | #Pipfile.lock
112 |
113 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow
114 | __pypackages__/
115 |
116 | # Celery stuff
117 | celerybeat-schedule
118 | celerybeat.pid
119 |
120 | # SageMath parsed files
121 | *.sage.py
122 |
123 | # Environments
124 | .env
125 | .venv
126 | env/
127 | venv/
128 | ENV/
129 | env.bak/
130 | venv.bak/
131 |
132 | # Spyder project settings
133 | .spyderproject
134 | .spyproject
135 |
136 | # Rope project settings
137 | .ropeproject
138 |
139 | # mkdocs documentation
140 | /site
141 |
142 | # mypy
143 | .mypy_cache/
144 | .dmypy.json
145 | dmypy.json
146 |
147 | # Pyre type checker
148 | .pyre/
149 |
150 |
151 | #vsc
152 | .vscode/
--------------------------------------------------------------------------------
/mk_pred_core_3boxi2019.sh:
--------------------------------------------------------------------------------
1 | #!/usr/bin/bash
2 |
3 | cdir=models/configurations/;
4 | sdirDef=submission.3boxi2019;
5 | gpuDef=0;
6 |
7 | sdir=$sdirDef;
8 | gpu=$gpuDef;
9 |
10 | cbase="$1"; shift;
11 | cpt="$1"; shift;
12 | if [ -n "$1" ]; then
13 | sdir="$1"; shift;
14 | fi
15 | if [ -n "$1" ]; then
16 | gpu="$1"; shift;
17 | fi
18 |
19 | out=$sdir.zip;
20 |
21 |
22 | cat <"$cout";
85 | python train.py --gpus $gpu --mode predict --config_path "$cnew" \
86 | --checkpoint "$cpt"
87 | rm $cout;
88 | done
89 | done
90 |
91 | if pushd $sdir; then
92 | echo /=== output summary
93 | ls -l */*
94 | fl=`find -type f|grep -v '/[.]'|sed -e 's/ /%20/g'`;
95 | for f in $fl; do
96 | f="${f//%20/ }";
97 | echo Compressing $f ...
98 | gzip -9f "$f" &
99 | done
100 | wait;
101 | echo ...zip packing $sdir
102 | [ -s "../$out" ] && rm -f "../$out";
103 | zip -0mr ../$out * -x .\* \*/.\*
104 | popd
105 | rmdir $sdir;
106 | ls -l $out;
107 | else
108 | echo "Cannot change to $sdir - aborting"
109 | exit
110 | fi
111 | echo \\=== done
112 |
--------------------------------------------------------------------------------
/mk_pred_core.sh:
--------------------------------------------------------------------------------
1 | #!/usr/bin/bash
2 |
3 | cdir=models/configurations/;
4 | sdirDef=submission.core;
5 | gpuDef=0;
6 |
7 | sdir=$sdirDef;
8 | gpu=$gpuDef;
9 |
10 | cbase="$1"; shift;
11 | cpt="$1"; shift;
12 | if [ -n "$1" ]; then
13 | sdir="$1"; shift;
14 | fi
15 | if [ -n "$1" ]; then
16 | gpu="$1"; shift;
17 | fi
18 |
19 | out=$sdir.zip;
20 |
21 |
22 | cat <"$cout";
86 | python train.py --gpus $gpu --mode predict --config_path "$cnew" \
87 | --checkpoint "$cpt"
88 | rm $cout;
89 | done
90 | done
91 |
92 | if pushd $sdir; then
93 | echo /=== output summary
94 | ls -l */*
95 | fl=`find -type f|grep -v '/[.]'|sed -e 's/ /%20/g'`;
96 | for f in $fl; do
97 | f="${f//%20/ }";
98 | echo Compressing $f ...
99 | gzip -9f "$f" &
100 | done
101 | wait;
102 | echo ...zip packing $sdir
103 | [ -s "../$out" ] && rm -f "../$out";
104 | zip -0mr ../$out * -x .\* \*/.\*
105 | popd
106 | rmdir $sdir;
107 | ls -l $out;
108 | else
109 | echo "Cannot change to $sdir - aborting"
110 | exit
111 | fi
112 | echo \\=== done
113 |
--------------------------------------------------------------------------------
/mk_heldout_core.sh:
--------------------------------------------------------------------------------
1 | #!/usr/bin/bash
2 |
3 | cdir=models/configurations/;
4 | sdirDef=submission.heldout.core;
5 | gpuDef=0;
6 |
7 | sdir=$sdirDef;
8 | gpu=$gpuDef;
9 |
10 | cbase="$1"; shift;
11 | cpt="$1"; shift;
12 | if [ -n "$1" ]; then
13 | sdir="$1"; shift;
14 | fi
15 | if [ -n "$1" ]; then
16 | gpu="$1"; shift;
17 | fi
18 |
19 | out=$sdir.zip;
20 |
21 |
22 | cat <"$cout";
86 | python train.py --gpus $gpu --mode heldout --config_path "$cnew" \
87 | --checkpoint "$cpt"
88 | rm $cout;
89 | done
90 | done
91 |
92 | if pushd $sdir; then
93 | echo /=== output summary
94 | ls -l */*
95 | fl=`find -type f|grep -v '/[.]'|sed -e 's/ /%20/g'`;
96 | for f in $fl; do
97 | f="${f//%20/ }";
98 | echo Compressing $f ...
99 | gzip -9f "$f" &
100 | done
101 | wait;
102 | echo ...zip packing $sdir
103 | [ -s "../$out" ] && rm -f "../$out";
104 | zip -0mr ../$out * -x .\* \*/.\*
105 | popd
106 | rmdir $sdir;
107 | ls -l $out;
108 | else
109 | echo "Cannot change to $sdir - aborting"
110 | exit
111 | fi
112 | echo \\=== done
113 |
--------------------------------------------------------------------------------
/mk_pred_transfer.sh:
--------------------------------------------------------------------------------
1 | #!/usr/bin/bash
2 |
3 | cdir=models/configurations/;
4 | sdirDef=submission.transfer;
5 | gpuDef=0;
6 |
7 | sdir=$sdirDef;
8 | gpu=$gpuDef;
9 |
10 | cbase="$1"; shift;
11 | cpt="$1"; shift;
12 | if [ -n "$1" ]; then
13 | sdir="$1"; shift;
14 | fi
15 | if [ -n "$1" ]; then
16 | gpu="$1"; shift;
17 | fi
18 |
19 | out=$sdir.zip;
20 |
21 |
22 | cat <"$cout";
85 | python train.py --gpus $gpu --mode predict --config_path "$cnew" \
86 | --checkpoint "$cpt"
87 | rm $cout;
88 | done
89 | done
90 | for y in 2021; do
91 | d=$sdir/$y;
92 | [ -d "$d" ] || mkdir "$d";
93 | for r in boxi_0015 boxi_0034 boxi_0076 \
94 | roxi_0004 roxi_0005 roxi_0006 roxi_0007 \
95 | roxi_0008 roxi_0009 roxi_0010; do
96 | echo /=== $r $y for $cpt
97 | cin=$cdir$cbase;
98 | cnew=${cbase%.yaml}-$$-$r-$y.yaml;
99 | cout=$cdir$cnew;
100 | sed -e "s/%REGION%/$r/g" -e "s/%YEAR%/$y/" -e "s/%SDIR%/$sdir/" \
101 | <"$cin" >"$cout";
102 | python train.py --gpus $gpu --mode predict --config_path "$cnew" \
103 | --checkpoint "$cpt"
104 | rm $cout;
105 | done
106 | done
107 |
108 | if pushd $sdir; then
109 | echo /=== output summary
110 | ls -l */*
111 | fl=`find -type f|grep -v '/[.]'|sed -e 's/ /%20/g'`;
112 | for f in $fl; do
113 | f="${f//%20/ }";
114 | echo Compressing $f ...
115 | gzip -9f "$f" &
116 | done
117 | wait;
118 | echo ...zip packing $sdir
119 | [ -s "../$out" ] && rm -f "../$out";
120 | zip -0mr ../$out * -x .\* \*/.\*
121 | popd
122 | rmdir $sdir;
123 | ls -l $out;
124 | else
125 | echo "Cannot change to $sdir - aborting"
126 | exit
127 | fi
128 | echo \\=== done
129 |
--------------------------------------------------------------------------------
/mk_heldout_transfer.sh:
--------------------------------------------------------------------------------
1 | #!/usr/bin/bash
2 |
3 | cdir=models/configurations/;
4 | sdirDef=submission.heldout.transfer;
5 | gpuDef=0;
6 |
7 | sdir=$sdirDef;
8 | gpu=$gpuDef;
9 |
10 | cbase="$1"; shift;
11 | cpt="$1"; shift;
12 | if [ -n "$1" ]; then
13 | sdir="$1"; shift;
14 | fi
15 | if [ -n "$1" ]; then
16 | gpu="$1"; shift;
17 | fi
18 |
19 | out=$sdir.zip;
20 |
21 |
22 | cat <"$cout";
85 | python train.py --gpus $gpu --mode heldout --config_path "$cnew" \
86 | --checkpoint "$cpt"
87 | rm $cout;
88 | done
89 | done
90 | for y in 2021; do
91 | d=$sdir/$y;
92 | [ -d "$d" ] || mkdir "$d";
93 | for r in boxi_0015 boxi_0034 boxi_0076 \
94 | roxi_0004 roxi_0005 roxi_0006 roxi_0007 \
95 | roxi_0008 roxi_0009 roxi_0010; do
96 | echo /=== $r $y for $cpt
97 | cin=$cdir$cbase;
98 | cnew=${cbase%.yaml}-$$-$r-$y.yaml;
99 | cout=$cdir$cnew;
100 | sed -e "s/%REGION%/$r/g" -e "s/%YEAR%/$y/" -e "s/%SDIR%/$sdir/" \
101 | <"$cin" >"$cout";
102 | python train.py --gpus $gpu --mode heldout --config_path "$cnew" \
103 | --checkpoint "$cpt"
104 | rm $cout;
105 | done
106 | done
107 |
108 | if pushd $sdir; then
109 | echo /=== output summary
110 | ls -l */*
111 | fl=`find -type f|grep -v '/[.]'|sed -e 's/ /%20/g'`;
112 | for f in $fl; do
113 | f="${f//%20/ }";
114 | echo Compressing $f ...
115 | gzip -9f "$f" &
116 | done
117 | wait;
118 | echo ...zip packing $sdir
119 | [ -s "../$out" ] && rm -f "../$out";
120 | zip -0mr ../$out * -x .\* \*/.\*
121 | popd
122 | rmdir $sdir;
123 | ls -l $out;
124 | else
125 | echo "Cannot change to $sdir - aborting"
126 | exit
127 | fi
128 | echo \\=== done
129 |
--------------------------------------------------------------------------------
/utils/evaluate.py:
--------------------------------------------------------------------------------
1 | # Weather4cast 2022 Starter Kit
2 | #
3 | # Copyright (C) 2022
4 | # Institute of Advanced Research in Artificial Intelligence (IARAI)
5 |
6 | # This file is part of the Weather4cast 2022 Starter Kit.
7 | #
8 | # The Weather4cast 2022 Starter Kit is free software: you can redistribute it
9 | # and/or modify it under the terms of the GNU General Public License as
10 | # published by the Free Software Foundation, either version 3 of the License,
11 | # or (at your option) any later version.
12 | #
13 | # The Weather4cast 2022 Starter Kit is distributed in the hope that it will be
14 | # useful, but WITHOUT ANY WARRANTY; without even the implied warranty of
15 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 | # GNU General Public License for more details.
17 | #
18 | # You should have received a copy of the GNU General Public License
19 | # along with this program. If not, see .
20 |
21 | # Contributors: Aleksandra Gruca, Pedro Herruzo, David Kreil, Stephen Moran
22 |
23 |
24 | from sklearn.metrics import confusion_matrix
25 | import numpy as np
26 | import torch as t
27 | import sys
28 |
29 | def get_confusion_matrix(y_true, y_pred):
30 | """get confusion matrix from y_true and y_pred
31 |
32 | Args:
33 | y_true (numpy array): ground truth
34 | y_pred (numpy array): prediction
35 |
36 | Returns:
37 | confusion matrix
38 | """
39 |
40 | labels = [0,1]
41 | return confusion_matrix(y_true, y_pred, labels=labels).ravel()
42 |
43 | def recall_precision_f1_acc(y, y_hat):
44 | """ returns metrics for recall, precision, f1, accuracy
45 |
46 | Args:
47 | y (numpy array): ground truth
48 | y_hat (numpy array): prediction
49 |
50 | Returns:
51 | recall(float): recall/TPR
52 | precision(float): precision/PPV
53 | F1(float): f1-score
54 | acc(float): accuracy
55 | csi(float): critical success index
56 | """
57 |
58 | # pytorch to numpy
59 | y, y_hat = [o.cpu() for o in [y, y_hat]]
60 | y, y_hat = [np.asarray(o) for o in [y, y_hat]]
61 |
62 | cm = get_confusion_matrix(y.ravel(), y_hat.ravel())
63 | if len(cm)==4:
64 | tn, fp, fn, tp = cm
65 | recall, precision, F1, acc, csi = 0, 0, 0, 0, 0
66 |
67 | if (tp + fn) > 0:
68 | recall = tp / (tp + fn)
69 |
70 | if (tp + fp) > 0:
71 | precision = tp / (tp + fp)
72 |
73 | if (precision + recall) > 0:
74 | F1 = 2 * (precision * recall) / (precision + recall)
75 |
76 | if (tp + fn + fp) > 0:
77 | csi = tp / (tp + fn + fp)
78 |
79 | if (tn+fp+fn+tp) > 0:
80 | acc = (tn + tp) / (tn+fp+fn+tp)
81 | else:
82 | print("FATAL ERROR: cannot create confusion matrix")
83 | print("EXITING....")
84 | sys.exit()
85 |
86 | return recall, precision, F1, acc, csi
87 |
88 |
89 | def iou_class(y_pred: t.Tensor, y_true: t.Tensor):
90 | #y_true, y_pred = [o.cpu() for o in [y_true, y_pred]]
91 | #y_true, y_pred = [np.asarray(o) for o in [y_true, y_pred]]
92 | y_pred = y_pred.int()
93 | y_true = y_true.int()
94 | # Outputs: BATCH X H X W
95 |
96 | intersection = (y_pred & y_true).float().sum() # Will be zero if Truth=0 or Prediction=0
97 | union = (y_pred | y_true).float().sum() # Will be zero if both are 0
98 |
99 | if union>0:
100 | iou = intersection / union
101 | else:
102 | iou = 0
103 |
104 | iou = iou.cpu()
105 | return iou
106 |
--------------------------------------------------------------------------------
/w4cNew.yml:
--------------------------------------------------------------------------------
1 | name: w4cNew
2 | channels:
3 | - pytorch
4 | - conda-forge
5 | - defaults
6 | dependencies:
7 | - _libgcc_mutex=0.1=conda_forge
8 | - _openmp_mutex=4.5=2_kmp_llvm
9 | - absl-py=1.2.0=pyhd8ed1ab_0
10 | - aiohttp=3.8.1=py38h0a891b7_1
11 | - aiosignal=1.2.0=pyhd8ed1ab_0
12 | - async-timeout=4.0.2=pyhd8ed1ab_0
13 | - attrs=22.1.0=pyh71513ae_1
14 | - blas=1.0=mkl
15 | - blinker=1.4=py_1
16 | - bottleneck=1.3.5=py38h7deecbd_0
17 | - brotlipy=0.7.0=py38h0a891b7_1004
18 | - bzip2=1.0.8=h7f98852_4
19 | - c-ares=1.18.1=h7f98852_0
20 | - ca-certificates=2022.07.19=h06a4308_0
21 | - cachetools=5.2.0=pyhd8ed1ab_0
22 | - certifi=2022.6.15=py38h06a4308_0
23 | - cffi=1.15.1=py38h4a40e3a_0
24 | - charset-normalizer=2.1.0=pyhd8ed1ab_0
25 | - click=8.1.3=py38h578d9bd_0
26 | - colorama=0.4.5=pyhd8ed1ab_0
27 | - cryptography=37.0.4=py38h2b5fc30_0
28 | - cudatoolkit=11.6.0=hecad31d_10
29 | - ffmpeg=4.3=hf484d3e_0
30 | - freetype=2.10.4=hca18f0e_2
31 | - frozenlist=1.3.1=py38h0a891b7_0
32 | - fsspec=2022.7.1=pyhd8ed1ab_0
33 | - future=0.18.2=py38h578d9bd_5
34 | - gmp=6.2.1=h58526e2_0
35 | - gnutls=3.6.13=h85f3911_1
36 | - google-auth=2.10.0=pyh6c4a22f_0
37 | - google-auth-oauthlib=0.4.6=pyhd8ed1ab_0
38 | - grpcio=1.46.3=py38ha0cdfde_0
39 | - h5py=3.7.0=py38h737f45e_0
40 | - hdf5=1.10.6=h3ffc7dd_1
41 | - idna=3.3=pyhd8ed1ab_0
42 | - importlib-metadata=4.11.4=py38h578d9bd_0
43 | - joblib=1.1.0=pyhd8ed1ab_0
44 | - jpeg=9e=h166bdaf_2
45 | - lame=3.100=h7f98852_1001
46 | - lcms2=2.12=hddcbb42_0
47 | - ld_impl_linux-64=2.36.1=hea4e1c9_2
48 | - lerc=3.0=h9c3ff4c_0
49 | - libblas=3.9.0=12_linux64_mkl
50 | - libcblas=3.9.0=12_linux64_mkl
51 | - libdeflate=1.10=h7f98852_0
52 | - libffi=3.4.2=h7f98852_5
53 | - libgcc-ng=12.1.0=h8d9b700_16
54 | - libgfortran-ng=12.1.0=h69a702a_16
55 | - libgfortran5=12.1.0=hdcd56e2_16
56 | - libiconv=1.17=h166bdaf_0
57 | - liblapack=3.9.0=12_linux64_mkl
58 | - libnsl=2.0.0=h7f98852_0
59 | - libpng=1.6.37=h753d276_4
60 | - libprotobuf=3.19.4=h780b84a_0
61 | - libsqlite=3.39.2=h753d276_1
62 | - libstdcxx-ng=12.1.0=ha89aaad_16
63 | - libtiff=4.3.0=h0fcbabc_4
64 | - libuuid=2.32.1=h7f98852_1000
65 | - libwebp-base=1.2.4=h166bdaf_0
66 | - libzlib=1.2.12=h166bdaf_2
67 | - llvm-openmp=14.0.4=he0ac6c6_0
68 | - lz4-c=1.9.3=h9c3ff4c_1
69 | - markdown=3.4.1=pyhd8ed1ab_0
70 | - markupsafe=2.1.1=py38h0a891b7_1
71 | - mkl=2021.4.0=h8d4b97c_729
72 | - mkl-service=2.4.0=py38h95df7f1_0
73 | - mkl_fft=1.3.1=py38h8666266_1
74 | - mkl_random=1.2.2=py38h1abd341_0
75 | - multidict=6.0.2=py38h0a891b7_1
76 | - ncurses=6.3=h27087fc_1
77 | - nettle=3.6=he412f7d_0
78 | - numexpr=2.8.3=py38h807cd23_0
79 | - numpy=1.23.1=py38h6c91a56_0
80 | - numpy-base=1.23.1=py38ha15fc14_0
81 | - oauthlib=3.2.0=pyhd8ed1ab_0
82 | - olefile=0.46=pyh9f0ad1d_1
83 | - openh264=2.1.1=h780b84a_0
84 | - openssl=1.1.1q=h7f8727e_0
85 | - packaging=21.3=pyhd8ed1ab_0
86 | - pandas=1.4.3=py38h6a678d5_0
87 | - pillow=6.2.2=py38h9776b28_0
88 | - pip=22.2.2=pyhd8ed1ab_0
89 | - protobuf=3.19.4=py38h709712a_0
90 | - pyasn1=0.4.8=py_0
91 | - pyasn1-modules=0.2.7=py_0
92 | - pycparser=2.21=pyhd8ed1ab_0
93 | - pydeprecate=0.3.2=pyhd8ed1ab_0
94 | - pyjwt=2.4.0=pyhd8ed1ab_0
95 | - pyopenssl=22.0.0=pyhd8ed1ab_0
96 | - pyparsing=3.0.9=pyhd8ed1ab_0
97 | - pysocks=1.7.1=py38h578d9bd_5
98 | - python=3.8.13=h582c2e5_0_cpython
99 | - python-dateutil=2.8.2=pyhd3eb1b0_0
100 | - python_abi=3.8=2_cp38
101 | - pytorch=1.12.1=py3.8_cuda11.6_cudnn8.3.2_0
102 | - pytorch-lightning=1.7.1=pyhd8ed1ab_0
103 | - pytorch-mutex=1.0=cuda
104 | - pytz=2022.1=py38h06a4308_0
105 | - pyu2f=0.1.5=pyhd8ed1ab_0
106 | - pyyaml=6.0=py38h0a891b7_4
107 | - readline=8.1.2=h0f457ee_0
108 | - requests=2.28.1=pyhd8ed1ab_0
109 | - requests-oauthlib=1.3.1=pyhd8ed1ab_0
110 | - rsa=4.9=pyhd8ed1ab_0
111 | - scikit-learn=1.1.2=py38h0b08f9b_0
112 | - scipy=1.9.0=py38hea3f02b_0
113 | - setuptools=59.5.0=py38h578d9bd_0
114 | - six=1.16.0=pyh6c4a22f_0
115 | - sqlite=3.39.2=h4ff8645_1
116 | - tbb=2021.5.0=h924138e_1
117 | - tensorboard=2.10.0=pyhd8ed1ab_0
118 | - tensorboard-data-server=0.6.0=py38h2b5fc30_2
119 | - tensorboard-plugin-wit=1.8.1=pyhd8ed1ab_0
120 | - threadpoolctl=3.1.0=pyh8a188c0_0
121 | - tk=8.6.12=h27826a3_0
122 | - torchaudio=0.12.1=py38_cu116
123 | - torchmetrics=0.9.3=pyhd8ed1ab_0
124 | - torchvision=0.13.1=py38_cu116
125 | - tqdm=4.64.0=pyhd8ed1ab_0
126 | - typing-extensions=4.3.0=hd8ed1ab_0
127 | - typing_extensions=4.3.0=pyha770c72_0
128 | - urllib3=1.26.11=pyhd8ed1ab_0
129 | - werkzeug=2.2.2=pyhd8ed1ab_0
130 | - wheel=0.37.1=pyhd8ed1ab_0
131 | - xz=5.2.6=h166bdaf_0
132 | - yaml=0.2.5=h7f98852_2
133 | - yarl=1.7.2=py38h0a891b7_2
134 | - zipp=3.8.1=pyhd8ed1ab_0
135 | - zlib=1.2.12=h166bdaf_2
136 | - zstd=1.5.2=h8a70e8d_4
137 |
--------------------------------------------------------------------------------
/models/configurations/config_baseline_stage1.yaml:
--------------------------------------------------------------------------------
1 | experiment: # Settings for the current experiment
2 | name: 'U-NET-252' # name of the current experiment for logging
3 | experiment_folder: 'lightning_logs/' # folder to save logs
4 | sub_folder: 'baseline' # logging sub-folder
5 | precision: 32 # bit precision of weights. 32 or 16
6 | logging: True # Toggle logging on/off
7 | dataset:
8 | # Available bands: 'IR_016', 'IR_039', 'IR_087', 'IR_097', 'IR_108', 'IR_120', 'IR_134', 'VIS006', 'VIS008', 'WV_062', 'WV_073'
9 | sat_bands: ['IR_016', 'IR_039', 'IR_087', 'IR_097', 'IR_108', 'IR_120', 'IR_134', 'VIS006', 'VIS008', 'WV_062', 'WV_073']
10 | regions: [ 'boxi_0015', 'boxi_0034', 'boxi_0076' ] # stage-1
11 | # or, e.g., ['boxi_0015'] if you want to train your model for one region only
12 | input_product: REFL-BT
13 | output_product: RATE
14 | out_channels: 1
15 | in_channels: 11
16 | swap_time_ch: True #swap the time and channels axis - if set to False (B, T, C, H, W)
17 | full_opera_context: 1512 # (105+42+105)*6;
18 | size_target_center: 252 # 42*6;
19 | len_seq_in: 4 # number of input slots per training sequence
20 | len_seq_predict: 32 # output slots; both still hard-coded in places!
