ecm200 / caltech_birds Goto Github PK

At the moment both the base and ignite training scripts are not using the args module of the cub_tools package.

This has certain advantages, like passing through command line arguments key value pairs to override the YACS YAML configuration file.

Use the following code to in the scripts to create the argument parser with some default arguments:

from cub_tools.args import get_parser

parser = get_parser()
parser.add_argument # add additional local arguments to this specific training script if needed.

args = parser.parse_args()

When running "pip install ." in the cub_tools directory, installation fails saying incorrect syntax.
Missing "," after dependencies list.
Incorrect specification of PyTorch version, need to remove the version specification.

Enable the use of training scripts through the implementation of a CLI interface by moving the training scripts into the cub_tools package (or creating a separate package installer), which references the training scripts as a CLI.

The keyword EARLY_STOPPING_PATIENCE should be sitting in TRAIN section of the configuration but is currently sitting in the root.

_C.EARLY_STOPPING_PATIENCE = 5

Should be the following:

_C.TRAIN.EARLY_STOPPING_PATIENCE = 5

Hi,

I am interested in doing this, as it overlaps with my interests.

However I believe I am having issues getting the necessary dependencies installed. Scikit-Image has been the worst part of it. I have attempted building Scitkit from source, pip installs, everything I can find in the past day.

Pip always hangs for when I am running the install command in the cub_tools directory. Everything else seems ok.

I am wondering if you can help me get the dependencies built to begin working this project.

The setup.py file of the cub_tools package does not have any package dependencies specified.

Need to add dependencies for installation.

Also need to create references.txt for full list of dependencies to run the notebooks.

When I run 'script/train_pytorchcv_InceptionV4_caltech_birds.py', I get an error of 'ModuleNotFoundError: No module named 'pytorchcv.utils''.

My pytorchcv version is 0.058, and I install through pip.

Thanks.

Use the standalone torchmetrics package for all metrics usage across the functions.

Use YACS library to implement YAML configuration files.

The get_parser() function currently overrides the --config CLI argument to a default filename.

This behaviour should be changed to allow no configuration file to be given, and thus allowing for instance where all configuration is applied using the CLI --opts arguments pathway.

setup.py depedencies need updating to match those of the conda_dependencies.yml file.

At the moment, installation issues will occur if installation is attempted of the package first.
This is related to the issues with PyTorch 1.8.1 through PYPI rather than CONDA.

This issue may have been fixed, but either way, the dependencies need updating.

Need to think about how to allow for custom events or more control over callbacks and events with the configuration file.

Originally posted by @ecm200 in #19 (comment)

Hi,

I am attempting to train the model using the code you have in the example notebooks. However I am running out of memory, and I am wondering if you can suggest how to alter the code to reduce memory usage during training.