I would like to find out the size of conda packages to delete the huge and seldom used ones. Which conda command should I use to find out the package size?
conda list will list the packages but does not show the package size.
I welcome other methods to find out package size.
I'm using Windows 10.
One way to get at this is by mining the JSON metadata files for each package inside an environment's conda-meta/ directory. There are two types of sizes listed:
Since you seem interested in total size of a package, let's do the simpler size. This will let us get a quick ranking of packages by their download size.
Commands
## activate the environment of interest
conda activate foo
## search all the JSONs for '"size":'
grep '"size":' ${CONDA_PREFIX}/conda-meta/*.json |\
## sort result
sort -k3rn |\
## show only filename
sed 's/.*conda-meta\///g' |\
## print with columns
column -t
Example Output
jaxlib-0.1.67-py39h6e9494a_0.json: "size": 38576847,
scipy-1.6.3-py39h056f1c0_0.json: "size": 19495906,
python-3.9.4-h9133fd0_0_cpython.json: "size": 13160553,
libopenblas-0.3.15-openmp_h5e1b9a4_1.json: "size": 9163719,
numpy-1.20.3-py39h7eed0ac_1.json: "size": 5732039,
tk-8.6.10-hb0a8c7a_1.json: "size": 3420669,
openssl-1.1.1k-h0d85af4_0.json: "size": 1985060,
sqlite-3.35.5-h44b9ce1_0.json: "size": 1810221,
libgfortran5-9.3.0-h6c81a4c_22.json: "size": 1766473,
pip-21.1.2-pyhd8ed1ab_0.json: "size": 1147500,
libcxx-11.1.0-habf9029_0.json: "size": 1055976,
setuptools-49.6.0-py39h6e9494a_3.json: "size": 972968,
ncurses-6.2-h2e338ed_4.json: "size": 901840,
jax-0.2.14-pyhd8ed1ab_0.json: "size": 571585,
llvm-openmp-11.1.0-hda6cdc1_1.json: "size": 274368,
readline-8.1-h05e3726_0.json: "size": 272444,
xz-5.2.5-haf1e3a3_1.json: "size": 233058,
certifi-2021.5.30-py39h6e9494a_0.json: "size": 144599,
ca-certificates-2021.5.30-h033912b_0.json: "size": 139088,
tzdata-2021a-he74cb21_0.json: "size": 123802,
zlib-1.2.11-h7795811_1010.json: "size": 104180,
absl-py-0.12.0-pyhd8ed1ab_0.json: "size": 98565,
tqdm-4.61.0-pyhd8ed1ab_0.json: "size": 81513,
opt_einsum-3.3.0-pyhd8ed1ab_1.json: "size": 54494,
libffi-3.3-h046ec9c_2.json: "size": 46425,
wheel-0.36.2-pyhd3deb0d_0.json: "size": 31381,
python-flatbuffers-2.0-pyhd8ed1ab_0.json: "size": 28606,
libgfortran-5.0.0-9_3_0_h6c81a4c_22.json: "size": 19280,
six-1.16.0-pyh6c4a22f_0.json: "size": 14259,
libblas-3.9.0-9_openblas.json: "size": 11762,
libcblas-3.9.0-9_openblas.json: "size": 11671,
liblapack-3.9.0-9_openblas.json: "size": 11671,
python_abi-3.9-1_cp39.json: "size": 3921,
The above output shows jaxlib is the largest package, followed by scipy, then the python interpreter itself. In this case, if I wanted to remove jaxlib it would also entail removing jax.
I think the above serves as a first approximation for ranking the packages by size. The size_in_bytes could possibly be more exact, but to be thorough one would need to also consider which individual files are hardlinked, since those actually shouldn't be counted against the package on a per environment level. For them, there is only one copy per system and it gets reused across environments.