21 | data_root: 'data'
22 | years: [ '2019' ] # stage-1
23 | splits_path: 'data/timestamps_and_splits_2019.csv' # stage-1
24 |
25 | preprocess_OPERA:
26 | # # var mean sd min median max length pos.weight.thr_0
27 | # 1 RATE 0.07165331 0.6302647 0 0 127.9399 5011989696 6.114572
28 | RATE:
29 | rainfall_rate-500X500:
30 | mask: [-9999000.0, inf, nan, max128] #, 0range0.1] # Mostly used for loss function: values added here are added to a mask and not used for loss
31 | map: [[lessthan0.0, 0], [greaterthan0.0, 1], [-8888000.0, 0], [-9999000.0, 0], [inf, 0], [nan, 0]] #Mostly used for input preprocessing # 1. map values # stage-1
32 | mean_std: [0.07165331, 0.6302647]
33 | range: [0, 128] # 2. we evaluate only pixels in this range
34 | standardise: False # 3. use log(x+1) instead & normalize (x/max)
35 | bin: False
36 | preprocess_HRIT: # 1: map values, 2: normalise in range per variable if process==True
37 | # # var mean sd min median max length
38 | # 1 IR_016 0.06605569 0.09920627 0 0.006609255 1.018736 4371869376
39 | # 2 IR_039 273.2187 15.98847 -2.968737 276.0403 336.2159 4371869376
40 | # 3 IR_087 268.3977 17.49075 -0.1731693 271.9306 326.3914 4371869376
41 | # 4 IR_097 246.1366 10.81174 -0.05971194 246.4856 301.0066 4371869376
42 | # 5 IR_108 270.1535 18.49373 -0.6266653 274.0552 338.0375 4371869376
43 | # 6 IR_120 268.7993 18.42736 -0.4006808 272.9807 337.3713 4371869376
44 | # 7 IR_134 250.6491 11.70623 -0.5645727 252.9884 300.8559 4371869376
45 | # 8 VIS006 0.06711527 0.1101766 0 0.01692321 1.002381 4371869376
46 | # 9 VIS008 0.08736397 0.1326554 0 0.01656201 1.100475 4371869376
47 | # 10 WV_062 232.1964 5.531017 -2.086555 232.3866 260.9901 4371869376
48 | # 11 WV_073 248.0414 9.495061 -0.4933934 250.0049 289.8742 4371869376
49 | IR_016:
50 | map: [[inf, 0], [nan, 0]]
51 | range: [0, 1.02]
52 | mean_std: [0.06605569, 0.09920627]
53 | standardise: True
54 | IR_039:
55 | map: [[inf, 0], [nan, 0]]
56 | range: [0, 350]
57 | mean_std: [273.2187, 15.98847]
58 | standardise: True
59 | IR_087:
60 | map: [[inf, 0], [nan, 0]]
61 | range: [0, 350]
62 | mean_std: [268.3977, 17.49075]
63 | standardise: True
64 | IR_097:
65 | map: [[inf, 0], [nan, 0]]
66 | range: [0, 350]
67 | mean_std: [246.1366, 10.81174]
68 | standardise: True
69 | IR_108:
70 | map: [[inf, 0], [nan, 0]]
71 | range: [0, 350]
72 | mean_std: [270.1535, 18.49373]
73 | standardise: True
74 | IR_120:
75 | map: [[inf, 0], [nan, 0]]
76 | range: [0, 350]
77 | mean_std: [268.7993, 18.42736]
78 | standardise: True
79 | IR_134:
80 | map: [[inf, 0], [nan, 0]]
81 | range: [0, 350]
82 | mean_std: [250.6491, 11.70623]
83 | standardise: True
84 | VIS006:
85 | map: [[inf, 0], [nan, 0]]
86 | range: [0, 1.02]
87 | mean_std: [0.06711527, 0.1101766]
88 | standardise: True
89 | VIS008:
90 | map: [[inf, 0], [nan, 0]]
91 | range: [0, 1.2]
92 | mean_std: [0.08736397, 0.1326554]
93 | standardise: True
94 | WV_062:
95 | map: [[inf, 0], [nan, 0]]
96 | range: [0, 300]
97 | mean_std: [232.1964, 5.531017]
98 | standardise: True
99 | WV_073:
100 | map: [[inf, 0], [nan, 0]]
101 | range: [0, 300]
102 | mean_std: [248.0414, 9.495061]
103 | standardise: True
104 |
105 | train: # model training settings
106 | batch_size: 16 # 40
107 | max_epochs: 90
108 | n_workers: 8 # 16
109 | loss: BCEWithLogitsLoss # requires pos_weight to be set
110 | pos_weight: 6.114572
111 | early_stopping: True
112 | patience: 5
113 | lr: 1e-4
114 | weight_decay: 2e-2
115 | init_filter_size: 32
116 | dropout_rate: 0.4
117 |
118 | model: # model definition settings
119 | model_name: 3D_UNET_base
120 | in_channels: 11
121 | gradient_clip_val: 2.0
122 | gradient_clip_algorithm: value
123 |
124 | predict: # model prediction settings
125 | region_to_predict: boxi_0015 # must match one of the names defined in 'dataset' / 'regions'
126 | year_to_predict: 2019
127 |
--------------------------------------------------------------------------------
/models/configurations/config_baseline_stage2-pred.yaml:
--------------------------------------------------------------------------------
1 | experiment: # Settings for the current experiment
2 | name: 'U-NET-252' # name of the current experiment for logging
3 | experiment_folder: 'lightning_logs/' # folder to save logs
4 | sub_folder: 'baseline' # logging sub-folder
5 | precision: 32 # bit precision of weights. 32 or 16
6 | logging: True # Toggle logging on/off
7 | dataset:
8 | # Available bands: 'IR_016', 'IR_039', 'IR_087', 'IR_097', 'IR_108', 'IR_120', 'IR_134', 'VIS006', 'VIS008', 'WV_062', 'WV_073'
9 | sat_bands: ['IR_016', 'IR_039', 'IR_087', 'IR_097', 'IR_108', 'IR_120', 'IR_134', 'VIS006', 'VIS008', 'WV_062', 'WV_073']
10 | regions: [ '%REGION%' ]
11 | input_product: REFL-BT
12 | output_product: RATE
13 | out_channels: 1
14 | in_channels: 11
15 | swap_time_ch: True #swap the time and channels axis - if set to False (B, T, C, H, W)
16 | full_opera_context: 1512 # (105+42+105)*6;
17 | size_target_center: 252 # 42*6;
18 | len_seq_in: 4 # number of input slots per training sequence
19 | len_seq_predict: 32 # output slots; both still hard-coded in places!
20 | # data_root: 'data/2019' # stage-1
21 | data_root: 'data' # stage-2
22 | years: [ '%YEAR%' ]
23 | # splits_path: 'data/timestamps_and_splits_2019.csv' # stage-1
24 | splits_path: 'data/timestamps_and_splits_stage2.csv' # stage-2
25 |
26 | preprocess_OPERA:
27 | # # var mean sd min median max length pos.weight.thr_0
28 | # 1 RATE 0.07165331 0.6302647 0 0 127.9399 5011989696 6.114572
29 | RATE:
30 | rainfall_rate-500X500:
31 | mask: [-9999000.0, inf, nan, max128] #, 0range0.1] # Mostly used for loss function: values added here are added to a mask and not used for loss
32 | # map: [[lessthan0.0, 0], [greaterthan0.0, 1], [-8888000.0, 0], [-9999000.0, 0], [inf, 0], [nan, 0]] #Mostly used for input preprocessing # 1. map values # stage-1
33 | map: [[lessthan0.2, 0], [greaterthan0.2, 1], [0.2,1], [-8888000.0, 0], [-9999000.0, 0], [inf, 0], [nan, 0]] #Mostly used for input preprocessing # 1. map values # stage-2
34 | # mean_std: [0.07165331, 0.6302647]
35 | range: [0, 128] # 2. we evaluate only pixels in this range
36 | standardise: False # 3. use log(x+1) instead & normalize (x/max)
37 | bin: False
38 | preprocess_HRIT: # 1: map values, 2: normalise in range per variable if process==True
39 | # # var mean sd min median max length
40 | # 1 IR_016 0.06605569 0.09920627 0 0.006609255 1.018736 4371869376
41 | # 2 IR_039 273.2187 15.98847 -2.968737 276.0403 336.2159 4371869376
42 | # 3 IR_087 268.3977 17.49075 -0.1731693 271.9306 326.3914 4371869376
43 | # 4 IR_097 246.1366 10.81174 -0.05971194 246.4856 301.0066 4371869376
44 | # 5 IR_108 270.1535 18.49373 -0.6266653 274.0552 338.0375 4371869376
45 | # 6 IR_120 268.7993 18.42736 -0.4006808 272.9807 337.3713 4371869376
46 | # 7 IR_134 250.6491 11.70623 -0.5645727 252.9884 300.8559 4371869376
47 | # 8 VIS006 0.06711527 0.1101766 0 0.01692321 1.002381 4371869376
48 | # 9 VIS008 0.08736397 0.1326554 0 0.01656201 1.100475 4371869376
49 | # 10 WV_062 232.1964 5.531017 -2.086555 232.3866 260.9901 4371869376
50 | # 11 WV_073 248.0414 9.495061 -0.4933934 250.0049 289.8742 4371869376
51 | IR_016:
52 | map: [[inf, 0], [nan, 0]]
53 | range: [0, 1.02]
54 | mean_std: [0.06605569, 0.09920627]
55 | standardise: True
56 | IR_039:
57 | map: [[inf, 0], [nan, 0]]
58 | range: [0, 350]
59 | mean_std: [273.2187, 15.98847]
60 | standardise: True
61 | IR_087:
62 | map: [[inf, 0], [nan, 0]]
63 | range: [0, 350]
64 | mean_std: [268.3977, 17.49075]
65 | standardise: True
66 | IR_097:
67 | map: [[inf, 0], [nan, 0]]
68 | range: [0, 350]
69 | mean_std: [246.1366, 10.81174]
70 | standardise: True
71 | IR_108:
72 | map: [[inf, 0], [nan, 0]]
73 | range: [0, 350]
74 | mean_std: [270.1535, 18.49373]
75 | standardise: True
76 | IR_120:
77 | map: [[inf, 0], [nan, 0]]
78 | range: [0, 350]
79 | mean_std: [268.7993, 18.42736]
80 | standardise: True
81 | IR_134:
82 | map: [[inf, 0], [nan, 0]]
83 | range: [0, 350]
84 | mean_std: [250.6491, 11.70623]
85 | standardise: True
86 | VIS006:
87 | map: [[inf, 0], [nan, 0]]
88 | range: [0, 1.02]
89 | mean_std: [0.06711527, 0.1101766]
90 | standardise: True
91 | VIS008:
92 | map: [[inf, 0], [nan, 0]]
93 | range: [0, 1.2]
94 | mean_std: [0.08736397, 0.1326554]
95 | standardise: True
96 | WV_062:
97 | map: [[inf, 0], [nan, 0]]
98 | range: [0, 300]
99 | mean_std: [232.1964, 5.531017]
100 | standardise: True
101 | WV_073:
102 | map: [[inf, 0], [nan, 0]]
103 | range: [0, 300]
104 | mean_std: [248.0414, 9.495061]
105 | standardise: True
106 |
107 | train: # model training settings
108 | batch_size: 16 # 40
109 | max_epochs: 90
110 | n_workers: 8 # 16
111 | loss: BCEWithLogitsLoss # requires pos_weight to be set
112 | pos_weight: 2.577389 # stage-2: 2.577389 inverse unmasked rain ratio; 13.32784 inverse rain ratio. Only used for BCEWithLogitsLoss
113 | # loss: mIoULoss
114 | early_stopping: True
115 | patience: 20
116 | lr: 1e-4
117 | weight_decay: 2e-2
118 | init_filter_size: 32
119 | dropout_rate: 0.4
120 |
121 | model: # model definition settings
122 | model_name: 3D_UNET_base
123 | in_channels: 11
124 | gradient_clip_val: 2.0
125 | gradient_clip_algorithm: value
126 |
127 | predict: # model prediction settings
128 | submission_out_dir: %SDIR%
129 | region_to_predict: %REGION%
130 | year_to_predict: %YEAR%
131 |
--------------------------------------------------------------------------------
/models/configurations/config_baseline_stage2.yaml:
--------------------------------------------------------------------------------
1 | experiment: # Settings for the current experiment
2 | name: 'U-NET-252' # name of the current experiment for logging
3 | experiment_folder: 'lightning_logs/' # folder to save logs
4 | sub_folder: 'baseline' # logging sub-folder
5 | precision: 32 # bit precision of weights. 32 or 16
6 | logging: True # Toggle logging on/off
7 | dataset:
8 | # Available bands: 'IR_016', 'IR_039', 'IR_087', 'IR_097', 'IR_108', 'IR_120', 'IR_134', 'VIS006', 'VIS008', 'WV_062', 'WV_073'
9 | sat_bands: ['IR_016', 'IR_039', 'IR_087', 'IR_097', 'IR_108', 'IR_120', 'IR_134', 'VIS006', 'VIS008', 'WV_062', 'WV_073']
10 | # regions: [ 'boxi_0015', 'boxi_0034', 'boxi_0076' ] # stage-1
11 | regions: [ 'boxi_0015', 'boxi_0034', 'boxi_0076',
12 | 'roxi_0004', 'roxi_0005', 'roxi_0006', 'roxi_0007' ] # stage-2
13 | # or, e.g., ['boxi_0015'] if you want to train your model for one region only
14 | input_product: REFL-BT
15 | output_product: RATE
16 | out_channels: 1
17 | in_channels: 11
18 | swap_time_ch: True #swap the time and channels axis - if set to False (B, T, C, H, W)
19 | full_opera_context: 1512 # (105+42+105)*6;
20 | size_target_center: 252 # 42*6;
21 | len_seq_in: 4 # number of input slots per training sequence
22 | len_seq_predict: 32 # output slots; both still hard-coded in places!
23 | # data_root: 'data/2019' # stage-1
24 | data_root: 'data'
25 | years: [ '2019', '2020' ] # stage-2
26 | # years: [ '2019' ] # stage-1
27 | # splits_path: 'data/timestamps_and_splits_2019.csv' # stage-1
28 | splits_path: 'data/timestamps_and_splits_stage2.csv' # stage-2
29 |
30 | preprocess_OPERA:
31 | # # var mean sd min median max length pos.weight.thr_0
32 | # 1 RATE 0.07165331 0.6302647 0 0 127.9399 5011989696 6.114572
33 | RATE:
34 | rainfall_rate-500X500:
35 | mask: [-9999000.0, inf, nan, max128] #, 0range0.1] # Mostly used for loss function: values added here are added to a mask and not used for loss
36 | # map: [[lessthan0.0, 0], [greaterthan0.0, 1], [-8888000.0, 0], [-9999000.0, 0], [inf, 0], [nan, 0]] #Mostly used for input preprocessing # 1. map values # stage-1
37 | map: [[lessthan0.2, 0], [greaterthan0.2, 1], [0.2,1], [-8888000.0, 0], [-9999000.0, 0], [inf, 0], [nan, 0]] #Mostly used for input preprocessing # 1. map values # stage-2
38 | # mean_std: [0.07165331, 0.6302647]
39 | range: [0, 128] # 2. we evaluate only pixels in this range
40 | standardise: False # 3. use log(x+1) instead & normalize (x/max)
41 | bin: False
42 | preprocess_HRIT: # 1: map values, 2: normalise in range per variable if process==True
43 | # # var mean sd min median max length
44 | # 1 IR_016 0.06605569 0.09920627 0 0.006609255 1.018736 4371869376
45 | # 2 IR_039 273.2187 15.98847 -2.968737 276.0403 336.2159 4371869376
46 | # 3 IR_087 268.3977 17.49075 -0.1731693 271.9306 326.3914 4371869376
47 | # 4 IR_097 246.1366 10.81174 -0.05971194 246.4856 301.0066 4371869376
48 | # 5 IR_108 270.1535 18.49373 -0.6266653 274.0552 338.0375 4371869376
49 | # 6 IR_120 268.7993 18.42736 -0.4006808 272.9807 337.3713 4371869376
50 | # 7 IR_134 250.6491 11.70623 -0.5645727 252.9884 300.8559 4371869376
51 | # 8 VIS006 0.06711527 0.1101766 0 0.01692321 1.002381 4371869376
52 | # 9 VIS008 0.08736397 0.1326554 0 0.01656201 1.100475 4371869376
53 | # 10 WV_062 232.1964 5.531017 -2.086555 232.3866 260.9901 4371869376
54 | # 11 WV_073 248.0414 9.495061 -0.4933934 250.0049 289.8742 4371869376
55 | IR_016:
56 | map: [[inf, 0], [nan, 0]]
57 | range: [0, 1.02]
58 | mean_std: [0.06605569, 0.09920627]
59 | standardise: True
60 | IR_039:
61 | map: [[inf, 0], [nan, 0]]
62 | range: [0, 350]
63 | mean_std: [273.2187, 15.98847]
64 | standardise: True
65 | IR_087:
66 | map: [[inf, 0], [nan, 0]]
67 | range: [0, 350]
68 | mean_std: [268.3977, 17.49075]
69 | standardise: True
70 | IR_097:
71 | map: [[inf, 0], [nan, 0]]
72 | range: [0, 350]
73 | mean_std: [246.1366, 10.81174]
74 | standardise: True
75 | IR_108:
76 | map: [[inf, 0], [nan, 0]]
77 | range: [0, 350]
78 | mean_std: [270.1535, 18.49373]
79 | standardise: True
80 | IR_120:
81 | map: [[inf, 0], [nan, 0]]
82 | range: [0, 350]
83 | mean_std: [268.7993, 18.42736]
84 | standardise: True
85 | IR_134:
86 | map: [[inf, 0], [nan, 0]]
87 | range: [0, 350]
88 | mean_std: [250.6491, 11.70623]
89 | standardise: True
90 | VIS006:
91 | map: [[inf, 0], [nan, 0]]
92 | range: [0, 1.02]
93 | mean_std: [0.06711527, 0.1101766]
94 | standardise: True
95 | VIS008:
96 | map: [[inf, 0], [nan, 0]]
97 | range: [0, 1.2]
98 | mean_std: [0.08736397, 0.1326554]
99 | standardise: True
100 | WV_062:
101 | map: [[inf, 0], [nan, 0]]
102 | range: [0, 300]
103 | mean_std: [232.1964, 5.531017]
104 | standardise: True
105 | WV_073:
106 | map: [[inf, 0], [nan, 0]]
107 | range: [0, 300]
108 | mean_std: [248.0414, 9.495061]
109 | standardise: True
110 |
111 | train: # model training settings
112 | batch_size: 40 # 16
113 | max_epochs: 90
114 | n_workers: 16 # 8
115 | loss: BCEWithLogitsLoss # requires pos_weight to be set
116 | pos_weight: 2.577389 # stage-2: 2.577389 inverse unmasked rain ratio; 13.32784 inverse rain ratio. Only used for BCEWithLogitsLoss
117 | # loss: mIoULoss
118 | early_stopping: True
119 | patience: 20
120 | lr: 1e-4
121 | weight_decay: 2e-2
122 | init_filter_size: 32
123 | dropout_rate: 0.4
124 |
125 | model: # model definition settings
126 | model_name: 3D_UNET_base
127 | in_channels: 11
128 | gradient_clip_val: 2.0
129 | gradient_clip_algorithm: value
130 |
131 | predict: # model prediction settings
132 | region_to_predict: boxi_0015 # must match one of the names defined in 'dataset' / 'regions'
133 | year_to_predict: 2019
134 |
--------------------------------------------------------------------------------
/models/unet.patch:
--------------------------------------------------------------------------------
1 | --- unet.py 2022-08-20 12:22:39.713834077 +0200
2 | +++ baseline_UNET3D.py 2022-08-20 12:22:03.482141847 +0200
3 | @@ -1,35 +1,32 @@
4 | -# ELEKTRONN3 - Neural Network Toolkit
5 | +# Weather4cast 2022 Starter Kit
6 | #
7 | -# Copyright (c) 2017 - now
8 | -# Max Planck Institute of Neurobiology, Munich, Germany
9 | -# Author: Martin Drawitsch
10 | -
11 | -"""
12 | -This is a modified version of the U-Net CNN architecture for biomedical
13 | -image segmentation. U-Net was originally published in
14 | -https://arxiv.org/abs/1505.04597 by Ronneberger et al.
