TorchOk
The toolkit for fast Deep Learning experiments in Computer Vision
What is it?
The toolkit consists of:
- Popular neural network models and custom modules implementations used in our company
- Metrics used in CV such that mIoU, mAP, etc.
- Commonly used datasets and data loaders
The framework is based on PyTorch and utilizes PyTorch Lightning for training pipeline routines.
Installation
Docker
One of the ways to install TorchOk is to use Docker:
docker build -t torchok --build-arg SSH_PUBLIC_KEY="" . docker run -d --name <username>_torchok --gpus=all -v <path/to/workdir>:/workdir -p <ssh_port>:22 -p <jupyter_port>:8888 -p <tensorboard_port>:6006 torchok
Conda
To remove previous installation of TorchOk environment, run:
conda remove --name torchok --all
To install TorchOk locally, run:
conda env create -f environment.yml
This will create a new conda environment torchok with all dependencies.
Getting started
Training is configured by YAML configuration files which each forked project should store inside configs folder (see configs/cifar10.yml for example). The configuration supports environment variables substitution, so that you can easily change base directory paths without changing the config file for each environment. The most common environment variables are:
SM_CHANNEL_TRAINING — directory to all training data
SM_OUTPUT_DATA_DIR — directory where logs for all runs will be stored SM_NUM_CPUS - number of used CPUs for dataloader
Start training locally
Download CIFAR10 dataset running all cells in notebooks/Cifar10.ipynb, the dataset will appear in data/cifar10 folder.
docker exec -it torchok bash
cd torchok
SM_NUM_CPUS=8 SM_CHANNEL_TRAINING=./data/cifar10 SM_OUTPUT_DATA_DIR=/tmp python train.py --config config/classification_resnet_example.yml
Start SageMaker Training Jobs
Start the job using one of the AWS SageMaker instances. You have 2 ways to provide data inside your training container:
- Slow downloaded S3 bucket:
s3://. Volume size is needed to be set when you use S3 bucket. For other cases it can be omitted./ - Fast FSx access:
fsx://. To create FSx filesystem follow this instructions/ /
Example with S3:
python run_sagemaker.py --config configs/cifar10.yml --input_path s3://sagemaker-mlflow-main/cifar10 --instance_type ml.g4dn.xlarge --volume_size 5
Example with FSx:
python run_sagemaker.py --input_path fsx://fs-0f79df302dcbd29bd/z6duzbmv/tz_jpg --config configs/siloiz_pairwise_xbm_resnet50_512d.yml --instance_type ml.g4dn.xlarge
In case something isn't working inside the Sagemaker container you can debug your model locally. Specify local_gpu instance type when starting the job:
python run_sagemaker.py --config configs/cifar10.yml --instance_type local_gpu --volume_size 5 --input_path file://../data/cifar10
Run tests
docker exec -it torchok bash
cd torchok
python -m unittest discover -s tests/ -p "test_*.py"
Differences in configs sagemaker vs local machine
1. Path to data folder
sagemaker
data:
dataset_name: ExampleDataset
common_params:
data_folder: "${SM_CHANNEL_TRAINING}"
local machine
data:
dataset_name: ExampleDataset
common_params:
data_folder: "/path/to/data"
2. Path to artifacts dir
sagemaker
log_dir: '/opt/ml/checkpoints'
local machine
log_dir: '/tmp/logs'
3. Restore path
do_restore is a special indicator which was designed to be used for SageMaker spot instances training. With this indicator you can debug your model locally and be free to leave the restore_path pointing to some common directory like /opt/ml/checkpoints, where TorchOk will search the checkpoints for.
sagemaker
restore_path: '/opt/ml/checkpoints'
do_restore: '${SM_USER_ENTRY_POINT}'
local machine
restore_path: '/opt/ml/checkpoints'
do_restore: '${SM_USER_ENTRY_POINT}'
Mlflow
To have more convenient logs it is recommended to name your experiment as project_name-developer_name, so that all your experiments related to this project will be under one tag in mlflow
experiment_name: &experiment_name fips-roman
State all the model parameters in mlflow.runName in logger params
logger:
logger: mlflow
experiment_name: *experiment_name
tags:
mlflow.runName: "siloiz_contrastive_xbm_resnet50_512d"
save_dir: "s3://sagemaker-mlflow-main/mlruns"
secrets_manager:
region: "eu-west-1"
mlflow_secret: "acme/mlflow"
57 Oct 22, 2022
242 Jan 08, 2023
278 Jan 01, 2023
406 Jan 04, 2023
25 Nov 24, 2022
55 Dec 24, 2022
3 Oct 17, 2021
433 Nov 24, 2022
117 Jan 09, 2023
8 Nov 24, 2022
37 Sep 15, 2022
17 Dec 12, 2022
25.5k Jan 07, 2023
73 Dec 09, 2022
722 Jan 08, 2023
1.9k Jan 06, 2023
153 Nov 28, 2022
1.2k Jan 05, 2023
4 Feb 11, 2022
18 Aug 27, 2022