14 KiB

Raw Blame History

MIM: MIM Installs OpenMMLab Packages

MIM provides a unified API for launching and installing OpenMMLab projects and their extensions, and managing the OpenMMLab model zoo.

Major Features

Package Management

You can use MIM to manage OpenMMLab codebases, install or uninstall them conveniently.
Checkpoint Management

You can use MIM to access all checkpoints in OpenMMLab, download checkpoints, look up checkpoints that meet your need.
Script Calling

You can call training scripts, testing scripts and any other scripts under the tools directory of a specific codebase conveniently at anywhere. Calling scripts via MIM is more flexible and efficient (The command will be much shorter, check abbreviation.md).

Installation

Please refer to installation.md for installation.

Command

1. install

command

# install latest version of mmcv-full
> mim install mmcv-full  # wheel
# install 1.3.1
> mim install mmcv-full==1.3.1
# install master branch
> mim install mmcv-full -f https://github.com/open-mmlab/mmcv.git

# install latest version of mmcls
> mim install mmcls
# install 0.11.0
> mim install mmcls==0.11.0  # v0.11.0
# install master branch
> mim install mmcls -f https://github.com/open-mmlab/mmclassification.git
# install local repo
> git clone https://github.com/open-mmlab/mmclassification.git
> cd mmclassification
> mim install .

# install extension based on OpenMMLab
mim install mmcls-project -f https://github.com/xxx/mmcls-project.git

api

from mim import install

# install mmcv
install('mmcv-full')

# install mmcls
# install mmcls will automatically install mmcv if it is not installed
install('mmcv-full', find_url='https://github.com/open-mmlab/mmcv.git')
install('mmcv-full==1.3.1', find_url='https://github.com/open-mmlab/mmcv.git')

# install extension based on OpenMMLab
install('mmcls-project', find_url='https://github.com/xxx/mmcls-project.git')

2. uninstall

[![asciicast](https://asciinema.org/a/416948.svg)](https://asciinema.org/a/416948)

command

# uninstall mmcv
> mim uninstall mmcv-full

# uninstall mmcls
> mim uninstall mmcls

api

from mim import uninstall

# uninstall mmcv
uninstall('mmcv-full')

# uninstall mmcls
uninstall('mmcls)

3. list

command
```
> mim list
> mim list --all
```

api

from mim import list_package

list_package()
list_package(True)

4. search

command

> mim search mmcls
> mim search mmcls==0.11.0 --remote
> mim search mmcls --config resnet18_b16x8_cifar10
> mim search mmcls --model resnet
> mim search mmcls --dataset cifar-10
> mim search mmcls --valid-field
> mim search mmcls --condition 'bs>45,epoch>100'
> mim search mmcls --condition 'bs>45 epoch>100'
> mim search mmcls --condition '128<bs<=256'
> mim search mmcls --sort bs epoch
> mim search mmcls --field epoch bs weight
> mim search mmcls --exclude-field weight paper

api

from mim import get_model_info

get_model_info('mmcls')
get_model_info('mmcls==0.11.0', local=False)
get_model_info('mmcls', models=['resnet'])
get_model_info('mmcls', training_datasets=['cifar-10'])
get_model_info('mmcls', filter_conditions='bs>45,epoch>100')
get_model_info('mmcls', filter_conditions='bs>45 epoch>100')
get_model_info('mmcls', filter_conditions='128<bs<=256')
get_model_info('mmcls', sorted_fields=['bs', 'epoch'])
get_model_info('mmcls', shown_fields=['epoch', 'bs', 'weight'])

5. download

command

> mim download mmcls --config resnet18_b16x8_cifar10
> mim download mmcls --config resnet18_b16x8_cifar10 --dest .

api

from mim import download

download('mmcls', ['resnet18_b16x8_cifar10'])
download('mmcls', ['resnet18_b16x8_cifar10'], dest_dir='.')