15 | -
16 | -A pure-3D variant of U-Net has been proposed by Çiçek et al.
17 | -in https://arxiv.org/abs/1606.06650, but the below implementation
18 | -is based on the original U-Net paper, with several improvements.
19 | -
20 | -This code is based on https://github.com/jaxony/unet-pytorch
21 | -(c) 2017 Jackson Huang, released under MIT License,
22 | -which implements (2D) U-Net with user-defined network depth
23 | -and a few other improvements of the original architecture.
24 | -
25 | -Major differences of this version from Huang's code:
26 | -
27 | -- Operates on 3D image data (5D tensors) instead of 2D data
28 | -- Uses 3D convolution, 3D pooling etc. by default
29 | -- planar_blocks architecture parameter for mixed 2D/3D convnets
30 | - (see UNet class docstring for details)
31 | -- Improved tests (see the bottom of the file)
32 | -- Cleaned up parameter/variable names and formatting, changed default params
33 | -- Updated for PyTorch 1.3 and Python 3.6 (earlier versions unsupported)
34 | -- (Optional DEBUG mode for optional printing of debug information)
35 | -- Extended documentation
36 | -"""
37 | +# Copyright (C) 2022
38 | +# Institute of Advanced Research in Artificial Intelligence (IARAI)
39 | +
40 | +# Baseline model for the Neurips 2022 Weather4cast Competition -
41 | +# an adaptation of the CNN ELEKTRONN3 model available under MIT license on
42 | +# https://raw.githubusercontent.com/ELEKTRONN/elektronn3/f754796d861f1cfe1c19dfc7819087972573ce40/elektronn3/models/unet.py
43 | +#
44 | +# The adaptations are part of the Weather4cast 2022 Starter Kit.
45 | +
46 | +# The Weather4cast 2022 Starter Kit is free software: you can redistribute it
47 | +# and/or modify it under the terms of the GNU General Public License as
48 | +# published by the Free Software Foundation, either version 3 of the License,
49 | +# or (at your option) any later version.
50 | +#
51 | +# The Weather4cast 2022 Starter Kit is distributed in the hope that it will be
52 | +# useful, but WITHOUT ANY WARRANTY; without even the implied warranty of
53 | +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
54 | +# GNU General Public License for more details.
55 | +#
56 | +# You should have received a copy of the GNU General Public License
57 | +# along with this program. If not, see .
58 | +
59 | +# Contributors: Aleksandra Gruca, Pedro Herruzo, David Kreil, Stephen Moran
60 | +
61 | +
62 | +VERBOSE=False
63 | +##VERBOSE=True
64 |
65 | __all__ = ['UNet']
66 |
67 | @@ -204,12 +201,13 @@
68 | A helper Module that performs 2 convolutions and 1 MaxPool.
69 | A ReLU activation follows each convolution.
70 | """
71 | - def __init__(self, in_channels, out_channels, pooling=True, planar=False, activation='relu',
72 | + def __init__(self, in_channels, out_channels, dropout_rate, pooling=True, planar=False, activation='relu',
73 | normalization=None, full_norm=True, dim=3, conv_mode='same'):
74 | super().__init__()
75 |
76 | self.in_channels = in_channels
77 | self.out_channels = out_channels
78 | + self.dropout_rate = dropout_rate
79 | self.pooling = pooling
80 | self.normalization = normalization
81 | self.dim = dim
82 | @@ -232,21 +230,28 @@
83 | self.pool = nn.Identity()
84 | self.pool_ks = -123 # Bogus value, will never be read. Only to satisfy TorchScript's static type system
85 |
86 | + self.dropout = nn.Dropout3d(dropout_rate)
87 | +
88 | self.act1 = get_activation(activation)
89 | self.act2 = get_activation(activation)
90 |
91 | if full_norm:
92 | self.norm0 = get_normalization(normalization, self.out_channels, dim=dim)
93 | + if VERBOSE: print("DownConv, full_norm, norm0 =",normalization)
94 | else:
95 | self.norm0 = nn.Identity()
96 | + if VERBOSE: print("DownConv, no full_norm")
97 | self.norm1 = get_normalization(normalization, self.out_channels, dim=dim)
98 | + if VERBOSE: print("DownConv, norm1 =",normalization)
99 |
100 | def forward(self, x):
101 | y = self.conv1(x)
102 | y = self.norm0(y)
103 | + y = self.dropout(y)
104 | y = self.act1(y)
105 | y = self.conv2(y)
106 | y = self.norm1(y)
107 | + y = self.dropout(y)
108 | y = self.act2(y)
109 | before_pool = y
110 | y = self.pool(y)
111 | @@ -754,9 +759,10 @@
112 | """
113 | def __init__(
114 | self,
115 | - in_channels: int = 1,
116 | - out_channels: int = 2,
117 | - n_blocks: int = 3,
118 | + in_channels: int = 11,
119 | + out_channels: int = 32, ## NEW: number of time slots to predict
120 | + dropout_rate: float = 0.4,
121 | + n_blocks: int = 5,
122 | start_filts: int = 32,
123 | up_mode: str = 'transpose',
124 | merge_mode: str = 'concat',
125 | @@ -815,6 +821,7 @@
126 |
127 | self.out_channels = out_channels
128 | self.in_channels = in_channels
129 | + self.dropout_rate = dropout_rate
130 | self.start_filts = start_filts
131 | self.n_blocks = n_blocks
132 | self.normalization = normalization
133 | @@ -846,8 +853,9 @@
134 | down_conv = DownConv(
135 | ins,
136 | outs,
137 | + dropout_rate,
138 | pooling=pooling,
139 | - planar=planar,
140 | + planar=planar,
141 | activation=activation,
142 | normalization=normalization,
143 | full_norm=full_norm,
144 | @@ -877,9 +885,9 @@
145 | conv_mode=conv_mode,
146 | )
147 | self.up_convs.append(up_conv)
148 | -
149 | - self.conv_final = conv1(outs, self.out_channels, dim=dim)
150 | -
151 | + self.reduce_channels = conv1(outs*4, ## 4 = experiment / len_seq_in
152 | + self.out_channels, dim=dim)
153 | + self.dropout = nn.Dropout3d(dropout_rate)
154 | self.apply(self.weight_init)
155 |
156 | @staticmethod
157 | @@ -898,9 +906,11 @@
158 | i = 0 # Can't enumerate because of https://github.com/pytorch/pytorch/issues/16123
159 | for module in self.down_convs:
160 | x, before_pool = module(x)
161 | + before_pool = self.dropout(before_pool) # for skip connections
162 | encoder_outs.append(before_pool)
163 | i += 1
164 |
165 | + x = self.dropout(x) # at bottom of the U, as in the original U-Net
166 | # Decoding by UpConv and merging with saved outputs of encoder
167 | i = 0
168 | for module in self.up_convs:
169 | @@ -909,8 +919,16 @@
170 | i += 1
171 |
172 | # No softmax is used, so you need to apply it in the loss.
173 | - x = self.conv_final(x)
174 | - # Uncomment the following line to temporarily store output for
175 | + if VERBOSE: print("pre-reshape",x.shape)
176 | + xs = x.shape;
177 | + x = torch.reshape(x,(xs[0],xs[1]*xs[2],1,xs[3],xs[4]));
178 | + if VERBOSE: print("pre-reduce",x.shape)
179 | + x = self.reduce_channels(x)
180 | + if VERBOSE: print("post-reduce",x.shape)
181 | + xs = x.shape;
182 | + x = torch.reshape(x,(xs[0],1,xs[1],xs[3],xs[4]));
183 | + if VERBOSE: print("post-reshape",x.shape)
184 | + # Uncomment the following line to temporarily store output for
185 | # receptive field estimation using fornoxai/receptivefield:
186 | # self.feature_maps = [x] # Currently disabled to save memory
187 | return x
188 |
--------------------------------------------------------------------------------
/utils/w4c_dataloader.py:
--------------------------------------------------------------------------------
1 | # Weather4cast 2022 Starter Kit
2 | #
3 | # Copyright (C) 2022
4 | # Institute of Advanced Research in Artificial Intelligence (IARAI)
5 |
6 | # This file is part of the Weather4cast 2022 Starter Kit.
7 | #
8 | # The Weather4cast 2022 Starter Kit is free software: you can redistribute it
9 | # and/or modify it under the terms of the GNU General Public License as
10 | # published by the Free Software Foundation, either version 3 of the License,
11 | # or (at your option) any later version.
12 | #
13 | # The Weather4cast 2022 Starter Kit is distributed in the hope that it will be
14 | # useful, but WITHOUT ANY WARRANTY; without even the implied warranty of
15 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 | # GNU General Public License for more details.
17 | #
18 | # You should have received a copy of the GNU General Public License
19 | # along with this program. If not, see .
20 |
21 | # Contributors: Aleksandra Gruca, Pedro Herruzo, David Kreil, Stephen Moran
22 |
23 |
24 | # from cv2 import CAP_PROP_XI_ACQ_TRANSPORT_BUFFER_COMMIT
25 | import numpy as np
26 | from torch.utils.data import Dataset
27 | import os
28 | import sys
29 | import time
30 | from timeit import default_timer as timer
31 |
32 | from utils.data_utils import *
33 |
34 | # folder to load config file
35 | # CONFIG_PATH = "../"
36 |
37 | # VERBOSE = True
38 | VERBOSE = False
39 |
40 | """
41 | Assumptions:
42 | - the data is already cropped to the right dimensions
43 | - Data format - [Samples, C, T, W, H]
44 | """
45 |
46 |
47 | class RainData(Dataset):
48 | def __init__(
49 | self,
50 | data_split,
51 | project_root="",
52 | data_root="",
53 | input_product="REFL-BT",
54 | compute_seq=True,
55 | output_product="RATE",
56 | sat_bands=[],
57 | preprocess_OPERA=None,
58 | size_target_center=None,
59 | full_opera_context=None,
60 | preprocess_HRIT=None,
61 | path_to_sample_ids="",
62 | len_seq_in=4,
63 | len_seq_predict=32,
64 | regions=["boxi_0015"],
65 | regions_def={},
66 | generate_samples=False,
67 | latlon_path="",
68 | altitudes_path="",
69 | splits_path=None,
70 | swap_time_ch=False,
71 | years=None,
72 | **kwargs
73 | ):
74 | start = timer()
75 | # Data Dimensions
76 | self.len_seq_in = len_seq_in
77 | self.len_seq_predict = len_seq_predict
78 | self.channel_dim = 1 # where to concat channels in structure
79 |
80 | # type of data & processing variables
81 | self.sat_bands = sat_bands
82 | self.regions = regions
83 | self.input_product = input_product
84 | self.output_product = output_product
85 | self.preprocess_target = preprocess_OPERA
86 | self.size_target_center = size_target_center
87 | self.full_opera_context = full_opera_context
88 | self.crop = int(
89 | (self.full_opera_context - self.size_target_center) / 2
90 | ) # calculate centre of image to begin crop
91 | self.preprocess_input = preprocess_HRIT
92 | self.path_to_sample_ids = path_to_sample_ids
93 | self.regions_def = regions_def
94 | self.generate_samples = generate_samples
95 | self.path_to_sample_ids = path_to_sample_ids
96 | self.swap_time_ch = swap_time_ch
97 | self.years = years
98 |
99 | # data splits to load (training/validation/test)
100 | self.root = project_root
101 | self.data_root = data_root
102 | self.data_split = data_split
103 | self.splits_df = load_timestamps(splits_path)
104 | # prepare all elements to load - sample idx will use the object 'self.idx'
105 | self.idxs = load_sample_ids(
106 | self.data_split,
107 | self.splits_df,
108 | self.len_seq_in,
109 | self.len_seq_predict,
110 | self.regions,
111 | self.generate_samples,
112 | self.years,
113 | self.path_to_sample_ids
114 | )
115 |
116 | # LOAD DATASET
117 | self.in_ds = load_dataset(
118 | self.data_root, self.data_split, self.regions, years, self.input_product
119 | )
120 | if self.data_split not in ["test", "heldout"]:
121 | self.out_ds = load_dataset(
122 | self.data_root, self.data_split, self.regions, years, self.output_product
123 | )
124 | else:
125 | self.out_ds = []
126 |
127 | def __len__(self):
128 | """total number of samples (sequences of in:4-out:1 in our case) to train"""
129 | # print(len(self.idxs), "-------------------", self.data_split)
130 | return len(self.idxs)
131 |
132 | def load_in(self, in_seq, seq_r, metadata, loaded_input=False):
133 | in0 = time.time()
134 | input_data, in_masks = get_sequence(
135 | in_seq,
136 | self.data_root,
137 | self.data_split,
138 | seq_r,
139 | self.input_product,
140 | self.sat_bands,
141 | self.preprocess_input,
142 | self.swap_time_ch,
143 | self.in_ds,
144 | )
145 |
146 | if VERBOSE:
147 | print(np.shape(input_data), time.time() - in0, "in sequence time")
148 | return input_data, metadata
149 |
150 | def load_out(self, out_seq, seq_r, metadata):
151 | t1 = time.time()
152 | # GROUND TRUTH (OUTPUT)
153 | if self.data_split not in ["test", "heldout"]:
154 | output_data, out_masks = get_sequence(
155 | out_seq,
156 | self.data_root,
157 | self.data_split,
158 | seq_r,
159 | self.output_product,
160 | [],
161 | self.preprocess_target,
162 | self.swap_time_ch,
163 | self.out_ds,
164 | )
165 |
166 | # collapse time to channels
167 | metadata["target"]["mask"] = out_masks
168 | else: # Just return [] if its test/heldout data
169 | output_data = np.array([])
170 | if VERBOSE:
171 | print(time.time() - t1, "out sequence")
172 | return output_data, metadata
173 |
174 | def load_in_out(self, in_seq, out_seq=None, seq_r=None):
175 | metadata = {
176 | "input": {"mask": [], "timestamps": in_seq},
177 | "target": {"mask": [], "timestamps": out_seq},
178 | }
179 |
180 | t0 = time.time()
181 | input_data, metadata = self.load_in(in_seq, seq_r, metadata)
182 | output_data, metadata = self.load_out(out_seq, seq_r, metadata)
183 |
184 | if VERBOSE:
185 | print(time.time() - t0, "seconds")
186 | return input_data, output_data, metadata
187 |
188 | def __getitem__(self, idx):
189 | """load 1 sequence (1 sample)"""
190 | in_seq = self.idxs[idx][0]
191 | out_seq = self.idxs[idx][1]
192 | seq_r = self.idxs[idx][2]
193 | # # print("=== DEBUG in_seq: ",in_seq, file=sys.stderr);
194 | # print("=== DEBUG in_seq: ",in_seq);
195 | return self.load_in_out(in_seq, out_seq, seq_r)
196 |
197 |
198 | class Normalise(object):
199 | """Dataset Transform: "Normalise values for each band."""
200 |
201 | def __init__(self, mean, std):
202 | """Normalise values for each band
203 | Args:
204 | mean (list): mean value of bands
205 | std (list): standard deviation of bands
206 | """
207 | self.mean = mean
208 | self.std = std
209 | super().__init__()
210 |
211 | def __call__(self, sample):
212 | """Normalise values for each band
213 | Args:
214 | sample (Tensor, Tensor): sample and labels for sample as tensor
215 | Returns:
216 | sample (Tensor, Tensor): sample and labels for sample normalized
217 | """
218 | data, labels = sample
219 | # For every channel, subtract the mean, and divide by the standard deviation\
220 | # possible approach = loop through band dim and access right values in corresponding dims of mean / stdev
221 | for t, m, s in zip(data, self.mean, self.std):
222 | t.sub_(m).div_(s)
223 | return (data, labels)
224 |
--------------------------------------------------------------------------------
/train.py:
--------------------------------------------------------------------------------
1 | # Weather4cast 2022 Starter Kit
2 | #
3 | # Copyright (C) 2022
4 | # Institute of Advanced Research in Artificial Intelligence (IARAI)
5 |
6 | # This file is part of the Weather4cast 2022 Starter Kit.
7 | #
8 | # The Weather4cast 2022 Starter Kit is free software: you can redistribute it
9 | # and/or modify it under the terms of the GNU General Public License as
10 | # published by the Free Software Foundation, either version 3 of the License,
11 | # or (at your option) any later version.
12 | #
13 | # The Weather4cast 2022 Starter Kit is distributed in the hope that it will be
14 | # useful, but WITHOUT ANY WARRANTY; without even the implied warranty of
15 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 | # GNU General Public License for more details.
17 | #
18 | # You should have received a copy of the GNU General Public License
19 | # along with this program. If not, see .