If you are concerned about using up disk space, you can run this command to clean up all the temporary packages, zip files, etc that conda used to setup your environment.
conda clean --all
These files remain and can clutter up your disk over time.
use
du -h -s $(conda info --base)/envs/test-1
for the file size of the package
You can use pip module through subprocess communication (Python >= 3.7):
from pip._internal.commands.show import search_packages_info
import subprocess
import sys
import json
import pathlib
import math
import operator
TMPL = '{:<30} {:<30} {:<10} {:>9}'
def humanize(size):
"""Convert size in bytes to human readable output."""
try:
units = ['B', 'KB', 'MB', 'GB']
i = math.floor(math.log(size) / math.log(1024))
hsize = f"{round(size / 1024**i)}{units[i]}"
except ValueError:
hsize = '0B'
return hsize
def pip_list_ext():
"""List installed packages with additional information."""
# python -m pip list --format --json --verbose
pip_list = subprocess.run([sys.executable, '-m', 'pip', 'list',
'--format', 'json', '--verbose'],
capture_output=True)
if pip_list.returncode != 0:
raise RuntimeError('...')
query = [pkg['name'] for pkg in json.loads(pip_list.stdout)]
packages = []
for pkg in search_packages_info(query):
location = pathlib.Path(pkg['location'])
pkg['installer'] = pkg.get('installer', '')
pkg['files'] = pkg.get('files', [])
size = sum((location / f).stat().st_size
for f in pkg['files'] if (location / f).is_file())
pkg['size'] = size
packages.append(pkg)
return packages
if __name__ == '__main__':
packages = pip_list_ext()
print(TMPL.format('Package', 'Version', 'Installer', 'Size'))
print(TMPL.format('-'*30, '-'*30, '-'*10, '-'*9))
for pkg in packages:
print(TMPL.format(pkg['name'], pkg['version'],
pkg['installer'], humanize(pkg['size'])))
# TOP 10
N = 10
pkgs = sorted(packages, key=operator.itemgetter('size'), reverse=True)
print(f"\nTOP {N}:")
for i, pkg in enumerate(pkgs[:N], 1):
print(f"{i:>2}. {pkg['name']} ({humanize(pkg['size'])})")
Sample output
Package Version Installer Size
------------------------------ ------------------------------ ---------- ---------
appnope 0.1.2 conda 13KB
backcall 0.2.0 conda 58KB
certifi 2021.5.30 261KB
cffi 1.14.5 conda 780KB
cryptography 3.4.7 conda 2MB
decorator 5.0.9 conda 35KB
ipython 7.22.0 conda 4MB
ipython-genutils 0.2.0 conda 66KB
jedi 0.17.2 conda 5MB
mkl-fft 1.3.0 0B
mkl-random 1.2.1 conda 1MB
mkl-service 2.3.0 0B
numpy 1.20.2 conda 23MB
pandas 1.2.4 27MB
parso 0.7.0 conda 320KB
pexpect 4.8.0 conda 177KB
pickleshare 0.7.5 conda 23KB
pip 21.1.2 0B
prompt-toolkit 3.0.17 conda 2MB
ptyprocess 0.7.0 conda 40KB
pycparser 2.20 conda 547KB
Pygments 2.9.0 conda 6MB
pyOpenSSL 20.0.1 conda 212KB
python-dateutil 2.8.1 conda 619KB
pytz 2021.1 conda 971KB
setuptools 52.0.0.post20210125 0B
six 1.15.0 conda 63KB
traitlets 5.0.5 conda 798KB
wcwidth 0.2.5 conda 411KB
wheel 0.36.2 0B
TOP 10:
1. pandas (27MB)
2. numpy (23MB)
3. Pygments (6MB)
4. jedi (5MB)
5. ipython (4MB)
6. cryptography (2MB)
7. prompt-toolkit (2MB)
8. mkl-random (1MB)
9. pytz (971KB)
10. traitlets (798KB)
The du tool can tell you how much disk space is occupied by a particular directory.
For example, create an identical conda environment (Miniconda) and check its disk usage like this:
$ conda create -n testenv -y python numpy pandas
$ du -h -s $(conda info --base)/envs/testenv
1.2G /opt/miniconda/envs/testenv
This solution works for Windows 10.
Install Windows SysInternals so that you have the command utility du.
Run the du command to find out the packages size under site-packages in Anaconda and store result in csv file.
$ du -c -v Lib/site-packages/ > packages.csv
Open packages.csv using Excel. Sort the column for the package size in descending order. Now, you have the information on the packages with the largest package size.