6. train

command

# Train models on a single server with CPU by setting `gpus` to 0 and
# 'launcher' to 'none' (if applicable). The training script of the
# corresponding codebase will fail if it doesn't support CPU training.
> mim train mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 0
# Train models on a single server with one GPU
> mim train mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 1
# Train models on a single server with 4 GPUs and pytorch distributed
> mim train mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 4 \
    --launcher pytorch
# Train models on a slurm HPC with one 8-GPU node
> mim train mmcls resnet101_b16x8_cifar10.py --launcher slurm --gpus 8 \
    --gpus-per-node 8 --partition partition_name --work-dir tmp
# Print help messages of sub-command train
> mim train -h
# Print help messages of sub-command train and the training script of mmcls
> mim train mmcls -h

api

from mim import train

train(repo='mmcls', config='resnet18_b16x8_cifar10.py', gpus=0,
      other_args='--work-dir tmp')
train(repo='mmcls', config='resnet18_b16x8_cifar10.py', gpus=1,
      other_args='--work-dir tmp')
train(repo='mmcls', config='resnet18_b16x8_cifar10.py', gpus=4,
      launcher='pytorch', other_args='--work-dir tmp')
train(repo='mmcls', config='resnet18_b16x8_cifar10.py', gpus=8,
      launcher='slurm', gpus_per_node=8, partition='partition_name',
      other_args='--work-dir tmp')

7. test

command

# Test models on a single server with 1 GPU, report accuracy
> mim test mmcls resnet101_b16x8_cifar10.py --checkpoint \
    tmp/epoch_3.pth --gpus 1 --metrics accuracy
# Test models on a single server with 1 GPU, save predictions
> mim test mmcls resnet101_b16x8_cifar10.py --checkpoint \
    tmp/epoch_3.pth --gpus 1 --out tmp.pkl
# Test models on a single server with 4 GPUs, pytorch distributed,
# report accuracy
> mim test mmcls resnet101_b16x8_cifar10.py --checkpoint \
    tmp/epoch_3.pth --gpus 4 --launcher pytorch --metrics accuracy
# Test models on a slurm HPC with one 8-GPU node, report accuracy
> mim test mmcls resnet101_b16x8_cifar10.py --checkpoint \
    tmp/epoch_3.pth --gpus 8 --metrics accuracy --partition \
    partition_name --gpus-per-node 8 --launcher slurm
# Print help messages of sub-command test
> mim test -h
# Print help messages of sub-command test and the testing script of mmcls
> mim test mmcls -h

api

from mim import test
test(repo='mmcls', config='resnet101_b16x8_cifar10.py',
     checkpoint='tmp/epoch_3.pth', gpus=1, other_args='--metrics accuracy')
test(repo='mmcls', config='resnet101_b16x8_cifar10.py',
     checkpoint='tmp/epoch_3.pth', gpus=1, other_args='--out tmp.pkl')
test(repo='mmcls', config='resnet101_b16x8_cifar10.py',
     checkpoint='tmp/epoch_3.pth', gpus=4, launcher='pytorch',
     other_args='--metrics accuracy')
test(repo='mmcls', config='resnet101_b16x8_cifar10.py',
     checkpoint='tmp/epoch_3.pth', gpus=8, partition='partition_name',
     launcher='slurm', gpus_per_node=8, other_args='--metrics accuracy')

8. run

command

# Get the Flops of a model
> mim run mmcls get_flops resnet101_b16x8_cifar10.py
# Publish a model
> mim run mmcls publish_model input.pth output.pth
# Train models on a slurm HPC with one GPU
> srun -p partition --gres=gpu:1 mim run mmcls train \
    resnet101_b16x8_cifar10.py --work-dir tmp
# Test models on a slurm HPC with one GPU, report accuracy
> srun -p partition --gres=gpu:1 mim run mmcls test \
    resnet101_b16x8_cifar10.py tmp/epoch_3.pth --metrics accuracy
# Print help messages of sub-command run
> mim run -h
# Print help messages of sub-command run, list all available scripts in
# codebase mmcls
> mim run mmcls -h
# Print help messages of sub-command run, print the help message of
# training script in mmcls
> mim run mmcls train -h

api

from mim import run

run(repo='mmcls', command='get_flops',
    other_args='resnet101_b16x8_cifar10.py')
run(repo='mmcls', command='publish_model',
    other_args='input.pth output.pth')
run(repo='mmcls', command='train',
    other_args='resnet101_b16x8_cifar10.py --work-dir tmp')
run(repo='mmcls', command='test',
    other_args='resnet101_b16x8_cifar10.py tmp/epoch_3.pth --metrics accuracy')