20 |
21 | # Contributors: Aleksandra Gruca, Pedro Herruzo, David Kreil, Stephen Moran
22 |
23 |
24 | import argparse
25 |
26 | import pytorch_lightning as pl
27 | from pytorch_lightning.callbacks import ModelCheckpoint
28 | from pytorch_lightning.plugins import DDPPlugin
29 | from torch.utils.data import DataLoader
30 | from pytorch_lightning import loggers as pl_loggers
31 | from pytorch_lightning.callbacks.early_stopping import EarlyStopping
32 | import datetime
33 | import os
34 | import torch
35 |
36 | from models.unet_lightning import UNet_Lightning as UNetModel
37 | from utils.data_utils import load_config
38 | from utils.data_utils import get_cuda_memory_usage
39 | from utils.data_utils import tensor_to_submission_file
40 | from utils.w4c_dataloader import RainData
41 |
42 | class DataModule(pl.LightningDataModule):
43 | """ Class to handle training/validation splits in a single object
44 | """
45 | def __init__(self, params, training_params, mode):
46 | super().__init__()
47 | self.params = params
48 | self.training_params = training_params
49 | if mode in ['train']:
50 | print("Loading TRAINING/VALIDATION dataset -- as test")
51 | self.train_ds = RainData('training', **self.params)
52 | self.val_ds = RainData('validation', **self.params)
53 | print(f"Training dataset size: {len(self.train_ds)}")
54 | if mode in ['val']:
55 | print("Loading VALIDATION dataset -- as test")
56 | self.val_ds = RainData('validation', **self.params)
57 | if mode in ['predict']:
58 | print("Loading PREDICTION/TEST dataset -- as test")
59 | self.test_ds = RainData('test', **self.params)
60 | if mode in ['heldout']:
61 | print("Loading HELD-OUT dataset -- as test")
62 | self.test_ds = RainData('heldout', **self.params)
63 |
64 | def __load_dataloader(self, dataset, shuffle=True, pin=True):
65 | dl = DataLoader(dataset,
66 | batch_size=self.training_params['batch_size'],
67 | num_workers=self.training_params['n_workers'],
68 | shuffle=shuffle,
69 | pin_memory=pin, prefetch_factor=2,
70 | persistent_workers=False)
71 | return dl
72 |
73 | def train_dataloader(self):
74 | return self.__load_dataloader(self.train_ds, shuffle=True, pin=True)
75 |
76 | def val_dataloader(self):
77 | return self.__load_dataloader(self.val_ds, shuffle=False, pin=True)
78 |
79 | def test_dataloader(self):
80 | return self.__load_dataloader(self.test_ds, shuffle=False, pin=True)
81 |
82 |
83 | def load_model(Model, params, checkpoint_path=''):
84 | """ loads a model from a checkpoint or from scratch if checkpoint_path='' """
85 | p = {**params['experiment'], **params['dataset'], **params['train']}
86 | if checkpoint_path == '':
87 | print('-> Modelling from scratch! (no checkpoint loaded)')
88 | model = Model(params['model'], p)
89 | else:
90 | print(f'-> Loading model checkpoint: {checkpoint_path}')
91 | model = Model.load_from_checkpoint(checkpoint_path, UNet_params=params['model'], params=p)
92 | return model
93 |
94 | def get_trainer(gpus,params):
95 | """ get the trainer, modify here its options:
96 | - save_top_k
97 | """
98 | max_epochs=params['train']['max_epochs'];
99 | print("Trainig for",max_epochs,"epochs");
100 | checkpoint_callback = ModelCheckpoint(monitor='val_loss_epoch', save_top_k=3, save_last=True,
101 | filename='{epoch:02d}-{val_loss_epoch:.6f}')
102 |
103 | parallel_training = None
104 | ddpplugin = None
105 | if gpus[0] == -1:
106 | gpus = None
107 | elif len(gpus) > 1:
108 | parallel_training = 'ddp'
109 | ## ddpplugin = DDPPlugin(find_unused_parameters=True)
110 | print(f"====== process started on the following GPUs: {gpus} ======")
111 | date_time = datetime.datetime.now().strftime("%m%d-%H:%M")
112 | version = params['experiment']['name']
113 | version = version + '_' + date_time
114 |
115 | #SET LOGGER
116 | if params['experiment']['logging']:
117 | tb_logger = pl_loggers.TensorBoardLogger(save_dir=params['experiment']['experiment_folder'],name=params['experiment']['sub_folder'], version=version, log_graph=True)
118 | else:
119 | tb_logger = False
120 |
121 | if params['train']['early_stopping']:
122 | early_stop_callback = EarlyStopping(monitor="val_loss_epoch",
123 | patience=params['train']['patience'],
124 | mode="min")
125 | callback_funcs = [checkpoint_callback, early_stop_callback]
126 | else:
127 | callback_funcs = [checkpoint_callback]
128 |
129 | trainer = pl.Trainer(devices=gpus, max_epochs=max_epochs,
130 | gradient_clip_val=params['model']['gradient_clip_val'],
131 | gradient_clip_algorithm=params['model']['gradient_clip_algorithm'],
132 | accelerator="gpu",
133 | callbacks=callback_funcs,logger=tb_logger,
134 | profiler='simple',precision=params['experiment']['precision'],
135 | strategy="ddp"
136 | )
137 |
138 | return trainer
139 |
140 | def do_predict(trainer, model, predict_params, test_data):
141 | scores = trainer.predict(model, dataloaders=test_data)
142 | scores = torch.concat(scores)
143 | tensor_to_submission_file(scores,predict_params)
144 |
145 | def do_test(trainer, model, test_data):
146 | scores = trainer.test(model, dataloaders=test_data)
147 |
148 | def train(params, gpus, mode, checkpoint_path, model=UNetModel):
149 | """ main training/evaluation method
150 | """
151 | # ------------
152 | # model & data
153 | # ------------
154 | get_cuda_memory_usage(gpus)
155 | data = DataModule(params['dataset'], params['train'], mode)
156 | model = load_model(model, params, checkpoint_path)
157 | # ------------
158 | # Add your models here
159 | # ------------
160 |
161 | # ------------
162 | # trainer
163 | # ------------
164 | trainer = get_trainer(gpus, params)
165 | get_cuda_memory_usage(gpus)
166 | # ------------
167 | # train & final validation
168 | # ------------
169 | if mode == 'train':
170 | print("------------------")
171 | print("--- TRAIN MODE ---")
172 | print("------------------")
173 | trainer.fit(model, data)
174 |
175 |
176 | if mode == "val":
177 | # ------------
178 | # VALIDATE
179 | # ------------
180 | print("---------------------")
181 | print("--- VALIDATE MODE ---")
182 | print("---------------------")
183 | do_test(trainer, model, data.val_dataloader())
184 |
185 |
186 | if mode == 'predict' or mode == 'heldout':
187 | # ------------
188 | # PREDICT
189 | # ------------
190 | print("--------------------")
191 | print("--- PREDICT MODE ---")
192 | print("--------------------")
193 | print("REGIONS!:: ", params["dataset"]["regions"], params["predict"]["region_to_predict"])
194 | if params["predict"]["region_to_predict"] not in params["dataset"]["regions"]:
195 | print("EXITING... \"regions\" and \"regions to predict\" must indicate the same region name in your config file.")
196 | else:
197 | do_predict(trainer, model, params["predict"], data.test_dataloader())
198 |
199 | get_cuda_memory_usage(gpus)
200 |
201 | def update_params_based_on_args(options):
202 | config_p = os.path.join('models/configurations',options.config_path)
203 | params = load_config(config_p)
204 |
205 | if options.name != '':
206 | print(params['experiment']['name'])
207 | params['experiment']['name'] = options.name
208 | return params
209 |
210 | def set_parser():
211 | """ set custom parser """
212 |
213 | parser = argparse.ArgumentParser(description="")
214 | parser.add_argument("-f", "--config_path", type=str, required=False, default='./configurations/config_basline.yaml',
215 | help="path to config-yaml")
216 | parser.add_argument("-g", "--gpus", type=int, nargs='+', required=False, default=1,
217 | help="specify gpu(s): 1 or 1 5 or 0 1 2 (-1 for no gpu)")
218 | parser.add_argument("-m", "--mode", type=str, required=False, default='train',
219 | help="choose mode: train (default) / val / predict")
220 | parser.add_argument("-c", "--checkpoint", type=str, required=False, default='',
221 | help="init a model from a checkpoint path. '' as default (random weights)")
222 | parser.add_argument("-n", "--name", type=str, required=False, default='',
223 | help="Set the name of the experiment")
224 |
225 | return parser
226 |
227 | def main():
228 | parser = set_parser()
229 | options = parser.parse_args()
230 |
231 | params = update_params_based_on_args(options)
232 | train(params, options.gpus, options.mode, options.checkpoint)
233 |
234 | if __name__ == "__main__":
235 | main()
236 | """ examples of usage:
237 |
238 | 1) train from scratch on one GPU
239 | python train.py --gpus 2 --mode train --config_path config_baseline.yaml --name baseline_train
240 |
241 | 2) train from scratch on four GPUs
242 | python train.py --gpus 0 1 2 3 --mode train --config_path config_baseline.yaml --name baseline_train
243 |
244 | 3) fine tune a model from a checkpoint on one GPU
245 | python train.py --gpus 1 --mode train --config_path config_baseline.yaml --checkpoint "lightning_logs/PATH-TO-YOUR-MODEL-LOGS/checkpoints/YOUR-CHECKPOINT-FILENAME.ckpt" --name baseline_tune
246 |
247 | 4) evaluate a trained model from a checkpoint on two GPUs
248 | python train.py --gpus 0 1 --mode val --config_path config_baseline.yaml --checkpoint "lightning_logs/PATH-TO-YOUR-MODEL-LOGS/checkpoints/YOUR-CHECKPOINT-FILENAME.ckpt" --name baseline_validate
249 |
250 | 5) generate predictions (plese note that this mode works only for one GPU)
251 | python train.py --gpus 1 --mode predict --config_path config_baseline.yaml --checkpoint "lightning_logs/PATH-TO-YOUR-MODEL-LOGS/checkpoints/YOUR-CHECKPOINT-FILENAME.ckpt"
252 |
253 | 6) generate predictions for the held-out dataset (plese note that this mode works only for one GPU)
254 | python train.py --gpus 1 --mode heldout --config_path config_baseline.yaml --checkpoint "lightning_logs/PATH-TO-YOUR-MODEL-LOGS/checkpoints/YOUR-CHECKPOINT-FILENAME.ckpt"
255 |
256 | """
257 |
--------------------------------------------------------------------------------
/w4c.yml:
--------------------------------------------------------------------------------
1 | name: w4c
2 | channels:
3 | - pytorch
4 | - anaconda
5 | - conda-forge
6 | - defaults
7 | dependencies:
8 | - _libgcc_mutex=0.1=conda_forge
9 | - _openmp_mutex=4.5=2_kmp_llvm
10 | - absl-py=1.0.0=pyhd8ed1ab_0
11 | - aiohttp=3.8.1=py38h0a891b7_1
12 | - aiosignal=1.2.0=pyhd8ed1ab_0
13 | - alsa-lib=1.2.3.2=h166bdaf_0
14 | - anyio=3.5.0=py38h578d9bd_0
15 | - aom=3.3.0=h27087fc_1
16 | - argon2-cffi=21.3.0=pyhd8ed1ab_0
17 | - argon2-cffi-bindings=21.2.0=py38h0a891b7_2
18 | - asttokens=2.0.5=pyhd8ed1ab_0
19 | - async-timeout=4.0.2=pyhd8ed1ab_0
20 | - attr=2.5.1=h166bdaf_0
21 | - attrs=21.4.0=pyhd8ed1ab_0
22 | - babel=2.10.1=pyhd8ed1ab_0
23 | - backcall=0.2.0=pyh9f0ad1d_0
24 | - backports=1.0=py_2
25 | - backports.functools_lru_cache=1.6.4=pyhd8ed1ab_0
26 | - beautifulsoup4=4.11.1=pyha770c72_0
27 | - bleach=5.0.0=pyhd8ed1ab_0
28 | - blinker=1.4=py_1
29 | - brotli=1.0.9=h166bdaf_7
30 | - brotli-bin=1.0.9=h166bdaf_7
31 | - brotlipy=0.7.0=py38h0a891b7_1004
32 | - bzip2=1.0.8=h7f98852_4
33 | - c-ares=1.18.1=h7f98852_0
34 | - ca-certificates=2022.4.26=h06a4308_0
35 | - cached-property=1.5.2=hd8ed1ab_1
36 | - cached_property=1.5.2=pyha770c72_1
37 | - cachetools=5.0.0=pyhd8ed1ab_0
38 | - cairo=1.16.0=ha12eb4b_1010
39 | - cartopy=0.20.2=py38h51d8e34_4
40 | - cdo=2.0.5=h2e6804c_0
41 | - certifi=2022.5.18.1=py38h06a4308_0
42 | - cffi=1.15.0=py38h3931269_0
43 | - cftime=1.6.0=py38h71d37f0_1
44 | - charset-normalizer=2.0.12=pyhd8ed1ab_0
45 | - click=8.1.3=py38h578d9bd_0
46 | - colorama=0.4.4=pyh9f0ad1d_0
47 | - cryptography=36.0.2=py38h2b5fc30_1
48 | - cudatoolkit=11.5.1=hcf5317a_10
49 | - cudnn=8.2.1.32=h86fa8c9_0
50 | - curl=7.83.1=h7bff187_0
51 | - cycler=0.11.0=pyhd8ed1ab_0
52 | - dbus=1.13.6=h5008d03_3
53 | - debugpy=1.6.0=py38hfa26641_0
54 | - decorator=5.1.1=pyhd8ed1ab_0
55 | - defusedxml=0.7.1=pyhd8ed1ab_0
56 | - eccodes=2.25.0=hc08acdf_0
57 | - entrypoints=0.4=pyhd8ed1ab_0
58 | - executing=0.8.3=pyhd8ed1ab_0
59 | - expat=2.4.8=h27087fc_0
60 | - ffmpeg=4.4.1=hd7ab26d_2
61 | - fftw=3.3.10=nompi_h77c792f_102
62 | - flit-core=3.7.1=pyhd8ed1ab_0
63 | - font-ttf-dejavu-sans-mono=2.37=hab24e00_0
64 | - font-ttf-inconsolata=3.000=h77eed37_0
65 | - font-ttf-source-code-pro=2.038=h77eed37_0
66 | - font-ttf-ubuntu=0.83=hab24e00_0
67 | - fontconfig=2.14.0=h8e229c2_0
68 | - fonts-conda-ecosystem=1=0
69 | - fonts-conda-forge=1=0
70 | - fonttools=4.33.3=py38h0a891b7_0
71 | - freeglut=3.2.2=h9c3ff4c_1
72 | - freetype=2.10.4=h0708190_1
73 | - fribidi=1.0.10=h36c2ea0_0
74 | - frozenlist=1.3.0=py38h0a891b7_1
75 | - fsspec=2022.3.0=pyhd8ed1ab_0
76 | - future=0.18.2=py38h578d9bd_5
77 | - geos=3.10.2=h9c3ff4c_0
78 | - gettext=0.19.8.1=h73d1719_1008
79 | - giflib=5.2.1=h36c2ea0_2
80 | - gmp=6.2.1=h58526e2_0
81 | - gnutls=3.6.13=h85f3911_1
82 | - google-auth=2.6.6=pyh6c4a22f_0
83 | - google-auth-oauthlib=0.4.6=pyhd8ed1ab_0
84 | - graphite2=1.3.13=h58526e2_1001
85 | - grpcio=1.46.1=py38ha0cdfde_0
86 | - gst-plugins-base=1.20.2=hcf0ee16_0
87 | - gstreamer=1.20.2=hd4edc92_0
88 | - h5py=3.6.0=nompi_py38hfbb2109_100
89 | - harfbuzz=4.2.0=h40b6f09_0
90 | - hdf4=4.2.15=h10796ff_3
91 | - hdf5=1.12.1=nompi_h2386368_104
92 | - icu=69.1=h9c3ff4c_0
93 | - idna=3.3=pyhd8ed1ab_0
94 | - importlib-metadata=4.11.3=py38h578d9bd_1
95 | - importlib_metadata=4.11.3=hd8ed1ab_1
96 | - importlib_resources=5.7.1=pyhd8ed1ab_0
97 | - ipykernel=6.13.0=py38h7f3c49e_0
98 | - ipython=8.3.0=py38h578d9bd_0
99 | - ipython_genutils=0.2.0=py_1
100 | - ipywidgets=7.7.0=pyhd8ed1ab_0
101 | - jack=1.9.18=hfd4fe87_1001
102 | - jasper=2.0.33=ha77e612_0
103 | - jbig=2.1=h7f98852_2003
104 | - jedi=0.18.1=py38h578d9bd_1
105 | - jinja2=3.1.2=pyhd8ed1ab_0
106 | - joblib=1.1.0=pyhd3eb1b0_0
107 | - jpeg=9e=h166bdaf_1
108 | - json-c=0.15=h98cffda_0
109 | - json5=0.9.5=pyh9f0ad1d_0
110 | - jsonschema=4.5.1=pyhd8ed1ab_0
111 | - jupyter=1.0.0=py38h578d9bd_7
112 | - jupyter_client=7.3.1=pyhd8ed1ab_0
113 | - jupyter_console=6.4.3=pyhd8ed1ab_0
114 | - jupyter_core=4.10.0=py38h578d9bd_0
115 | - jupyter_server=1.17.0=pyhd8ed1ab_0
116 | - jupyterlab=3.4.1=pyhd8ed1ab_0
117 | - jupyterlab_pygments=0.2.2=pyhd8ed1ab_0
118 | - jupyterlab_server=2.13.0=pyhd8ed1ab_1
119 | - jupyterlab_widgets=1.1.0=pyhd8ed1ab_0
120 | - keyutils=1.6.1=h166bdaf_0
121 | - kiwisolver=1.4.2=py38h43d8883_1
122 | - krb5=1.19.3=h3790be6_0
123 | - lame=3.100=h7f98852_1001
124 | - lcms2=2.12=hddcbb42_0
125 | - ld_impl_linux-64=2.36.1=hea4e1c9_2
126 | - lerc=3.0=h9c3ff4c_0
127 | - libaec=1.0.6=h9c3ff4c_0
128 | - libblas=3.9.0=14_linux64_mkl
129 | - libbrotlicommon=1.0.9=h166bdaf_7
130 | - libbrotlidec=1.0.9=h166bdaf_7
131 | - libbrotlienc=1.0.9=h166bdaf_7
132 | - libcap=2.51=h166bdaf_1
133 | - libcblas=3.9.0=14_linux64_mkl
134 | - libclang=13.0.1=default_hc23dcda_0
135 | - libclang13=14.0.3=default_h3a83d3e_0
136 | - libcups=2.3.3=hf5a7f15_1
137 | - libcurl=7.83.1=h7bff187_0
138 | - libdb=6.2.32=h9c3ff4c_0
139 | - libdeflate=1.10=h7f98852_0
140 | - libdrm=2.4.109=h7f98852_0
141 | - libedit=3.1.20191231=he28a2e2_2
142 | - libev=4.33=h516909a_1
143 | - libevent=2.1.10=h9b69904_4
144 | - libffi=3.4.2=h7f98852_5
145 | - libflac=1.3.4=h27087fc_0
146 | - libgcc-ng=12.1.0=h8d9b700_16
147 | - libgfortran-ng=12.1.0=h69a702a_16
148 | - libgfortran5=12.1.0=hdcd56e2_16
149 | - libglib=2.70.2=h174f98d_4
150 | - libglu=9.0.0=he1b5a44_1001
151 | - libiconv=1.16=h516909a_0
152 | - liblapack=3.9.0=14_linux64_mkl
153 | - liblapacke=3.9.0=14_linux64_mkl
154 | - libllvm13=13.0.1=hf817b99_2
155 | - libllvm14=14.0.3=he0ac6c6_0
156 | - libnetcdf=4.8.1=nompi_h329d8a1_102
157 | - libnghttp2=1.47.0=h727a467_0
158 | - libnsl=2.0.0=h7f98852_0
159 | - libogg=1.3.4=h7f98852_1
160 | - libopencv=4.5.5=py38h2380011_9
161 | - libopus=1.3.1=h7f98852_1
162 | - libpciaccess=0.16=h516909a_0
163 | - libpng=1.6.37=h21135ba_2
164 | - libpq=14.2=hd57d9b9_0
165 | - libprotobuf=3.20.1=h6239696_0
166 | - libsndfile=1.0.31=h9c3ff4c_1
167 | - libsodium=1.0.18=h36c2ea0_1
168 | - libssh2=1.10.0=ha56f1ee_2
169 | - libstdcxx-ng=12.1.0=ha89aaad_16
170 | - libtiff=4.3.0=h542a066_3
171 | - libtool=2.4.6=h9c3ff4c_1008
172 | - libuuid=2.32.1=h7f98852_1000
173 | - libva=2.14.0=h7f98852_0
174 | - libvorbis=1.3.7=h9c3ff4c_0
175 | - libvpx=1.11.0=h9c3ff4c_3
176 | - libwebp=1.2.2=h3452ae3_0
177 | - libwebp-base=1.2.2=h7f98852_1
178 | - libxcb=1.13=h7f98852_1004
179 | - libxkbcommon=1.0.3=he3ba5ed_0
180 | - libxml2=2.9.12=h885dcf4_1
181 | - libzip=1.8.0=h4de3113_1
182 | - libzlib=1.2.11=h166bdaf_1014
183 | - llvm-openmp=14.0.3=he0ac6c6_0
184 | - lz4-c=1.9.3=h9c3ff4c_1
185 | - magics=4.11.0=he381006_1
186 | - magics-python=1.5.6=pyhd8ed1ab_0
187 | - magma=2.5.4=h6103c52_2
188 | - markdown=3.3.7=pyhd8ed1ab_0
189 | - markupsafe=2.1.1=py38h0a891b7_1
190 | - matplotlib=3.5.2=py38h578d9bd_0
191 | - matplotlib-base=3.5.2=py38h826bfd8_0
192 | - matplotlib-inline=0.1.3=pyhd8ed1ab_0
193 | - mistune=0.8.4=py38h497a2fe_1005
194 | - mkl=2022.0.1=h8d4b97c_803
195 | - multidict=6.0.2=py38h0a891b7_1
196 | - munkres=1.1.4=pyh9f0ad1d_0
197 | - mysql-common=8.0.29=haf5c9bc_0
198 | - mysql-libs=8.0.29=h28c427c_0
199 | - nbclassic=0.3.7=pyhd8ed1ab_0
200 | - nbclient=0.5.13=pyhd8ed1ab_0
201 | - nbconvert=6.5.0=pyhd8ed1ab_0
202 | - nbconvert-core=6.5.0=pyhd8ed1ab_0
203 | - nbconvert-pandoc=6.5.0=pyhd8ed1ab_0
204 | - nbformat=5.4.0=pyhd8ed1ab_0
205 | - nccl=2.12.12.1=h0800d71_0
206 | - ncurses=6.3=h27087fc_1
207 | - nest-asyncio=1.5.5=pyhd8ed1ab_0
208 | - netcdf4=1.5.8=nompi_py38h2823cc8_101
209 | - nettle=3.6=he412f7d_0
210 | - ninja=1.10.2=h4bd325d_1
211 | - notebook=6.4.11=pyha770c72_0
212 | - notebook-shim=0.1.0=pyhd8ed1ab_0
213 | - nspr=4.32=h9c3ff4c_1
214 | - nss=3.77=h2350873_0
215 | - numpy=1.22.3=py38h99721a1_2
216 | - oauthlib=3.2.0=pyhd8ed1ab_0
217 | - opencv=4.5.5=py38h578d9bd_9
218 | - openh264=2.1.1=h780b84a_0
219 | - openjpeg=2.4.0=hb52868f_1
220 | - openssl=1.1.1o=h7f8727e_0
221 | - ossuuid=1.6.2=hf484d3e_1000
222 | - packaging=21.3=pyhd8ed1ab_0
223 | - pandas=1.2.5=py38h295c915_0
224 | - pandoc=2.18=ha770c72_0
225 | - pandocfilters=1.5.0=pyhd8ed1ab_0
226 | - pango=1.50.7=hbd2fdc8_0
227 | - parso=0.8.3=pyhd8ed1ab_0
228 | - pcre=8.45=h9c3ff4c_0
229 | - pexpect=4.8.0=pyh9f0ad1d_2
230 | - pickleshare=0.7.5=py_1003
231 | - pillow=9.1.0=py38h0ee0e06_2
232 | - pip
233 | - pixman=0.40.0=h36c2ea0_0
234 | - proj=9.0.0=h93bde94_1
235 | - prometheus_client=0.14.1=pyhd8ed1ab_0
236 | - prompt-toolkit=3.0.29=pyha770c72_0
237 | - prompt_toolkit=3.0.29=hd8ed1ab_0
238 | - protobuf=3.20.1=py38hfa26641_0
239 | - psutil=5.9.0=py38h0a891b7_1
240 | - pthread-stubs=0.4=h36c2ea0_1001
241 | - ptyprocess=0.7.0=pyhd3deb0d_0
