cfm-task-models 0.1.0

CFM - model zoo

CFM-Task-Models

Popular CV models modified for various approaches of Compression for Machines (aka Coding for Machines)

Installation

The easiest way to install CFM-Task-Models is through PyPI, to do this simply use pip install cfm-task-models However, CFM-Task-Models has requirements that currently cannot be handled by pip, thus before using CFM-Task-Models for the first time please run miminstaller.py in your virutal environment. If using poetry consider: poetry run python -m cfm_task_models.miminstaller

Requirements

CFM-Task-Models currently relies on several tools from OpenMMLab, which require custom installation using the openmim installer. Openmim is a tool provided by OpenMMLab which installs their libraries based on the user's pytorch and cuda versions.

See pyproject.toml for standard requirements and miminstaller.py for OpenMMLAB requirements

Using conda

conda env create -f environment.yml

Usage

To test Swin_Transformer, run the following command from the root directory:

python models/Swin-Transformer/models/swin_transformer_v2.py

Semantic Segmentation

Dataset

To downald the ADE20K dataset, run the following command from the root directory:

python cfm_task_models/dataset_download.py --dataset-name ade20k_2016 --save-dir data/ade --unzip --delete

Model Zoo

The model configs can be found in configs/segmentation.

ADE20K dataset

Model	Backbone	config	checkpoint	mIoU
UperNet	Swin-T	UPerNet	download model	44.41
Mask2Former	Swin-T	Mask2Former	download model	48.66
KNet + UperNet	Swin-T	KNet	download model	45.84

Download from mim (#archive):

mim download mmsegmentation --config swin-tiny-patch4-window7-in1k-pre_upernet_8xb2-160k_ade20k-512x512 --dest ./configs/segmentation

mim download mmsegmentation --config mask2former_swin-t_8xb2-160k_ade20k-512x512 --dest ./configs/segmentation

mim download mmsegmentation --config knet-s3_swin-t_upernet_8xb2-adamw-80k_ade20k-512x512 --dest ./configs/segmentation

Evaluation

To evaluate the model on the ADE20K dataset, run the following command from the root directory:

python cfm_task_models/run_mmseg_swin_ade20k_eval.py --device cuda --checkpoint path/to/checkpoint --config path/to/config --output result/file/path --data-root path/to/data

For example : python cfm_task_models/run_mmseg_swin_ade20k_eval.py --device cuda --checkpoint checkpoints/mask2former_swin-t_8xb2-160k_ade20k-512x512_20221203_234230-7d64e5dd.pth --config configs/segmentation/mask2former_swin-t_8xb2-160k_ade20k-512x512.py --output result.json --data-root /home/data/ade/ADEChallengeData2016/