9. gridsearch

command

# Parameter search on a single server with CPU by setting `gpus` to 0 and
# 'launcher' to 'none' (if applicable). The training script of the
# corresponding codebase will fail if it doesn't support CPU training.
> mim gridsearch mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 0 \
    --search-args '--optimizer.lr 1e-2 1e-3'
# Parameter search with on a single server with one GPU, search learning
# rate
> mim gridsearch mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 1 \
    --search-args '--optimizer.lr 1e-2 1e-3'
# Parameter search with on a single server with one GPU, search
# weight_decay
> mim gridsearch mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 1 \
    --search-args '--optimizer.weight_decay 1e-3 1e-4'
# Parameter search with on a single server with one GPU, search learning
# rate and weight_decay
> mim gridsearch mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 1 \
    --search-args '--optimizer.lr 1e-2 1e-3 --optimizer.weight_decay 1e-3 \
    1e-4'
# Parameter search on a slurm HPC with one 8-GPU node, search learning
# rate and weight_decay
> mim gridsearch mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 8 \
    --partition partition_name --gpus-per-node 8 --launcher slurm \
    --search-args '--optimizer.lr 1e-2 1e-3 --optimizer.weight_decay 1e-3 \
    1e-4'
# Parameter search on a slurm HPC with one 8-GPU node, search learning
# rate and weight_decay, max parallel jobs is 2
> mim gridsearch mmcls resnet101_b16x8_cifar10.py --work-dir tmp --gpus 8 \
    --partition partition_name --gpus-per-node 8 --launcher slurm \
    --max-workers 2 --search-args '--optimizer.lr 1e-2 1e-3 \
    --optimizer.weight_decay 1e-3 1e-4'
# Print the help message of sub-command search
> mim gridsearch -h
# Print the help message of sub-command search and the help message of the
# training script of codebase mmcls
> mim gridsearch mmcls -h

api

from mim import gridsearch

gridsearch(repo='mmcls', config='resnet101_b16x8_cifar10.py', gpus=0,
           search_args='--optimizer.lr 1e-2 1e-3',
           other_args='--work-dir tmp')
gridsearch(repo='mmcls', config='resnet101_b16x8_cifar10.py', gpus=1,
           search_args='--optimizer.lr 1e-2 1e-3',
           other_args='--work-dir tmp')
gridsearch(repo='mmcls', config='resnet101_b16x8_cifar10.py', gpus=1,
           search_args='--optimizer.weight_decay 1e-3 1e-4',
           other_args='--work-dir tmp')
gridsearch(repo='mmcls', config='resnet101_b16x8_cifar10.py', gpus=1,
           search_args='--optimizer.lr 1e-2 1e-3 --optimizer.weight_decay'
                       '1e-3 1e-4',
           other_args='--work-dir tmp')
gridsearch(repo='mmcls', config='resnet101_b16x8_cifar10.py', gpus=8,
           partition='partition_name', gpus_per_node=8, launcher='slurm',
           search_args='--optimizer.lr 1e-2 1e-3 --optimizer.weight_decay'
                       ' 1e-3 1e-4',
           other_args='--work-dir tmp')
gridsearch(repo='mmcls', config='resnet101_b16x8_cifar10.py', gpus=8,
           partition='partition_name', gpus_per_node=8, launcher='slurm',
           max_workers=2,
           search_args='--optimizer.lr 1e-2 1e-3 --optimizer.weight_decay'
                       ' 1e-3 1e-4',
           other_args='--work-dir tmp')

Build custom projects with MIM

We provide some examples about how to build custom projects based on OpenMMLAB codebases and MIM in MIM-Example. In mmcls_custom_backbone, we define a custom backbone and a classification config file that uses the backbone. To train this model, you can use the command:

# The working directory is `mim-example/mmcls_custom_backbone`
PYTHONPATH=$PWD:$PYTHONPATH mim train mmcls custom_net_config.py --work-dir tmp --gpus 1

Contributing

We appreciate all contributions to improve mim. Please refer to CONTRIBUTING.md for the contributing guideline.

License

This project is released under the Apache 2.0 license.

14 KiB Raw Blame History

MIM: MIM Installs OpenMMLab Packages

Major Features

Installation

Command

Build custom projects with MIM

Contributing

License

14 KiB

Raw Blame History