242 | - pulseaudio=14.0=hb166930_4
243 | - pure_eval=0.2.2=pyhd8ed1ab_0
244 | - py-opencv=4.5.5=py38h7f3c49e_9
245 | - pyasn1=0.4.8=py_0
246 | - pyasn1-modules=0.2.7=py_0
247 | - pycparser=2.21=pyhd8ed1ab_0
248 | - pydeprecate=0.3.2=pyhd8ed1ab_0
249 | - pygments=2.12.0=pyhd8ed1ab_0
250 | - pyjwt=2.3.0=pyhd8ed1ab_1
251 | - pyopenssl=22.0.0=pyhd8ed1ab_0
252 | - pyparsing=3.0.9=pyhd8ed1ab_0
253 | - pyproj=3.3.1=py38h5b5ac8f_0
254 | - pyqt=5.12.3=py38ha8c2ead_4
255 | - pyrsistent=0.18.1=py38h0a891b7_1
256 | - pyshp=2.3.0=pyhd8ed1ab_0
257 | - pysocks=1.7.1=py38h578d9bd_5
258 | - python=3.8.13=h582c2e5_0_cpython
259 | - python-dateutil=2.8.2=pyhd8ed1ab_0
260 | - python-fastjsonschema=2.15.3=pyhd8ed1ab_0
261 | - python_abi=3.8=2_cp38
262 | - pytorch=1.11.0=cuda112py38habe9d5a_1
263 | - pytorch-gpu=1.11.0=cuda112py38h68407e5_1
264 | - pytorch-lightning=1.6.3=pyhd8ed1ab_0
265 | - pytorch-mutex=1.0=cuda
266 | - pytz=2022.1=pyhd8ed1ab_0
267 | - pyu2f=0.1.5=pyhd8ed1ab_0
268 | - pyyaml=6.0=py38h0a891b7_4
269 | - pyzmq=22.3.0=py38hfc09fa9_2
270 | - qt=5.12.9=h1304e3e_6
271 | - qtconsole=5.3.0=pyhd8ed1ab_0
272 | - qtconsole-base=5.3.0=pyhd8ed1ab_0
273 | - qtpy=2.1.0=pyhd8ed1ab_0
274 | - readline=8.1=h46c0cb4_0
275 | - requests=2.27.1=pyhd8ed1ab_0
276 | - requests-oauthlib=1.3.1=pyhd8ed1ab_0
277 | - rsa=4.8=pyhd8ed1ab_0
278 | - scikit-learn=1.0.2=py38h51133e4_1
279 | - scipy=1.8.0=py38h56a6a73_1
280 | - seaborn=0.11.2=pyhd3eb1b0_0
281 | - send2trash=1.8.0=pyhd8ed1ab_0
282 | - setuptools=59.5.0=py38h578d9bd_0
283 | - shapely=1.8.2=py38h97f7145_1
284 | - simplejson=3.17.6=py38h0a891b7_1
285 | - sip=6.5.1=py38h709712a_2
286 | - six=1.16.0=pyh6c4a22f_0
287 | - sleef=3.5.1=h9b69904_2
288 | - sniffio=1.2.0=py38h578d9bd_3
289 | - soupsieve=2.3.1=pyhd8ed1ab_0
290 | - sqlite=3.38.5=h4ff8645_0
291 | - stack_data=0.2.0=pyhd8ed1ab_0
292 | - svt-av1=0.9.1=h27087fc_0
293 | - tbb=2021.5.0=h924138e_1
294 | - tensorboard=2.9.0=pyhd8ed1ab_0
295 | - tensorboard-data-server=0.6.0=py38h2b5fc30_2
296 | - tensorboard-plugin-wit=1.8.1=pyhd8ed1ab_0
297 | - terminado=0.13.3=py38h578d9bd_1
298 | - threadpoolctl=2.2.0=pyh0d69192_0
299 | - tinycss2=1.1.1=pyhd8ed1ab_0
300 | - tk=8.6.12=h27826a3_0
301 | - toml=0.10.2=pyhd8ed1ab_0
302 | - torchaudio=0.11.0=py38_cu115
303 | - torchmetrics=0.8.2=pyhd8ed1ab_0
304 | - torchvision=0.12.0=cuda112py38h46b2766_1
305 | - tornado=6.1=py38h0a891b7_3
306 | - tqdm=4.64.0=pyhd8ed1ab_0
307 | - traitlets=5.2.0=pyhd8ed1ab_0
308 | - typing-extensions=4.2.0=hd8ed1ab_1
309 | - typing_extensions=4.2.0=pyha770c72_1
310 | - udunits2=2.2.28=hc3e0081_0
311 | - unicodedata2=14.0.0=py38h0a891b7_1
312 | - urllib3=1.26.9=pyhd8ed1ab_0
313 | - voila=0.3.5=pyhd8ed1ab_0
314 | - wcwidth=0.2.5=pyh9f0ad1d_2
315 | - webencodings=0.5.1=py_1
316 | - websocket-client=1.3.2=pyhd8ed1ab_0
317 | - websockets=10.3=py38h0a891b7_0
318 | - werkzeug=2.1.2=pyhd8ed1ab_1
319 | - wheel=0.37.1=pyhd8ed1ab_0
320 | - widgetsnbextension=3.6.0=py38h578d9bd_0
321 | - x264=1!161.3030=h7f98852_1
322 | - x265=3.5=h924138e_3
323 | - xorg-fixesproto=5.0=h7f98852_1002
324 | - xorg-inputproto=2.3.2=h7f98852_1002
325 | - xorg-kbproto=1.0.7=h7f98852_1002
326 | - xorg-libice=1.0.10=h7f98852_0
327 | - xorg-libsm=1.2.3=hd9c2040_1000
328 | - xorg-libx11=1.7.2=h7f98852_0
329 | - xorg-libxau=1.0.9=h7f98852_0
330 | - xorg-libxdmcp=1.1.3=h7f98852_0
331 | - xorg-libxext=1.3.4=h7f98852_1
332 | - xorg-libxfixes=5.0.3=h7f98852_1004
333 | - xorg-libxi=1.7.10=h7f98852_0
334 | - xorg-libxrender=0.9.10=h7f98852_1003
335 | - xorg-renderproto=0.11.1=h7f98852_1002
336 | - xorg-xextproto=7.3.0=h7f98852_1002
337 | - xorg-xproto=7.0.31=h7f98852_1007
338 | - xz=5.2.5=h516909a_1
339 | - yaml=0.2.5=h7f98852_2
340 | - yarl=1.7.2=py38h0a891b7_2
341 | - zeromq=4.3.4=h9c3ff4c_1
342 | - zipp=3.8.0=pyhd8ed1ab_0
343 | - zlib=1.2.11=h166bdaf_1014
344 | - zstd=1.5.2=ha95c52a_0
345 | - pip:
346 | - pyqt5-sip==4.19.18
347 | - pyqtchart==5.12
348 | - pyqtwebengine==5.12.1
349 | - xarray==2022.3.0
350 | prefix: ./
351 |
--------------------------------------------------------------------------------
/models/unet_lightning.py:
--------------------------------------------------------------------------------
1 | # Weather4cast 2022 Starter Kit
2 | #
3 | # Copyright (C) 2022
4 | # Institute of Advanced Research in Artificial Intelligence (IARAI)
5 |
6 | # This file is part of the Weather4cast 2022 Starter Kit.
7 | #
8 | # The Weather4cast 2022 Starter Kit is free software: you can redistribute it
9 | # and/or modify it under the terms of the GNU General Public License as
10 | # published by the Free Software Foundation, either version 3 of the License,
11 | # or (at your option) any later version.
12 | #
13 | # The Weather4cast 2022 Starter Kit is distributed in the hope that it will be
14 | # useful, but WITHOUT ANY WARRANTY; without even the implied warranty of
15 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 | # GNU General Public License for more details.
17 | #
18 | # You should have received a copy of the GNU General Public License
19 | # along with this program. If not, see .
20 |
21 | # Contributors: Aleksandra Gruca, Pedro Herruzo, David Kreil, Stephen Moran
22 |
23 |
24 | import pytorch_lightning as pl
25 | from sklearn.metrics import confusion_matrix
26 | import torch
27 | import torch.nn as nn
28 | import torch.nn.functional as F
29 | from utils.evaluate import *
30 | import numpy as np;
31 |
32 | #models
33 | from models.baseline_UNET3D import UNet as Base_UNET3D # 3_3_2 model selection
34 |
35 | VERBOSE = False
36 | # VERBOSE = True
37 |
38 | class UNet_Lightning(pl.LightningModule):
39 | def __init__(self, UNet_params: dict, params: dict,
40 | **kwargs):
41 | super(UNet_Lightning, self).__init__()
42 |
43 | self.in_channels = params['in_channels']
44 | self.start_filts = params['init_filter_size']
45 | self.dropout_rate = params['dropout_rate']
46 | self.model = Base_UNET3D(in_channels=self.in_channels, start_filts = self.start_filts, dropout_rate = self.dropout_rate)
47 |
48 | self.save_hyperparameters()
49 | self.params = params
50 | #self.example_input_array = np.zeros((44,252,252))
51 |
52 | self.val_batch = 0
53 |
54 | self.prec = 7
55 |
56 | pos_weight = torch.tensor(params['pos_weight']);
57 | if VERBOSE: print("Positive weight:",pos_weight);
58 |
59 | self.loss = params['loss']
60 | self.bs = params['batch_size']
61 | self.loss_fn = {
62 | 'smoothL1': nn.SmoothL1Loss(), 'L1': nn.L1Loss(), 'mse': F.mse_loss,
63 | 'BCELoss': nn.BCELoss(),
64 | 'BCEWithLogitsLoss': nn.BCEWithLogitsLoss(pos_weight=pos_weight), 'CrossEntropy': nn.CrossEntropyLoss(), 'DiceBCE': DiceBCELoss(), 'DiceLoss': DiceLoss(),
65 | 'mIoULoss': mIoU()
66 | }[self.loss]
67 | self.main_metric = {
68 | 'smoothL1': 'Smooth L1',
69 | 'L1': 'L1',
70 | 'mse': 'MSE', # mse [log(y+1)-yhay]'
71 | 'BCELoss': 'BCE', # binary cross-entropy
72 | 'BCEWithLogitsLoss': 'BCE with logits',
73 | 'CrossEntropy': 'cross-entropy',
74 | 'DiceBCE': 'Dice BCE',
75 | 'DiceLoss': 'Dice loss',
76 | 'mIoULoss': 'Modified IoU'
77 | }[self.loss]
78 |
79 | self.relu = nn.ReLU() # None
80 | t = f"============== n_workers: {params['n_workers']} | batch_size: {params['batch_size']} \n"+\
81 | f"============== loss: {self.loss} | weight: {pos_weight} (if using BCEwLL)"
82 | print(t)
83 |
84 | def on_fit_start(self):
85 | """ create a placeholder to save the results of the metric per variable """
86 | metric_placeholder = {self.main_metric: -1}
87 | self.logger.log_hyperparams(self.hparams, metric_placeholder)
88 |
89 | def forward(self, x):
90 | x = self.model(x)
91 | #if self.loss =='BCELoss':
92 | #x = self.relu(x)
93 | return x
94 |
95 | def retrieve_only_valid_pixels(self, x, m):
96 | """ we asume 1s in mask are invalid pixels """
97 | ##print(f"x: {x.shape} | mask: {m.shape}")
98 | return x[~m]
99 |
100 | def get_target_mask(self, metadata):
101 | mask = metadata['target']['mask']
102 | #print("mask---->", mask.shape)
103 | return mask
104 |
105 | def _compute_loss(self, y_hat, y, agg=True, mask=None):
106 |
107 | if self.loss == "mIoULoss":
108 | y_hat = 0.5 * torch.tanh(y_hat/2) + 0.5
109 | if mask is not None:
110 | y_hat[mask] = 0
111 | y[mask] = 0
112 | # print("================================================================================")
113 | # print(y_hat.shape, y_hat.min(), y_hat.max())
114 | # print(y.shape, y.min(), y.max())
115 | if agg:
116 | loss = self.loss_fn(y_hat, y)
117 | else:
118 | loss = self.loss_fn(y_hat, y, reduction='none')
119 | return loss
120 |
121 | def training_step(self, batch, batch_idx, phase='train'):
122 | x, y, metadata = batch
123 | if VERBOSE:
124 | print('x', x.shape, 'y', y.shape, '----------------- batch')
125 | y_hat = self.forward(x)
126 | if VERBOSE:
127 | print('y_hat', y_hat.shape, 'y', y.shape, '----------------- model')
128 | mask = self.get_target_mask(metadata)
129 | loss = self._compute_loss(y_hat, y, mask=mask)
130 | #recall, precision, F1, acc = recall_precision_f1_acc(y, y_hat)
131 | #LOGGING
132 | #values = {'{phase}_loss': loss, '{phase}_acc': acc, '{phase}_recall': recall, '{phase}_precision': precision, 'F1': F1}
133 | self.log(f'{phase}_loss', loss,batch_size=self.bs, sync_dist=True)
134 | return loss
135 |
136 | def validation_step(self, batch, batch_idx, phase='val'):
137 | #data_start = timer()
138 | x, y, metadata = batch
139 | #data_end = timer()
140 | if VERBOSE:
141 | print('x', x.shape, 'y', y.shape, '----------------- batch')
142 | y_hat = self.forward(x)
143 | mask = self.get_target_mask(metadata)
144 | if VERBOSE:
145 | print('y_hat', y_hat.shape, 'y', y.shape, '----------------- model')
146 |
147 | loss = self._compute_loss(y_hat, y, mask=mask)
148 |
149 | # todo: add the same plot as in `test_step`
150 |
151 | if self.loss == "BCEWithLogitsLoss" or self.loss == "mIoULoss":
152 | print("applying thresholds to y_hat logits")
153 | # set the logits threshold equivalent to sigmoid(x)>=0.5
154 | idx_gt0 = y_hat>=0
155 | y_hat[idx_gt0] = 1
156 | y_hat[~idx_gt0] = 0
157 |
158 | if mask is not None:
159 | y_hat[mask]=0
160 | y[mask]=0
161 |
162 | recall, precision, F1, acc, csi = recall_precision_f1_acc(y, y_hat)
163 | iou = iou_class(y_hat, y)
164 |
165 | #LOGGING
166 | self.log(f'{phase}_loss', loss, batch_size=self.bs, sync_dist=True)
167 | values = {'val_acc': acc, 'val_recall': recall,
168 | 'val_precision': precision, 'val_F1': F1, 'val_iou': iou,
169 | 'val_CSI': csi
170 | # , 'val_N': float(x.shape[0])
171 | }
172 | self.log_dict(values, batch_size=self.bs, sync_dist=True)
173 |
174 | return {'loss': loss.cpu(), 'N': x.shape[0],
175 | 'iou': iou}
176 |
177 | def validation_epoch_end(self, outputs, phase='val'):
178 | print("Validation epoch end average over batches: ",
179 | [batch['N'] for batch in outputs]);
180 | avg_loss = np.average([batch['loss'] for batch in outputs],
181 | weights=[batch['N'] for batch in outputs]);
182 | avg_iou = np.average([batch['iou'] for batch in outputs],
183 | weights=[batch['N'] for batch in outputs]);
184 | values={f"{phase}_loss_epoch": avg_loss,
185 | f"{phase}_iou_epoch": avg_iou}
186 | self.log_dict(values, batch_size=self.bs, sync_dist=True)
187 | self.log(self.main_metric, avg_loss, batch_size=self.bs, sync_dist=True)
188 |
189 |
190 | def test_step(self, batch, batch_idx, phase='test'):
191 | x, y, metadata = batch
192 | if VERBOSE:
193 | print('x', x.shape, 'y', y.shape, '----------------- batch')
194 | y_hat = self.forward(x)
195 | mask = self.get_target_mask(metadata)
196 | if VERBOSE:
197 | print('y_hat', y_hat.shape, 'y', y.shape, '----------------- model')
198 | loss = self._compute_loss(y_hat, y, mask=mask)
199 | ## todo: add the same plot as in `test_step`
200 | if self.loss == "BCEWithLogitsLoss" or self.loss == "mIoULoss":
201 | print("applying thresholds to y_hat logits")
202 | # set the logits threshold equivalent to sigmoid(x)>=0.5
203 | idx_gt0 = y_hat>=0
204 | y_hat[idx_gt0] = 1
205 | y_hat[~idx_gt0] = 0
206 |
207 | if mask is not None:
208 | y_hat[mask]=0
209 | y[mask]=0
210 |
211 | recall, precision, F1, acc, csi = recall_precision_f1_acc(y, y_hat)
212 | iou = iou_class(y_hat, y)
213 |
214 | #LOGGING
215 | self.log(f'{phase}_loss', loss, batch_size=self.bs, sync_dist=True)
216 | values = {'test_acc': acc, 'test_recall': recall, 'test_precision': precision, 'test_F1': F1, 'test_iou': iou, 'test_CSI': csi}
217 | self.log_dict(values, batch_size=self.bs, sync_dist=True)
218 |
219 | return 0, y_hat
220 |
221 | def predict_step(self, batch, batch_idx, phase='predict'):
222 | x, y, metadata = batch
223 | y_hat = self.model(x)
224 | mask = self.get_target_mask(metadata)
225 | if VERBOSE:
226 | print('y_hat', y_hat.shape, 'y', y.shape, '----------------- model')
227 | if self.loss == "BCEWithLogitsLoss" or self.loss == "mIoULoss":
228 | print("applying thresholds to y_hat logits")
229 | # set the logits threshold equivalent to sigmoid(x)>=0.5
230 | idx_gt0 = y_hat>=0
231 | y_hat[idx_gt0] = 1
232 | y_hat[~idx_gt0] = 0
233 | return y_hat
234 |
235 | def configure_optimizers(self):
236 | if VERBOSE: print("Learning rate:",self.params["lr"], "| Weight decay:",self.params["weight_decay"])
237 | optimizer = torch.optim.AdamW(self.parameters(),
238 | lr=float(self.params["lr"]),weight_decay=float(self.params["weight_decay"]))
239 | return optimizer
240 |
241 | def seq_metrics(self, y_true, y_pred):
242 | text = ''
243 | cm = confusion_matrix(y_true, y_pred).ravel()
244 | if len(cm)==4:
245 | tn, fp, fn, tp = cm
246 | recall, precision, F1 = 0, 0, 0
247 |
248 | if (tp + fn) > 0:
249 | recall = tp / (tp + fn)
250 | r = f'r: {recall:.2f}'
251 |
252 | if (tp + fp) > 0:
253 | precision = tp / (tp + fp)
254 | p = f'p: {precision:.2f}'
255 |
256 | if (precision + recall) > 0:
257 | F1 = 2 * (precision * recall) / (precision + recall)
258 | f = f'F1: {F1:.2f}'
259 |
260 | acc = (tn + tp) / (tn+fp+fn+tp)
261 | text = f"{r} | {p} | {f} | acc: {acc} "
262 |
263 | return text
264 |
265 |
266 | def main():
267 | print("running")
268 | if __name__ == 'main':
269 | main()
270 |
271 | #PyTorch
272 | class DiceLoss(nn.Module):
273 | def __init__(self, weight=None, size_average=True):
274 | super(DiceLoss, self).__init__()
275 |
276 | def forward(self, targets, inputs, smooth=1):
277 |
278 | #comment out if your model contains a sigmoid or equivalent activation layer
279 | inputs = F.sigmoid(inputs)
280 |
281 | #flatten label and prediction tensors
282 | inputs = inputs.view(-1)
283 | targets = targets.view(-1)
284 |
285 | intersection = (inputs * targets).sum()
286 | print('intersection', intersection)
287 | print('inputs', inputs.sum())
288 | print('targets', targets.sum())
289 | dice = (2.*intersection + smooth)/(inputs.sum() + targets.sum() + smooth)
290 | print(1-dice)
291 | return 1 - dice
292 |
293 | class DiceBCELoss(nn.Module):
294 | def __init__(self, weight=None, size_average=True):
295 | super(DiceBCELoss, self).__init__()
296 |
297 | def forward(self, inputs, targets, smooth=1):
298 |
299 | #comment out if your model contains a sigmoid or equivalent activation layer
300 | inputs = F.sigmoid(inputs)
301 |
302 | #flatten label and prediction tensors
303 | inputs = inputs.view(-1)
304 | targets = targets.view(-1)
305 |
306 | intersection = (inputs * targets).sum()
307 | dice_loss = 1 - (2.*intersection + smooth)/(inputs.sum() + targets.sum() + smooth)
308 | BCE = F.binary_cross_entropy(inputs, targets, reduction='mean')
309 | Dice_BCE = BCE + dice_loss
310 |
311 | return Dice_BCE
312 |
313 |
314 | ## adapted from
315 | ## https://github.com/amirhosseinh77/UNet-AerialSegmentation/blob/main/losses.py
316 | ## by Amirhossein Heydarian (GPL 3)
317 | ##
318 | class mIoU(nn.Module):
319 | def __init__(self, n_classes=1):
320 | super(mIoU, self).__init__()
321 | self.classes = n_classes
322 |
323 | def forward(self, inputs, target):
324 | # inputs => N x Classes x H x W
325 | # target_oneHot => N x Classes x H x W
326 |
327 | N = inputs.size()[0]
328 | inter = inputs * target
329 | ## Sum over all pixels N x C x H x W => N x C
330 | inter = inter.view(N,self.classes,-1).sum(2)
331 | #Denominator
332 | union= inputs + target - (inputs*target)
333 | ## Sum over all pixels N x C x H x W => N x C
334 | union = union.view(N,self.classes,-1).sum(2)
335 | loss = torch.nan_to_num(inter/union)
336 |
337 | ## Return average loss over classes and batch
338 | return 1-loss.mean()
339 |
340 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | 
2 |
3 | # [Weather4cast](https://www.iarai.ac.at/weather4cast/) - Super-Resolution Rain Movie Prediction under Spatio-Temporal Shifts
4 | - Predict super-resolution rain movies in various regions of Europe
5 | - Transfer learning across space and time under strong shifts
6 | - Exploit data fusion to model ground-radar and multi-band satellite images
7 |
8 | ## Contents
9 | [Weather4cast: Super-Resolution Rain Movie Prediction under Spatio-Temporal Shifts](#weather4cast-multi-sensor-weather-forecasting-competition--benchmark-dataset)
10 | - [Contents](#contents)
11 | - [Introduction](#introduction)
12 | - [Get the data](#get-the-data)
13 | - [Submission guide](#submission-guide)
14 | - [Starter kit](#starter-kit)
15 | - [Environment](#environment)
16 | - [Training](#training)
17 | - [Validation](#validation)
18 | - [TensorBoard](#tensorboard)
19 | - [Generating a submission](#generating-a-submission)
20 | - [Automated generation of submissions](#automated-generation-of-submissions-helper-scripts)
21 | - [Code and Scientific Abstract](#code-and-scientific-abstract)
22 | - [Cite](#citation)
23 | - [Credits](#credits)
24 |
25 | ## Introduction
26 | The aim of the 2022 edition of the Weather4cast competition is to predict future high resolution rainfall events from lower resolution satellite radiances. Ground-radar reflectivity measurements are used to calculate pan-European composite rainfall rates by the [Operational Program for Exchange of Weather Radar Information (OPERA)](https://www.eumetnet.eu/activities/observations-programme/current-activities/opera/) radar network. While these are more precise, accurate, and of higher resolution than satellite data, they are expensive to obtain and not available in many parts of the world. We thus want to learn how to predict this high value rain rates from radiation measured by geostationary satellites operated by the [European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT)](https://www.eumetsat.int/).
