Changing default directory for torch models

[jagin (Jarosław Gilewski)]

Hi,

If I run for example

learn = create_cnn(data, models.resnet18, metrics=accuracy)

the model is loaded into /root/.torch/models as it is shown here:

Downloading: "https://download.pytorch.org/models/resnet18-5c106cde.pth" to /root/.torch/models/resnet18-5c106cde.pth 100%|██████████| 46827520/46827520 [00:00<00:00, 73938844.51it/s]

My question is: can I change the directory where the model is loaded?

Regards

Jarek

[zenlytix]

Pl try this.

How to set Config.data_path() to pwd?

You can override that in your code like follows (replace with desired value and name of your config file): Config.DEFAULT_CONFIG = { 'data_path': '/tmp/fastai/data', 'model_path': '/tmp/fastai/models' } Config.create('/tmp/myconfig.yml') Config.DEFAULT_CONFIG_PATH = '/tmp/myconfig.yml' In subsequent notebooks you just need to set the DEFAULT_CONFIG_PATH to your config file.

[jagin (Jarosław Gilewski)]

I have ~/.fastai/config.yml setup so the data and model are loaded into the /workspace/fastai/data and /workspace/fastai/model respectively.

It works properly for path = untar_data(URLs.PETS); path for example. The result is PosixPath('/workspace/fastai/data/oxford-iiit-pet') but still create_cnn loads the model into ~/.torch/models/

[gamo (Gabriel)]

You can override the default pytorch model_dir by changing the env variable,
os.environ['TORCH_MODEL_ZOO'] = '<path>'

[jagin (Jarosław Gilewski)]

That’s what I need. Thank you!

[AlexeyRB (Alexey)]

You can also use os.rename(old_path, new_path) if your files are uploaded incorrectly. This will move your files to the new directory. For example:

folders = [‘black’, ‘teddys’, ‘grizzly’]
files = [‘urls_black.csv’, ‘urls_teddys.csv’, ‘urls_grizzly.csv’]
for folder, file in zip(folders, files):
path = ‘data/bears’
os.rename(os.path.join(os.getcwd(), file),
os.path.join(os.getcwd(), path, file))

[gil3co (Gil Cohen)]

Great advice!
Please notice that “TORCH_MODEL_ZOO” is deprecated, use “TORCH_HOME” instead.