27 |
28 | # Prediction task
29 | Competition participants should predict rainfall locations for the next 8 hours in 32 time slots from an input sequence of 4 time slots of the preceeding hour. The input sequence consists of four 11-band spectral satellite images. These 11 channels show slightly noisy satellite radiances covering so-called visible (VIS), water vapor (WV), and infrared (IR) bands. Each satellite image covers a 15 minute period and its pixels correspond to a spatial area of about 12km x 12km. The prediction output is a sequence of 32 images representing rain rates from ground-radar reflectivities. Output images also have a temporal resolution of 15 minutes but have higher spatial resolution, with each pixel corresponding to a spatial area of about 2km x 2km. So in addition to predicting the weather in the future, converting satellite inputs to ground-radar outputs, this adds a super-resolution task due to the coarser spatial resolution of the satellite data
30 |
31 | ### Weather4cast 2022 - Stage 1
32 | For Stage 1 of the competition we provide data from three Eureopean regions selected based on their preciptation characteristics. The task is to predict rain events 8 hours into the future from a 1 hour sequence of satellite images. The models should output binary pixels, with 1 and 0 indicating *rain* or *no rain* respectively. Rain rates computed from OPERA ground-radar reflectivities provide a ground truth. Although we provide the rain rates, at this stage, only rain/no rain needs to be predicted for each pixel.
33 |
34 | For Stage 1 we provide data from one year only, covering February to December 2019.
35 |
36 | ### Weather4cast 2022 - Stage 2
37 | In Stage 2 additional data will be provided for 2020 and 2021. Years 2019 and 2020 can then be used for training, while test sets from 2021 assess model robustness to temporal shifts. Additional regions with different climatological characteristics test model robustsness under spatial shifts. There are thus additional files for the new regions and years and thus the folder structure for stage 2 has been expanded accordingly to include additional sub-folders with the data for 2020 and 2021. In total there then are 7 regions with full training data in both 2019 and 2020. Three additional regions provide a spatial transfer learning challenge in years 2019 and 2020. For all ten regions, the year 2021 provides a temporal transfer learning challenge. For the seven regions with extensive training data in 2019 and 2020 this constitutes a pure temporal transfer learning challenge. The three additional regions 2021 provide a combined spatial and temporal transfer learning challenge.
38 |
39 |
40 | ## Get the data
41 | You need to register for the competition and accept its Terms and Conditions to access the data:
42 |
43 | - Competition Data: [Join and get the data](https://www.iarai.ac.at/weather4cast/get-data-2022/)
44 |
45 | Data are provided in [HDF-5](https://docs.h5py.org/en/stable/quick.html) files, separately for each year and data type. In our canonical folder structure `year/datatype/` the HRIT folder holds the satellite data and the OPERA folder provides the ground radar data. The file names reflect the different regions (`boxi_####`) and data splits (`train`, `validation`, and `test`). Ground truth for the test data split is of course withheld.
46 |
47 | After downloading the data, your data files should thus be arranged in folders of the following structure:
48 | ```
49 | 2019/
50 | +-- HRIT/ ... sub-folder for satellite radiance datasets
51 | +-- boxi_0015.test.reflbt0.ns.h5
52 | +-- boxi_0015.train.reflbt0.ns.h5
53 | +-- boxi_0015.val.reflbt0.ns.h5
54 | +-- boxi_00XX…….
55 | +-- OPERA/ ... sub-folder for OPERA ground-radar rain rates
56 | +-- boxi_0015.train.rates.crop.h5
57 | +-- boxi_0015.val.rates.crop.h5
58 | +-- boxi_00XX…….
59 | 2020/
60 | +-- HRIT/ ... sub-folder for satellite radiance datasets
61 | +-- boxi_0015.test.reflbt0.ns.h5
62 | +-- boxi_0015.train.reflbt0.ns.h5
63 | +-- boxi_0015.val.reflbt0.ns.h5
64 | +-- boxi_00XX…….
65 | +-- OPERA/ ... sub-folder for OPERA ground-radar rain rates
66 | +-- boxi_0015.train.rates.crop.h5
67 | +-- boxi_0015.val.rates.crop.h5
68 | +-- boxi_00XX…….
69 | ```
70 |
71 | Each HDF file provides a set of (multi-channel) images:
72 |
73 | - **boxi_00XX.train.reflbt0.ns.h5** provides *REFL-BT*, which is a tensor of shape `(20308, 11, 252, 252)` representing 20,308 images with 11 channels of satellite radiances for region XX. These are the input training data. The order of the channels in the H5 file corresonds to the following order of the satellite channels: `IR_016, IR_039, IR_087, IR_097, IR_108, IR_120,IR_134, VIS006, VIS008, WV_062, WV_073`.
74 |
75 | - **boxi_00XX.train.rates.crop.h5** provides *rates.crop*, which is a tensor of shape `(20308, 11, 252, 252)` representing OPERA ground-radar rain rates for the corresponding satellite radiances from the train dataset. Model output should be 1 or 0 for rain or no-rain predictions respectively.
76 |
77 | - **boxi_00XX.val.reflbt0.ns.h5** provides *REFL-BT*, which is a tensor of shape `(2160, 11, 252, 252)` representing additional measured satellite radiances. This data can be used as input for independent model validation. There are 60 validation sequences and each validation sequence consists of images for 4 input time slots; while in addition we also provide images for the 32 output time slots please note that this is just to aid model development and that model predictions cannot use these. The validation data set thus holds 4x60 + 32x60 = 2,160 images in total.
78 |
79 | - **boxi_00XX.val.rates.crop.h5** provides *rates.crop*, which is a tensor of shape `(2160, 1, 252, 252)` representing OPERA ground-radar rain rates for the corresponding satellite radiances from the validation dataset. Model output should be 1 or 0 for rain or no-rain predictions respectively.
80 |
81 | - **boxi_00XX.test.reflbt0.ns.h5** provides *REFL-BT*, which is a tensor of a shape `(240, 11, 252, 252)` representing additional satellite radiances. This dataset gives the input data for your model predictions for submission to the leaderboard. There are 60 input sequences in total, as each test sequence consists of images for 4 time slots (4x60 = 240). Note that no satellite radiances are provided for the future, so this is a true prediction task.
82 |
83 | Both input satellite radiances and output OPERA ground-radar rain rates are given for 252x252 pixel patches but please note that the spatial resolution of the satellite images is about six times lower than the resolution of the ground radar. This means that the 252x252 pixel ground radar patch corresponds to a 42x42 pixel center region in the coarser satellite resolution. The model target region thus is surrounded by a large area providing sufficient context as input for a prediction of future weather. In fact, fast storm clouds from one border of the input data would reach the center target region in about 7-8h.
84 |
85 | 
86 |
87 | - Stage-2 Competition: [Join and get the data](https://www.iarai.ac.at/weather4cast/get-data-2022/)
88 | ## Submission guide
89 | For submissions you need to upload a ZIP format archive of HDF-5 files that follows the folder structure below. Optionally, each HDF-5 file can be compressed by gzip, allowing for simple parallelization of the compression step. You need to include model predictions for all the regions. For each region, an HDF file should provide *submission*, a tensor of type `float32` and shape `(60, 1, 32, 252, 252)`, representing your predictions for the 60 test samples of a region. You need to follow the file naming convention shown in the example below to indicate the target region. Predictions for different years need to be placed in separate folders as shown below. The folder structure must be preserved in the submitted ZIP file. Please note that for Stage 1 we only ask for predictions for the year 2019, and predictions are simply 1 or 0 to indicate *rain* or *no rain* events respectively. For the Stage 2 Core Challenge, we ask for predictions for a total of 7 regions in both 2019 and 2020. For the Stage 2 Transfer Learning Challenge, predictions for 3 regions are required in years 2019 and 2020, and for all 10 regions in 2021. To simplify compilation of predictions, we now provide helper scripts in the Starter Kit.
90 |
91 | ```
92 | +-- 2019 –
93 | +-- boxi_0015.pred.h5.gz ...1 file per region for 60 test-sequence predictions of 32 time-slots each
94 | +-- boxi_00XX….
95 | +-- 2020 –
96 | +-- boxi_0015.pred.h5.gz
97 | +-- boxi_00XX….
98 | ```
99 |
100 | ## Starter kit
101 | This repository provides a starter kit accompanying the Weather4cast 2022 competition that includes example code to get you up to speed quickly. Please note that its use is entirely optional. The sample code includes a dataloader, some helper scripts, and a Unet-3D baseline model, some parameters of which can be set in a configuration file.
102 |
103 | To obtain the baseline model, you will need the `wget` command installed - then you can run
104 | ```
105 | ./mk_baseline.sh
106 | ```
107 | to fetch and patch a basic 3D U-Net baseline model.
108 |
109 | You will need to download the competition data separately. The sample code assumes that the downloaded data are organized in the following folder structure (shown here for Stage-1 data, conversely for Stage-2):
110 |
111 | ```
112 | +-- data
113 | +-- 2019 –
114 | +-- HRIT --
115 | +-- boxi_0015.test.reflbt0.ns.h5
116 | +-- boxi_0015.train.reflbt0.ns.h5
117 | +-- boxi_0015.val.reflbt0.ns.h5
118 | +-- boxi_0034.test.reflbt0.ns.h5
119 | +-- boxi_0034.train.reflbt0.ns.h5
120 | +-- boxi_0034.val.reflbt0.ns.h5
121 | +-- boxi_0076.test.reflbt0.ns.h5
122 | +-- boxi_0076.train.reflbt0.ns.h5
123 | +-- boxi_0076.val.reflbt0.ns.h5
124 | +-- OPERA --
125 | +-- boxi_0015.train.rates.crop.h5
126 | +-- boxi_0015.val.rates.crop.h5
127 | +-- boxi_0034.train.rates.crop.h5
128 | +-- boxi_0034.val.rates.crop.h5
129 | +-- boxi_0076.train.rates.crop.h5
130 | +-- boxi_0076.val.rates.crop.h5
131 | ```
132 |
133 | The path to the parent folder `data` needs to be provided as the `data_root` parameter in the `config_baseline.yaml` file.
134 |
135 | ### Environment
136 | We provide Conda environments for the sample code which can be recreated locally. An environment with libraries current at release can be recreated from the file `w4cNew.yml`using the following command:
137 | ```
138 | conda env create -f w4cNew.yml
139 | ```
140 | If you want to use older libraries for compatibility reasons, we also provide an earlier environment in `w4c.yml`. Finally, if you want to create an environment in the future, we also provide a script `mk_env.sh` to get you started.
141 | Note that all this can easily run for an hour or more, depending on your machine and setup.
142 |
143 | To activate the environment please run
144 | ```
145 | conda activate w4cNew
146 | ```
147 | or
148 | ```
149 | conda activate w4c
150 | ```
151 | respectively.
152 |
153 | ### Training
154 |
155 | We provide a script `train.py` with all the necessary code to train and explore a modified version of a [3D variant of the U-Net](https://github.com/ELEKTRONN/elektronn3). The script supports training from scratch or fine tuning from a provided checkpoint. The same script can also be used to evaluate model predictions on the validation data split using the flag `--mode val`, or to generate submissions from the test data using the flag `--mode predict`.
156 | In all cases please ensure you have set the correct data path in `config_baseline.yaml` and activated the `w4c` environment.
157 |
158 | *Example invocations:*
159 | - Training the model on a single GPU:
160 | ```
161 | python train.py --gpus 0 --config_path config_baseline.yaml --name name_of_your_model
162 | ```
163 | If you have more than one GPU you can select which GPU to use, with numbering starting from zero.
164 | - Fine tuning the model on 4 GPUs starting with a given checkpoint:
165 | ```
166 | python train.py --gpus 0 1 2 3 --mode train --config_path config_baseline.yaml --checkpoint "lightning_logs/PATH-TO-YOUR-MODEL-LOGS/checkpoints/YOUR-CHECKPOINT-FILENAME.ckpt" --name baseline_tune
167 | ```
168 |
169 | ### Validation
170 | Training will create logs and checkpoint files that are saved in the `lightning_logs` directory. To validate your model from a checkpoint you can for example run the following command (here for two CPUs):
171 | ```
172 | python train.py --gpus 0 1 --mode val --config_path config_baseline.yaml --checkpoint "lightning_logs/PATH-TO-YOUR-MODEL-LOGS/checkpoints/YOUR-CHECKPOINT-FILENAME.ckpt" --name baseline_validate
173 | ```
174 |
175 | ### TensorBoard
176 | You can of course also use [TensorBoard](https://www.tensorflow.org/tensorboard) to track and visualize model evaluation metrics during the training process.
177 | The standard TensorBoard command line is:
178 | ```
179 | tensorboard --logdir ./lightning_logs
180 | ```
181 | This should confirm that TensorBoard has started. For the default port, you point your browser to http://localhost:6006.
182 |
183 | ### Generating a submission
184 | Submission files can be generated from a trained model based on the model paramters saved in the checkpoint file. To generate predictions from your model checkpoint you can run the `train.py` script as below:
185 | ```
186 | train.py --gpus 0 --mode predict --config_path config_baseline.yaml --checkpoint "lightning_logs/PATH-TO-YOUR-MODEL-LOGS/checkpoints/YOUR-CHECKPOINT-FILENAME.ckpt"
187 | ```
188 | The code currently does not support generating a prediction for more than one region/year at a time.
189 |
190 | The results are saved in a single HDF-5 file named `boxi_00XX.pred.h5` in the `./submssion/YEAR/` folder, where *boxi_00XX* is the name of the region defined in the *predict* section your config file. A sample configuration is shown below:
191 | ```
192 | predict:
193 | region_to_predict: boxi_0015
194 | year_to_predict: 2019
195 | ```
196 | To generate predictions for multiple regions this needs to be run with a separate configuration file for each region.
197 |
198 | After generating prediction files for all the regions, please pack them into a single ZIP file (keeping the `year/` folder structure) and submit them to the [respective Weather4cast leaderboards](https://www.iarai.ac.at/weather4cast/challenge/).
199 |
200 | ### Automated generation of submissions (helper scripts)
201 | Considering the much increased number of individual predictions to collect for a leaderboard submission in Stage-2, we now provide helper scripts `mk_pred_core.sh` and `./mk_pred_transfer.sh` that can be used to generate and compile all individual predictions from a single model. The scripts display help text and diagnostics. Note that the use of these scripts is entirely optional because you may prefer to apply different models for different regions. You can provide both an output directory and a GPU ID to generate multiple predictions in parallel. The script will typically run for 20-40 minutes on a recent GPU system.
202 |
203 | Example invocation for interactive use:
204 | ```
205 | ./mk_pred_core.sh config_baseline_stage2-pred.yaml 'lightning_logs/yourModelName/checkpoints/yourCheckPointName.ckpt' yourSubmissionName 0 2>&1 | tee yourSubmission.core.log
206 | ```
207 |
208 | ## Code and Scientific Abstract
209 | At the end of the competition paricpants are required to provide:
210 | 1. A short scientific paper (Scientific Abstract) with a sufficiently detailed description of your approach (4-6 pages plus references)
211 | 2. The code and models with their learned weights that you used for your predictions, with explanations of how to reproduce you submissions.
212 |
213 | We will notify participants of how to provide their scientific abstract. For the code, you will need to submit it to a public repository like GitHub, providing a link to download the model's learned weights. Ideally, your repository should at least contain:
214 | - a) A list of **dependencies**. In the case of using Python, we suggest using conda/pip to generate them: `conda env export > environment.yml`. Make sure that your code can be executed from a fresh environment using the provided list of requirements: `conda env create -f environment.yml`.
215 | - b) **Code**, **models**, and a **folder with all model weights**.
216 | - c) An **out-of-the-box script** to use your best model **to generate predictions**. The script will read the inputs for the model from a given path and region, using its test folder (like the one used for the leaderboard), and save the outputs on a given path. The path to the folder containing the weights to be loaded by the models can also be an argument of the script.
217 |
218 |
219 | ## Citation
220 |
221 | When using or referencing the Weather4cast Competition in general or the competition data please cite:
222 | ```
223 | @INPROCEEDINGS{9672063,
224 | author={Herruzo, Pedro and Gruca, Aleksandra and Lliso, Llorenç and Calbet, Xavier and Rípodas, Pilar and Hochreiter, Sepp and Kopp, Michael and Kreil, David P.},
225 | booktitle={2021 IEEE International Conference on Big Data (Big Data)},
226 | title={High-resolution multi-channel weather forecasting – First insights on transfer learning from the Weather4cast Competitions 2021},
227 | year={2021},
228 | volume={},
229 | number={},
230 | pages={5750-5757},
231 | doi={10.1109/BigData52589.2021.9672063}
232 | }
233 |
234 | @inbook{10.1145/3459637.3482044,
235 | author = {Gruca, Aleksandra and Herruzo, Pedro and R\'{\i}podas, Pilar and Kucik, Andrzej and Briese, Christian and Kopp, Michael K. and Hochreiter, Sepp and Ghamisi, Pedram and Kreil, David P.},
236 | title = {CDCEO'21 - First Workshop on Complex Data Challenges in Earth Observation},
237 | year = {2021},
238 | isbn = {9781450384469},
239 | publisher = {Association for Computing Machinery},
240 | address = {New York, NY, USA},
241 | url = {https://doi.org/10.1145/3459637.3482044},
242 | booktitle = {Proceedings of the 30th ACM International Conference on Information & Knowledge Management},
243 | pages = {4878–4879},
244 | numpages = {2}
245 | }
246 | ```
247 |
248 | ## Credits
249 | The competition is organized / supported by:
250 | - [Institute of Advanced Research in Artificial Intelligence, Austria](https://iarai.ac.at)
251 | - [Silesian University of Technology, Poland](https://polsl.pl)
252 | - [European Space Agency Φ-lab, Italy](https://philab.phi.esa.int/)
253 | - [Spanish State Meteorological Agency, AEMET, Spain](http://aemet.es/)
254 |
--------------------------------------------------------------------------------
/utils/data_utils.py:
--------------------------------------------------------------------------------
1 | # Weather4cast 2022 Starter Kit
2 | #
3 | # Copyright (C) 2022
4 | # Institute of Advanced Research in Artificial Intelligence (IARAI)
5 |
6 | # This file is part of the Weather4cast 2022 Starter Kit.
7 | #
8 | # The Weather4cast 2022 Starter Kit is free software: you can redistribute it
9 | # and/or modify it under the terms of the GNU General Public License as
10 | # published by the Free Software Foundation, either version 3 of the License,
11 | # or (at your option) any later version.
12 | #
13 | # The Weather4cast 2022 Starter Kit is distributed in the hope that it will be
14 | # useful, but WITHOUT ANY WARRANTY; without even the implied warranty of
15 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 | # GNU General Public License for more details.
17 | #
18 | # You should have received a copy of the GNU General Public License
19 | # along with this program. If not, see .
20 |
21 | # Contributors: Aleksandra Gruca, Pedro Herruzo, David Kreil, Stephen Moran
22 |
23 |
24 | from re import VERBOSE
25 | import numpy as np
26 | import pandas as pd
27 | import yaml
28 | import h5py
29 | import pickle
30 | from torch.utils.data import DataLoader
31 | import time
32 | import torch
33 | import os
34 |
35 | from datetime import datetime, timedelta
36 |
37 | VERBOSE = False
38 | # VERBOSE = True
39 |
40 |
41 | # __________________________________________________CREATING/LOADING SAMPLE IDS____________________________________________________
42 |
43 |
44 | def load_sample_ids(
45 | data_split, splits_df, len_seq_in, len_seq_predict, regions, generate_pkl, years, path=""
46 | ):
47 | """For loading the sample idxs of the dataset. If a pkl file is found, it will be loaded. If not, it will be generated.
48 | If you want to save a .pkl file, set generate_pkl to True in the yaml file.
49 |
50 | Args:
51 | data_split (string): The data split to load the sample idxs for.
52 | splits_df (DataFrame): The dataframe specifying what split each timepoint belongs to. Used to generate samples
53 | len_seq_in (int): The length of the input sequence.
54 | len_seq_predict (int): The length of the prediction sequence.
55 | regions (list): the regions to create samples for
56 | generate_pkl (bool): If True, a pkl file will be generated. If False, a pkl file will be loaded.
57 | path (str, optional): The path to load sample idxs from. Defaults to ''.
58 |
59 | Returns:
60 | dict: A dictionary of the sample idxs.
61 | """
62 |
63 | print("|YEARS]", years)
64 |
65 | # if generate_pkl:
66 | # idxs_r = generate_and_cache_sequences(
67 | # data_split, splits_df, len_seq_in, len_seq_predict, regions, years, path
68 | # )
69 | # # elif ".pkl" in path: # read pre-computed (fastest)
70 | # # print("load_sample_ids using pkl cache:", path)
71 | # # idxs_r = read_samples_ids(path, data_split)
72 | # else:
73 | idxs_r = generate_sample_ids(
74 | data_split, splits_df, len_seq_in, len_seq_predict, regions, years
75 | )
76 | return idxs_r
77 |
78 |
79 | def generate_and_cache_sequences(
80 | split, splits_df, len_seq_in, len_seq_predict, regions, path_name, years
81 | ):
82 | """Generates and saves the the sample idxs for the dataset. This generates the sample idxs for all data splits.
83 |
84 | Args:
85 | split (string): The data split to return the sample idxs for.
86 | splits_df (DataFrame): The dataframe specifying what split each timepoint belongs to. Used to generate samples
87 | len_seq_in (_type_): The length of the input sequence.
88 | len_seq_predict (_type_): The length of the prediction sequence.
89 | regions (_type_): The regions to create samples for
90 | path_name (_type_): The path to save the sample idxs to.
91 |
92 | Returns:
93 | list: A list of the sample idxs.
94 | """
95 |
96 | samples = {"training": [], "validation": [], "test": []}
97 |
98 | for data_split in ["training", "validation", "test"]:
99 | idxs = generate_sample_ids(
100 | data_split, splits_df, len_seq_in, len_seq_predict, regions, years
101 | )
102 | samples[data_split] = idxs
103 | if VERBOSE:
104 | print(f"{len(idxs)} {data_split} samples")
105 |
106 | qq = samples.copy()
107 | with open(path_name, "wb") as f:
108 | pickle.dump(qq, f)
109 | return samples[split]
110 |
111 |
112 | def read_samples_ids(path, data_split):
113 | """read in sample idxs if they are already generated
114 |
115 | Args:
116 | path (String): path to pkl file with sample idxs
117 | data_split (String): data split to load sample idxs for
118 |
119 | Returns:
120 | list: list of sample idxs
121 | """
122 | with open(path, "rb") as f:
123 | loaded_dict = pickle.load(f)
124 | return loaded_dict[data_split]
125 |
126 |
127 | def generate_sample_ids(data_split, splits_df, len_seq_in, len_seq_predict, regions, years):
128 | """generates sample idxs for a given data split
129 |
130 | Args:
131 | data_split (String): data split to load sample idxs for
132 | splits_df (DataFrame): the dataframe specifying what split each timepoint belongs to. Used to generate samples
133 | len_seq_in (int): the length of the input sequence.
134 | len_seq_predict (int): the length of the prediction sequence.
135 | regions (list): the regions to create samples for
136 |
137 | Returns:
138 | list: list of sample idxs for a specific data split
139 | """
140 | if data_split == "training":
141 | idxs_r = get_training_idxs(
142 | splits_df, len_seq_in, len_seq_predict, data_split, regions[0], years
143 | )
144 | for r in range(1, len(regions)): # append rest of regions to df
145 | idxs = get_training_idxs(
146 | splits_df, len_seq_in, len_seq_predict, data_split, regions[r], years
147 | )
148 | idxs_r = idxs_r + idxs
149 |
150 | elif data_split == "validation": # Validation/Testing/Heldout Data
151 | idxs_r = get_validation_idxs(
152 | splits_df, len_seq_in, len_seq_predict, data_split, regions[0], years
153 | )
154 | for r in range(1, len(regions)): # append rest of regions to df
155 | idxs = get_validation_idxs(
156 | splits_df, len_seq_in, len_seq_predict, data_split, regions[r], years
157 | )
158 | idxs_r = idxs_r + idxs
159 | else: # testing
160 | idxs_r = get_test_heldout_idxs(splits_df, len_seq_in, data_split, regions[0], years)
161 | for r in range(1, len(regions)): # append rest of regions to df
162 | idxs = get_test_heldout_idxs(splits_df, len_seq_in, data_split, regions[r], years)
163 | idxs_r = idxs_r + idxs
164 | return idxs_r
165 |
166 |
167 | def get_training_idxs(df, len_seq_in, len_seq_predict, data_split, region, years):
168 | """get sample idxs for training data split
169 | Args:
170 | df (DataFrame): _description_
171 | len_seq_in (int): length of input sequence
172 | len_seq_predict (int): length of prediction sequence
173 | data_split (String): data split to load sample idxs for
174 | region (String): region to create samples for
175 | Returns:
176 | list: list of sample idxs for training data split
177 | """
178 | idxs = []
179 |
180 | df = df[df['split'] == data_split]
181 | df = df[df['all_vars'] == 1]
182 | dfs= pd.DataFrame(columns = df.columns);
183 |
184 | print("years //// ", years)
185 | for year in years:
186 | dfs=pd.concat([dfs,df[df['date'].dt.year==int(year)]]);
187 | if VERBOSE:
188 | print(f"Keeping {dfs.shape[0]} {data_split} rows of {df.shape[0]} for {region} in {years}");
189 | df=dfs;
190 |
191 | # non testing
192 | for i in range(df.shape[0] - len_seq_in - len_seq_predict + 1):
193 | # date check
194 |
195 | s_id = df.iloc[i]["date_time_str"]
196 | e_id = df.iloc[i + len_seq_in + len_seq_predict - 1]["date_time_str"]
197 |
198 | dd1, mm1, yy1 = get_day_month_year(s_id)
199 | h1, m1, s1 = get_hours_minutes_seconds(s_id)
200 | start_dt = datetime(yy1, mm1, dd1, h1, m1, s1)
201 |
202 | dd2, mm2, yy2 = get_day_month_year(e_id)
203 | h2, m2, s2 = get_hours_minutes_seconds(e_id)
204 | end_dt = datetime(yy2, mm2, dd2, h2, m2, s2)
205 |
206 | if start_dt + timedelta(hours=8, minutes=45, seconds=0) == end_dt:
207 | # print(get_future_time(df.iloc[i]['time'], (len_seq_predict * 15)))
208 | in_seq = [i + j for j in range(len_seq_in)]
209 | out_seq = [i + len_seq_in + j for j in range(len_seq_predict)]
210 | idxs.append([in_seq, out_seq, region])
211 |
212 | return idxs
213 |
214 |
215 | def get_test_heldout_idxs(df, len_seq_in, data_split, region, years):
216 | """get sample idxs for test data split
217 |
218 | Args:
219 | df (DataFrame): _description_
220 | len_seq_in (int): length of input sequence
221 | len_seq_predict (int): length of prediction sequence
222 | data_split (String): data split to load sample idxs for
223 | region (String): region to create samples for
224 |
225 | Returns:
226 | list: list of sample idxs for test data split
227 | """
228 | idxs = []
229 |
230 | split_type = f"{data_split}_in"
231 | df = df[df["split_type"] == split_type]
232 | df = df[df["all_vars"] == 1]
233 | dfs= pd.DataFrame(columns = df.columns);
234 | for year in years:
235 | dfs=pd.concat([dfs,df[df['date'].dt.year==int(year)]]);
236 | if VERBOSE:
237 | print(f"Keeping {dfs.shape[0]} {data_split} rows of {df.shape[0]} for {region} in {years}");
238 | df=dfs;
239 |
240 |
241 | for start_index in range(0, df.shape[0], len_seq_in):
242 | in_seq = [start_index + i for i in range(len_seq_in)]
243 | test_seq = [in_seq, [], region]
244 | idxs.append(test_seq)
245 | return idxs
246 |
247 |
248 | def get_validation_idxs(df, len_seq_in, len_seq_out, data_split, region, years):
249 | """get sample idxs for validation data split
250 |
251 | Args:
252 | df (DataFrame): timepoint dataframe
253 | len_seq_in (int): length of input sequence
254 | len_seq_predict (int): length of prediction sequence
255 | data_split (String): data split to load sample idxs for
256 | region (String): region to create samples for
257 |
258 | Returns:
259 | list: list of sample idxs for validation data split
260 | """
261 | idxs = []
262 | split_type = f"{data_split}_in"
263 | df = df[df["split"] == data_split]
264 | df = df[df["all_vars"] == 1]
265 | dfs= pd.DataFrame(columns = df.columns);
266 | for year in years:
267 | dfs=pd.concat([dfs,df[df['date'].dt.year==int(year)]]);
268 | if VERBOSE:
269 | print(f"Keeping {dfs.shape[0]} {data_split} rows of {df.shape[0]} for {region} in {years}");
270 | df=dfs;
271 |
272 | for i in range(df.shape[0]):
273 | if df.iloc[i]["split_type"] == split_type:
274 | input_output_seq = get_io_validation_times(
275 | df, i, len_seq_in, len_seq_out, region
276 | )
277 | if input_output_seq:
278 | idxs.append(input_output_seq)
279 | return idxs
280 |
281 |
282 | def get_io_validation_times(df, start_index, len_seq_in, len_seq_predict, region):
283 | """get input and output times for validation data split. Checks for valid sequences
284 |
285 | Args:
286 | df (int): timepoint dataframe
287 | start_index (int): startng index of sequence
288 | len_seq_in (int): length of input sequence
289 | len_seq_predict (int): length of prediction sequence
290 | region (String): region to create sequences for
291 |
292 | Returns:
293 | list: list of input, output times and region for validation data split
294 | """
295 | split_type_in = f"validation_in"
296 | end_index = start_index + len_seq_in + len_seq_predict - 1
297 | if end_index < len(df):
298 | seq = df[start_index:end_index]
299 | in_seq = seq[:len_seq_in]
300 | out_seq = seq[len_seq_in:]
301 | # check all input is of the same type
302 | if len(in_seq[in_seq["split_type"] == split_type_in]) != len(in_seq):
303 | return []
304 | # convert input/output sequence to lists and return
305 | else:
306 | in_seq = [start_index + i for i in range(len_seq_in)]
307 | out_seq = [start_index + len_seq_in + i for i in range(len_seq_predict)]
308 | return [in_seq, out_seq, region]
309 | else:
310 | return []
311 |
312 |
313 | # ______________________________________LOADING DATA_______________________________________
314 |
315 |
316 | def load_dataset(root, data_split, regions, years, product):
317 | """load dataset from root folder and return as list
318 | Args:
319 | root (String): root folder to load dataset from
320 | data_split (String): data split to load dataset for
321 | regions (list): list of regions to load dataset for
322 | product (String): product to load dataset for e.g. satellite or radar
323 | Returns:
324 | list : full dataset
325 | """
326 | dataset = {}
327 | if data_split == "validation":
328 | data_split = "val"
329 | elif data_split == "training":
330 | data_split = "train"
331 |
332 | if VERBOSE:
333 | print(f"Data split {data_split}")
334 | print(f"Regions {regions}")
335 | print(f"Years {years}")
336 | print(f"Product {product}")
337 |
338 | for region in regions:
339 | if VERBOSE: print(f"Region: {region}")
340 | yearly_records = {}
341 | for year in years:
342 | if VERBOSE: print(f"Year: {year}")
343 | if product == "RATE":
344 | file = f"{region}.{data_split}.rates.crop.h5"
345 | path = f"{root}/{year}/OPERA/{file}"
346 | f = h5py.File(path, "r")
347 | ds = f["rates.crop"]
348 | yearly_records[year] = ds
349 | # f.close()
350 | else:
351 | file = f"{region}.{data_split}.reflbt0.ns.h5"
352 | path = f"{root}/{year}/HRIT/{file}"
353 | f = h5py.File(path, "r")
354 | ds = f["REFL-BT"]
355 | yearly_records[year] = ds
356 | # f.close()
357 | # close file memory leak?
358 | # when done reading, close files?
359 |
360 | if VERBOSE: print(f"Done reading {region}... skipping concatenation")
361 | dataset[region] = yearly_records
362 |
363 | if VERBOSE: print(f"{data_split} Data read ... returning to dataset.)")
364 | if VERBOSE: print(f"Read {len(dataset)} regions, with each containing {len(dataset[list(dataset.keys())[0]])} years.")
365 | return dataset
366 |
367 |
368 | def get_sequence(
369 | seq,
370 | root,
371 | data_split,
372 | region,
373 | product,
374 | bands,
375 | preprocess=None,
376 | swap_time_ch=False,
377 | ds=None,
378 | ):
379 | """get data and metadata for a given sequence.
380 |
381 | Args:
382 | seq (list): list containing sequence of idxs to be retrieved
383 | root (String): root for data
384 | data_split (String): data split to load sequence for
385 | region (String): region to load sequence for
386 | product (String): product to load sequence for e.g. satellite or radar
387 | bands (list): list of bands to load sequence for
388 | preprocess (list, optional): preprocessing settings. Defaults to None.
389 | swap_time_ch (bool, optional): swap time and channel axis if set to True. Defaults to False.
390 | ds (numpy array, optional): full dataset to read from. Defaults to None.
391 | Returns:
392 | prod_seq: sequence of data and metadata
393 | mask_seq: corresponding sequence of masks
394 | """
395 |
396 | prod_seq = []
397 | mask_seq = []
398 | for s in seq:
399 | prods, masks = get_file(
400 | s, root, data_split, region, product, bands, preprocess, ds
401 | )
402 | prod_seq.append(prods)
403 | mask_seq.append(masks)
404 | # return format - time x channels x width x height
405 | mask_seq = np.asarray(mask_seq)
406 |
407 | # Swapping Axes to give shape channels x time x width x height
408 | if swap_time_ch:
409 | prod_seq = np.swapaxes(prod_seq, 0, 1)
410 | mask_seq = np.swapaxes(mask_seq, 0, 1)
411 | return np.array(prod_seq), mask_seq
412 |
413 |
414 | def get_file(
415 | sample_id, root, data_split, region, product, bands, preprocess=None, ds=None
416 | ):
417 | """Read/Preprocess data and metadata for single timepoint.
418 |
419 | Args:
420 | sample_id (int): idx to be retrieved
421 | root (String): root for data
422 | data_split (String): data split to load data point for
423 | region (String): region to load data point for
424 | product (String): product to load data point for e.g. satellite or radar
425 | bands (list): satellite bands to load data point for
426 | preprocess (list, optional): preprocessing settings. Defaults to None.
427 | ds (numpy array, optional): full dataset to read from. Defaults to None.
428 |
429 | Returns:
430 | x(numpy array): data for single timepoint
431 | masks(numpy array): masks for single timepoint
432 | """
433 |
434 | if VERBOSE:
435 | print("\n", "_____________READING FILE_________________________")
436 | file_t = time.time()
437 | # Get file containing all neede channels for a given time - (1xCxWxH)
438 | prods = []
439 | masks = []
440 |
441 | x = read_file(sample_id, ds, region)
442 | x = np.float32(x)
443 | if VERBOSE:
444 | print(time.time() - file_t, "reading file time")
445 | if product == "RATE":
446 | for b in ["rainfall_rate-500X500"]:
447 | if preprocess[product][b]["mask"]:
448 | mask = create_mask(x, preprocess[product][b]["mask"])
449 | masks.append(mask)
450 | masks = masks[0]
451 | if preprocess is not None:
452 | x = preprocess_OPERA(x, preprocess[product][b])
453 | if VERBOSE:
454 | print(time.time() - file_t, "OPERA preprocess time")
455 | else:
456 | for j, b in enumerate(bands):
457 | if preprocess is not None:
458 | x[j] = preprocess_HRIT(x[j], preprocess[b])
459 | if VERBOSE:
460 | print(time.time() - file_t, "HRIT preprocess time")
461 |
462 | # return format - channels x width x height
463 | if VERBOSE:
464 | print(time.time() - file_t, "Total File Read Time")
465 | return x, masks
466 |
467 |
468 | def read_file(sample_id, ds, region):
469 | previous_year_samples = 0
470 | for year in ds[region]:
471 | # print(sample_id, previous_year_samples)
472 | i = sample_id - previous_year_samples;
473 | if i < len(ds[region][year]):
474 | return ds[region][year][i]
475 | previous_year_samples += len(ds[region][year])
476 |
477 | raise Exception(f"Sample {sample_id} not found for {region}")
478 |
479 | # ___________________PREPROCESSING FUNCTIONS_______________________________________
480 |
481 |
482 | def crop_numpy(x, crop):
483 | """crop numpy array
484 |
485 | Args:
486 | x (numpy array): array to be cropped
487 | crop (int): crop size
488 |
489 | Returns:
490 | data(numpy array): cropped array
491 | """
492 | return x[:, :, crop:-crop, crop:-crop]
493 |
494 |
495 | def preprocess_fn(x, preprocess, verbose=False):
496 | """Funciton to precprocess data
497 |
498 | Assumption: the mask has been previously created (otherwise won't be recovered)
499 | # 1. map values
500 | # 2. clip values out of range
501 | Args:
502 | x (numpy array): data to be preprocessed
503 | preprocess (list): preprocessing settings
504 | verbose (bool, optional): verbose preprocessing. Defaults to False.
505 |
506 | Returns:
507 | data(numpy array): preprocessed data
508 | """
509 |
510 | if verbose:
511 | print(0, np.unique(x))
512 | # 1 Map values
513 | for q, v in preprocess["map"]:
514 |
515 | if isinstance(q, str):
516 | if q == "nan":
517 | x[np.isnan(x)] = v
518 | elif q == "inf":
519 | x[np.isinf(x)] = v
520 | elif "greaterthan" in q:
521 | greater_v = float(q.partition("greaterthan")[2])
522 | x[x > greater_v] = v
523 | elif "lessthan" in q:
524 | less_v = float(q.partition("lessthan")[2])
525 | x[x < less_v] = v
526 | else:
527 | x[x == q] = v
528 | if verbose:
529 | print(1, np.unique(x, return_counts=True))
530 |
531 | # 2 Clip values out of range
532 | m, M = preprocess["range"]
533 | x[x < m] = m
534 | x[x > M] = M
535 | if verbose:
536 | print(2, np.unique(x, return_counts=True))
537 |
538 | return x, M
539 |
540 |
541 | def preprocess_HRIT(x, preprocess, verbose=False):
542 | """Preprocess HRIT data - standardize data, map values, clip values out of range
543 |
544 | Assumption: the mask has been previously created (otherwise won't be recovered)
545 | Args:
546 | x (numpy array): data to be preprocessed
547 | preprocess (list): preprocessing settings
548 | verbose (bool, optional): verbose preprocessing. Defaults to False.
549 |
550 | Returns:
551 | x(list): preprocessed data
552 | """
553 | # 1, 2
554 | if verbose:
555 | print("HRIT file:")
556 | x, M = preprocess_fn(x, preprocess, verbose=verbose)
557 |
558 | # 3 - mean_std - standardisation
559 | if preprocess["standardise"]:
560 | # x = x/M
561 | mean, stdev = preprocess["mean_std"]
562 | x = x - mean
563 | x = x / stdev
564 | if verbose:
565 | print(3, np.uniquze(x, return_counts=True))
566 | return x
567 |
568 |
569 | def preprocess_OPERA(x, preprocess, verbose=False):
570 | """Precprocess OPERA data - standardize data, map values, clip values out of range
571 | Assumption: the mask has been previously created (otherwise won't be recovered)
572 | # 1. map values
573 | # 2. clip values out of range
574 |
575 | Args:
576 | x (numpy array): data to be preprocessed
577 | preprocess (list): preprocessing settings
578 | verbose (bool, optional): verbose preprocessing. Defaults to False.
579 |
580 | Returns:
581 | x(numpy array): preprocessed data
582 | """
583 | # 1, 2
584 | if verbose:
585 | print("OPERA file:")
586 | x, M = preprocess_fn(x, preprocess, verbose=verbose)
587 |
588 | # 3 - mean_std - standardisation
589 | if preprocess["standardise"]:
590 | mean, stdev = preprocess["mean_std"]
591 | x = x - mean
592 | x = x / stdev
593 | if preprocess["bin"]:
594 | bins = np.arange(0, 128, 0.2)
595 | x = np.digitize(x, bins)
596 | if verbose:
597 | print(3, np.unique(x, return_counts=True))
598 | return x
599 |
600 |
601 | # __________________________________GENERAL HELPER FUNCTIONS__________________________________
602 |
603 |
604 | def standardise_time_strings(time):
605 | if len(time) < 6:
606 | time = time.rjust(6, "0")
607 | else:
608 | return time
609 | return time
610 |
611 |
612 | def get_hours_minutes_seconds(date_time_str):
613 | h = date_time_str[9:11]
614 | m = date_time_str[11:13]
615 | s = date_time_str[13:15]
616 | return int(h), int(m), int(s)
617 |
618 |
619 | def get_day_month_year(date_time_str):
620 | yy = date_time_str[0:4]
621 | mm = date_time_str[4:6]
622 | dd = date_time_str[6:8]
623 | return int(dd), int(mm), int(yy)
624 |
625 |
626 | def time_2_channels(w, height, width):
627 | """collapse time dimension into channels - (B, C, T, H, W) -> (B, C*T, H, W)
628 |
629 | Args:
630 | w (numpy array): data
631 | height (int): image height in pixels
632 | width (int): image width in pixels
633 |
634 | Returns:
635 | w(numpy array): data with time dimension collapsed into channels
636 | """
637 | w = np.reshape(w, (-1, height, width))
638 | return w
639 |
640 |
641 | def channels_2_time(w, seq_time_bins, n_channels, height, width):
642 | """Recover time dimension from channels - (B, C*T, H, W) -> (B, T, C, H, W)
643 |
644 | Args:
645 | w (numpy array): data
646 | seq_time_bins (_type_): number of time bins
647 | n_channels (int): number of channels
648 | height (int): image height in pixels
649 | width (int): image width in pixels
650 |
651 | Returns:
652 | _type_: _description_
653 | """
654 | w = np.reshape(w, (seq_time_bins, n_channels, height, width))
655 | return w
656 |
657 |
658 | def load_timestamps(
659 | path, types={"time": "str", "date_str": "str", "split_type": "str"}
660 | ):
661 | """load timestamps from a from csv file
662 |
663 | Args:
664 | path (String): path to the csv file
665 | types (dict, optional): types to cast columns of dataframe to. Defaults to {'time': 'str', 'date_str': 'str', 'split_type': 'str'}.
666 |
667 | Returns:
668 | Dataframe: dataframe with timestamps
669 | """
670 |
671 | df = pd.read_csv(path, index_col=False, dtype=types)
672 |
673 | df.sort_values(by=["date_str", "time"], inplace=True)
674 |
675 | # to datetime type
676 | df["date"] = pd.to_datetime(df["date"])
677 |
678 | # convert times to strings
679 | df["time"] = df["time"].astype(str)
680 | df["time"] = df["time"].apply(standardise_time_strings)
681 |
682 | df["date_str"] = df["date_str"].astype(str)
683 | # create date_time string
684 | df["date_time_str"] = df["date_str"] + "T" + df["time"]
685 |
686 | return df
687 |
688 |
689 | def load_config(config_path):
690 | """Load confgiuration file
691 |
692 | Args:
693 | config_path (String): path to configuration file
694 |
695 | Returns:
696 | dict: configuration file
697 | """
698 | with open(config_path) as file:
699 | config = yaml.safe_load(file)
700 | return config
701 |
702 |
703 | def create_mask(data, mask_values, verbose=False):
704 | """Create a mask from a data array and a list of values to mask out
705 | - " 1's indicate (any kind of) missing value
706 | Args:
707 | data (numpy array): data to be masked
708 | mask_values (list): list of values to mask out
709 | verbose (bool, optional): verbose preprocessing. Defaults to False.
710 |
711 | Returns:
712 | mask(numpy array): array of mask corresponding to the data
713 | """
714 |
715 | mask = np.full(data.shape, False, dtype=bool)
716 | for m in mask_values:
717 | if isinstance(m, str):
718 | if m == "nan":
719 | filter = np.isnan(data)
720 | elif m == "inf":
721 | filter = np.isinf(data)
722 | elif "max" in m:
723 | max_v = int(m.partition("max")[2])
724 | filter = data > max_v
725 | elif "range" in m:
726 | range = m.partition("range")
727 | l = float(range[0])
728 | h = float(range[2])
729 | filter = (data >= l) & (data < h)
730 | else:
731 | filter = data == m
732 | mask = np.logical_or(mask, filter)
733 |
734 | if verbose:
735 | n_missing = np.unique(filter, return_counts=True)
736 | print("-----> total masked values:", n_missing, "\n")
737 | return mask
738 |
739 |
740 | def get_mean_std(dataset, batch_size=10):
741 | """Get the mean and stdev of the different image bands
742 | Args:
743 | times (list): timepoints to get mean from
744 | batch_size (int, optional): batch size for dataloader. Defaults to 1000.
745 | Returns:
746 | mean (list): list of band means
747 | stdev (list): list of band stdevs
748 | """
749 | loader = DataLoader(dataset, batch_size=batch_size)
750 | nimages = 0
751 | mean = 0
752 | std = 0
753 | count = 0
754 | for batch, _, _ in loader:
755 | print(count)
756 | count += 1
757 | # Rearrange batch to be the shape of [B, C, T * W * H]
758 | batch = batch.view(batch.size(0), batch.size(1), -1)
759 | # Update total number of images
760 | nimages += batch.size(0)
761 |
762 | # Compute mean of T * W * H
763 | # Sum over samples
764 | mean += batch.mean(2).sum(0)
765 | std += batch.std(2).sum(0)
766 | # Final step - divide by number of images to get average of each band
767 | mean /= nimages
768 | std /= nimages
769 | print(np.array(mean))
770 | return mean, std
771 |
772 |
773 | def get_mean_std_opera(dataset, batch_size=10):
774 | """Get the mean and stdev of the different image bands
775 | Args:
776 | times (list): timepoints to get mean from
777 | batch_size (int, optional): batch size for dataloader. Defaults to 1000.
778 | Returns:
779 | mean (list): list of band means
780 | stdev (list): list of band stdevs
781 | """
782 | loader = DataLoader(dataset, batch_size=batch_size)
783 | nimages = 0
784 | mean = 0.0
785 | std = 0.0
786 | count = 0
787 | for _, _, metadata in loader:
788 | print(count)
789 | count += 1
790 | batch = metadata["OPERA_input"]["data"]
791 | # Rearrange batch to be the shape of [B, C, T * W * H]
792 | batch = batch.view(batch.size(0), batch.size(1), -1)
793 | # Update total number of images
794 | nimages += batch.size(0)
795 |
796 | # Compute mean of T * W * H
797 | # Sum over samples
798 | mean += batch.mean(2).sum(0)
799 | std += batch.std(2).sum(0)
800 |
801 | # Final step - divide by number of images to get average of each band
802 | mean /= nimages
803 | std /= nimages
804 |
805 | return mean, std
806 |
807 |
808 | # ------------------------------------- GPU Functions------------------------------------------
809 |
810 |
811 | def get_cuda_memory_usage(gpus):
812 | """Get the GPU memory usage
813 |
814 | Args:
815 | gpus (list): list of GPUs
816 | """
817 | for gpu in gpus:
818 | r = torch.cuda.memory_reserved(gpu)
819 | a = torch.cuda.memory_allocated(gpu)
820 | f = r - a # free inside reserved
821 | print("GPU", gpu, "CUDA memory reserved:", r, "allocated:", a, "free:", f)
822 |
823 |
824 | # --------------------------------------- Crop files--------------------------------------------
825 |
826 |
827 | def crop_xarray(product, x_start, y_start, size):
828 | """Crop an xarray dataset
829 | - crop a squared region size
830 | - provide upper-left corner with (x_start, y_start)
831 | Args:
832 | product (xarray dataset): dataset to crop
833 | x_start (int): x start coordinate
834 | y_start (int): y start coordinate
835 | size (int): size of the crop
836 | Returns:
837 | xarray dataset: cropped dataset
838 | """
839 | return product.isel(
840 | x=slice(x_start, x_start + size), y=slice(y_start, y_start + size)
841 | )
842 |
843 |
844 | def prepare_crop(location, opera_k=1):
845 | """prepare the crop for both context and region of interest (roi)"""
846 |
847 | position = np.asarray(location["up_left"])
848 | context, roi = location["context"], location["roi"]
849 |
850 | # scale if opera
851 | if opera_k != 1:
852 | context, roi, position = [opera_k * s for s in [context, roi, position]]
853 | context_size = context + roi + context
854 |
855 | context_shape = {
856 | "x_start": position[0],
857 | "y_start": position[1],
858 | "size": context_size,
859 | }
860 | position = position + context
861 | roi_shape = {"x_start": position[0], "y_start": position[1], "size": roi}
862 | return context_shape, roi_shape
863 |
864 | # --------------------------------------- Save predictions --------------------------------------------
865 |
866 |
867 | def tensor_to_submission_file(predictions, predict_params):
868 | """saves prediction tesnor to submission .h5 file
869 |
870 | Args:
871 | predictions (numpy array): data cube of predictions
872 | predict_params (dict): dictionary of parameters for prediction
873 | """
874 |
875 | path = os.path.join(predict_params["submission_out_dir"],
876 | str(predict_params["year_to_predict"]))
877 | if not os.path.exists(path):
878 | os.makedirs(path)
879 |
880 | submission_file_name = predict_params["region_to_predict"] + ".pred.h5"
881 | submission_path = os.path.join(path, submission_file_name)
882 | h5f = h5py.File(submission_path, "w")
883 | h5f.create_dataset("submission", data=predictions.squeeze())
884 | h5f.close()
885 |
--------------------------------------------------------------------------------
/COPYING:
--------------------------------------------------------------------------------
1 | GNU GENERAL PUBLIC LICENSE
2 | Version 3, 29 June 2007
3 |
4 | Copyright (C) 2007 Free Software Foundation, Inc.
5 | Everyone is permitted to copy and distribute verbatim copies
6 | of this license document, but changing it is not allowed.
7 |
8 | Preamble
9 |
10 | The GNU General Public License is a free, copyleft license for
11 | software and other kinds of works.
12 |
13 | The licenses for most software and other practical works are designed
14 | to take away your freedom to share and change the works. By contrast,
15 | the GNU General Public License is intended to guarantee your freedom to
16 | share and change all versions of a program--to make sure it remains free
17 | software for all its users. We, the Free Software Foundation, use the
18 | GNU General Public License for most of our software; it applies also to
19 | any other work released this way by its authors. You can apply it to
20 | your programs, too.
21 |
22 | When we speak of free software, we are referring to freedom, not
23 | price. Our General Public Licenses are designed to make sure that you
24 | have the freedom to distribute copies of free software (and charge for
25 | them if you wish), that you receive source code or can get it if you
26 | want it, that you can change the software or use pieces of it in new
27 | free programs, and that you know you can do these things.
28 |
29 | To protect your rights, we need to prevent others from denying you
30 | these rights or asking you to surrender the rights. Therefore, you have
31 | certain responsibilities if you distribute copies of the software, or if
32 | you modify it: responsibilities to respect the freedom of others.
33 |
34 | For example, if you distribute copies of such a program, whether
35 | gratis or for a fee, you must pass on to the recipients the same
36 | freedoms that you received. You must make sure that they, too, receive
37 | or can get the source code. And you must show them these terms so they
38 | know their rights.
39 |
40 | Developers that use the GNU GPL protect your rights with two steps:
41 | (1) assert copyright on the software, and (2) offer you this License
42 | giving you legal permission to copy, distribute and/or modify it.
43 |
44 | For the developers' and authors' protection, the GPL clearly explains
45 | that there is no warranty for this free software. For both users' and
46 | authors' sake, the GPL requires that modified versions be marked as
47 | changed, so that their problems will not be attributed erroneously to
48 | authors of previous versions.
49 |
50 | Some devices are designed to deny users access to install or run
51 | modified versions of the software inside them, although the manufacturer
52 | can do so. This is fundamentally incompatible with the aim of
53 | protecting users' freedom to change the software. The systematic
54 | pattern of such abuse occurs in the area of products for individuals to
55 | use, which is precisely where it is most unacceptable. Therefore, we
56 | have designed this version of the GPL to prohibit the practice for those
57 | products. If such problems arise substantially in other domains, we
58 | stand ready to extend this provision to those domains in future versions
59 | of the GPL, as needed to protect the freedom of users.
60 |
61 | Finally, every program is threatened constantly by software patents.
62 | States should not allow patents to restrict development and use of
63 | software on general-purpose computers, but in those that do, we wish to
64 | avoid the special danger that patents applied to a free program could
65 | make it effectively proprietary. To prevent this, the GPL assures that
66 | patents cannot be used to render the program non-free.
67 |
68 | The precise terms and conditions for copying, distribution and
69 | modification follow.
70 |
71 | TERMS AND CONDITIONS
72 |
73 | 0. Definitions.
74 |
75 | "This License" refers to version 3 of the GNU General Public License.
76 |
77 | "Copyright" also means copyright-like laws that apply to other kinds of
78 | works, such as semiconductor masks.
79 |
80 | "The Program" refers to any copyrightable work licensed under this
81 | License. Each licensee is addressed as "you". "Licensees" and
82 | "recipients" may be individuals or organizations.
83 |
84 | To "modify" a work means to copy from or adapt all or part of the work
85 | in a fashion requiring copyright permission, other than the making of an
86 | exact copy. The resulting work is called a "modified version" of the
87 | earlier work or a work "based on" the earlier work.
88 |
89 | A "covered work" means either the unmodified Program or a work based
90 | on the Program.
91 |
92 | To "propagate" a work means to do anything with it that, without
93 | permission, would make you directly or secondarily liable for
94 | infringement under applicable copyright law, except executing it on a
95 | computer or modifying a private copy. Propagation includes copying,
96 | distribution (with or without modification), making available to the
97 | public, and in some countries other activities as well.
98 |
99 | To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies. Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 |
103 | An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License. If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 |
112 | 1. Source Code.
113 |
114 | The "source code" for a work means the preferred form of the work
115 | for making modifications to it. "Object code" means any non-source
116 | form of a work.
117 |
118 | A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 |
123 | The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form. A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 |
134 | The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities. However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work. For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 |
147 | The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 |
151 | The Corresponding Source for a work in source code form is that
152 | same work.
153 |
154 | 2. Basic Permissions.
155 |
156 | All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met. This License explicitly affirms your unlimited
159 | permission to run the unmodified Program. The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work. This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 |
164 | You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force. You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright. Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 |
175 | Conveying under any other circumstances is permitted solely under
176 | the conditions stated below. Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 |
179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 |
181 | No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 |
187 | When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 |
195 | 4. Conveying Verbatim Copies.
196 |
197 | You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 |
205 | You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 |
208 | 5. Conveying Modified Source Versions.
209 |
210 | You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 |
214 | a) The work must carry prominent notices stating that you modified
215 | it, and giving a relevant date.
216 |
217 | b) The work must carry prominent notices stating that it is
218 | released under this License and any conditions added under section
219 | 7. This requirement modifies the requirement in section 4 to
220 | "keep intact all notices".
221 |
222 | c) You must license the entire work, as a whole, under this
223 | License to anyone who comes into possession of a copy. This
224 | License will therefore apply, along with any applicable section 7
225 | additional terms, to the whole of the work, and all its parts,
226 | regardless of how they are packaged. This License gives no
227 | permission to license the work in any other way, but it does not
228 | invalidate such permission if you have separately received it.
229 |
230 | d) If the work has interactive user interfaces, each must display
231 | Appropriate Legal Notices; however, if the Program has interactive
232 | interfaces that do not display Appropriate Legal Notices, your
233 | work need not make them do so.
234 |
235 | A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit. Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 |
245 | 6. Conveying Non-Source Forms.
246 |
247 | You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 |
252 | a) Convey the object code in, or embodied in, a physical product
253 | (including a physical distribution medium), accompanied by the
254 | Corresponding Source fixed on a durable physical medium
255 | customarily used for software interchange.
256 |
257 | b) Convey the object code in, or embodied in, a physical product
258 | (including a physical distribution medium), accompanied by a
259 | written offer, valid for at least three years and valid for as
260 | long as you offer spare parts or customer support for that product
261 | model, to give anyone who possesses the object code either (1) a
262 | copy of the Corresponding Source for all the software in the
263 | product that is covered by this License, on a durable physical
264 | medium customarily used for software interchange, for a price no
265 | more than your reasonable cost of physically performing this
266 | conveying of source, or (2) access to copy the
267 | Corresponding Source from a network server at no charge.
268 |
269 | c) Convey individual copies of the object code with a copy of the
270 | written offer to provide the Corresponding Source. This
271 | alternative is allowed only occasionally and noncommercially, and
272 | only if you received the object code with such an offer, in accord
273 | with subsection 6b.
274 |
275 | d) Convey the object code by offering access from a designated
276 | place (gratis or for a charge), and offer equivalent access to the
277 | Corresponding Source in the same way through the same place at no
278 | further charge. You need not require recipients to copy the
279 | Corresponding Source along with the object code. If the place to
280 | copy the object code is a network server, the Corresponding Source
281 | may be on a different server (operated by you or a third party)
282 | that supports equivalent copying facilities, provided you maintain
283 | clear directions next to the object code saying where to find the
284 | Corresponding Source. Regardless of what server hosts the
285 | Corresponding Source, you remain obligated to ensure that it is
286 | available for as long as needed to satisfy these requirements.
287 |
288 | e) Convey the object code using peer-to-peer transmission, provided
289 | you inform other peers where the object code and Corresponding
290 | Source of the work are being offered to the general public at no
291 | charge under subsection 6d.
292 |
293 | A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 |
297 | A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling. In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage. For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product. A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 |
310 | "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source. The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 |
318 | If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information. But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 |
329 | The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed. Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 |
337 | Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 |
343 | 7. Additional Terms.
344 |
345 | "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law. If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 |
354 | When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it. (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.) You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 |
361 | Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 |
365 | a) Disclaiming warranty or limiting liability differently from the
366 | terms of sections 15 and 16 of this License; or
367 |
368 | b) Requiring preservation of specified reasonable legal notices or
369 | author attributions in that material or in the Appropriate Legal
370 | Notices displayed by works containing it; or
371 |
372 | c) Prohibiting misrepresentation of the origin of that material, or
373 | requiring that modified versions of such material be marked in
374 | reasonable ways as different from the original version; or
375 |
376 | d) Limiting the use for publicity purposes of names of licensors or
377 | authors of the material; or
378 |
379 | e) Declining to grant rights under trademark law for use of some
380 | trade names, trademarks, or service marks; or
381 |
382 | f) Requiring indemnification of licensors and authors of that
383 | material by anyone who conveys the material (or modified versions of
384 | it) with contractual assumptions of liability to the recipient, for
385 | any liability that these contractual assumptions directly impose on
386 | those licensors and authors.
387 |
388 | All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10. If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term. If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 |
398 | If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 |
403 | Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 |
407 | 8. Termination.
408 |
409 | You may not propagate or modify a covered work except as expressly
410 | provided under this License. Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 |
415 | However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 |
422 | Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 |
429 | Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License. If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 |
435 | 9. Acceptance Not Required for Having Copies.
436 |
437 | You are not required to accept this License in order to receive or
438 | run a copy of the Program. Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance. However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work. These actions infringe copyright if you do
443 | not accept this License. Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 |
446 | 10. Automatic Licensing of Downstream Recipients.
447 |
448 | Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License. You are not responsible
451 | for enforcing compliance by third parties with this License.
452 |
453 | An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations. If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 |
463 | You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License. For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 |
471 | 11. Patents.
472 |
473 | A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based. The
475 | work thus licensed is called the contributor's "contributor version".
476 |
477 | A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version. For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 |
487 | Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 |
492 | In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement). To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 |
499 | If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients. "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 |
513 | If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 |
521 | A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License. You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 |
536 | Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 |
540 | 12. No Surrender of Others' Freedom.
541 |
542 | If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License. If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all. For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 |
552 | 13. Use with the GNU Affero General Public License.
553 |
554 | Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work. The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 |
563 | 14. Revised Versions of this License.
564 |
565 | The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time. Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 |
570 | Each version is given a distinguishing version number. If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation. If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 |
579 | If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 |
584 | Later license versions may give you additional or different
585 | permissions. However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 |
589 | 15. Disclaimer of Warranty.
590 |
591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 |
600 | 16. Limitation of Liability.
601 |
602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 |
612 | 17. Interpretation of Sections 15 and 16.
613 |
614 | If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 |
621 | END OF TERMS AND CONDITIONS
622 |
623 | How to Apply These Terms to Your New Programs
624 |
625 | If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 |
629 | To do so, attach the following notices to the program. It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 |
634 |
635 | Copyright (C)
636 |
637 | This program is free software: you can redistribute it and/or modify
638 | it under the terms of the GNU General Public License as published by
639 | the Free Software Foundation, either version 3 of the License, or
640 | (at your option) any later version.
641 |
642 | This program is distributed in the hope that it will be useful,
643 | but WITHOUT ANY WARRANTY; without even the implied warranty of
644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645 | GNU General Public License for more details.
646 |
647 | You should have received a copy of the GNU General Public License
648 | along with this program. If not, see .
649 |
650 | Also add information on how to contact you by electronic and paper mail.
651 |
652 | If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 |
655 | Copyright (C)
656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 | This is free software, and you are welcome to redistribute it
658 | under certain conditions; type `show c' for details.
659 |
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License. Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 |
664 | You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | .
668 |
669 | The GNU General Public License does not permit incorporating your program
670 | into proprietary programs. If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library. If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License. But first, please read
674 | .
675 |
--------------------------------------------------------------------------------