This repository contains the code of our method presented at ICIP'22.
This repository contains the following files:
|-- Dockerfile
|-- commands.sh # bash file containing commands to run for reproducing experiments
|-- config
| -- parser.py # Argument parser
|-- logs # log files
|-- main.py
|-- readme.md # This file
|-- results # Folder containing experiments results
|-- requirements.txt
|-- src
|-- buffers # Memory strategies
| |-- buffer.py
| |-- reservoir.py
|-- datasets
| |-- __init__.py
| |-- cifar10.py
| |-- cifar100.py
| |-- split_cifar10.py
| |-- split_cifar100.py
| |-- split_tiny.py
| |-- tinyImageNet.py
|-- learners # Methods compared in the publications
| |-- base.py
| |-- baselines
| | |-- er.py
| | |-- scr.py
| |-- ce.py
| |-- sscr.py
| |-- supcon.py
|-- models # Model used
| |-- resnet.py
|-- utils
|-- data.py
|-- losses.py
|-- metrics.py
|-- name_match.py # Matching parser arguments and classes
|-- tensorboard.py
|-- utils.pyTo run the code, you have to install the necessary dependencies. It is advise to use either a virtual environment or docker
pip install torch==1.9.0
pip install -r requirements.txt
docker build -t ossgcl .
docker run -it --rm --name [CONTAINER_NAME] -v $PWD:/workspace -v [LOCAL_DATA_FOLDER]:/data --gpus all --shm-size=8G ossgcl
Command line usage of the current repository with available parameters.
usage: main.py [-h] [--epochs EPOCHS] [--start-epoch START_EPOCH]
[-b BATCH_SIZE] [--learning-rate LEARNING_RATE]
[--momentum MOMENTUM] [--weight-decay WEIGHT_DECAY]
[--optim {Adam,SGD}] [--seed SEED] [--memory-only]
[--resume RESUME] [--tag TAG] [--tb-root TB_ROOT]
[--print-freq PRINT_FREQ] [--verbose] [--logs LOGS]
[--logs-root LOGS_ROOT] [--results-root RESULTS_ROOT]
[--ckpt-root CKPT_ROOT] [--data-root-dir DATA_ROOT_DIR]
[--min-crop MIN_CROP] [--dataset {cifar10,cifar100,tiny}]
[--training-type {uni,inc}] [--n-classes N_CLASSES]
[--img-size IMG_SIZE] [--num-workers NUM_WORKERS]
[--n-tasks N_TASKS]
[--labels-order LABELS_ORDER [LABELS_ORDER ...]]
[--nb-channels NB_CHANNELS] [--temperature TEMPERATURE]
[--alpha ALPHA] [--mem-size MEM_SIZE]
[--mem-batch-size MEM_BATCH_SIZE] [--buffer {reservoir}]
[--drop-method {random}] [--learner {ER,CE,SCR,SSCL,SC}]
[--eval-mem] [--eval-random] [--n-runs N_RUNS]
[--run-id RUN_ID] [--lab-pc LAB_PC] [--mem-iters MEM_ITERS]
[--review] [--rv-iters RV_ITERS]
Pytorch Continual learning repo.
optional arguments:
-h, --help show this help message and exit
--epochs EPOCHS number of total epochs to run
--start-epoch START_EPOCH
manual epoch number (useful on restarts)
-b BATCH_SIZE, --batch-size BATCH_SIZE
Batch size for stream data (default: 10)
--learning-rate LEARNING_RATE, -lr LEARNING_RATE
Initial learning rate
--momentum MOMENTUM momentum
--weight-decay WEIGHT_DECAY, --wd WEIGHT_DECAY
weight decay (default: 0)
--optim {Adam,SGD}
--seed SEED Random seed to use.
--memory-only, -mo Training using only the memory ?
--resume RESUME, -r RESUME
Resume old training. Load model given as resume
parameter
--tag TAG, -t TAG Base name for graphs and checkpoints
--tb-root TB_ROOT Where do you want tensorboards graphs ?
--print-freq PRINT_FREQ, -p PRINT_FREQ
Number of batches between each print (default: 200)
--verbose, -v Verbose logs.
--logs LOGS, -l LOGS Name of the log file
--logs-root LOGS_ROOT
Defalt root folder for writing logs.
--results-root RESULTS_ROOT
Where you want to save the results ?
--ckpt-root CKPT_ROOT
Directory where to save the model.
--data-root-dir DATA_ROOT_DIR
Root dir containing the dataset to train on.
--min-crop MIN_CROP Minimum size for cropping in data augmentation. range
(0-1)
--dataset {cifar10,cifar100,tiny}, -d {cifar10,cifar100,tiny}
Dataset to train on
--training-type {uni,inc}
How to feed the data to the network (incremental
(continual learning) or uniform (offline))
--n-classes N_CLASSES
Number of classes in database.
--img-size IMG_SIZE Size of the square input image
--num-workers NUM_WORKERS, -w NUM_WORKERS
Number of workers to use for dataloader.
--n-tasks N_TASKS How many tasks do you want ?
--labels-order LABELS_ORDER [LABELS_ORDER ...]
In which order to you want to see the labels ? Random
if not specified.
--nb-channels NB_CHANNELS
Number of channels for the input image.
--temperature TEMPERATURE
temperature parameter for softmax
--alpha ALPHA Weight for stream in SemiCon loss. Section 3.2 in paper.
--mem-size MEM_SIZE Memory size for continual learning
--mem-batch-size MEM_BATCH_SIZE, -mbs MEM_BATCH_SIZE
How many images do you want to retrieve from the
memory/ltm
--buffer {reservoir} Type of buffer.
--drop-method {random}
How to drop images from memory when adding new ones.
--learner {ER,CE,SCR,SSCL,SC}
What learner do you want ?
--eval-mem
--eval-random
--n-runs N_RUNS Number of runs, with different seeds each time.
--run-id RUN_ID Id of the current run in multi run.
--lab-pc LAB_PC Number of labeled images per class when evaluating
outside of memory.
--mem-iters MEM_ITERS
Number of times to make a grad update on memory at
each step
--review, -rv Review trick on last memory.
--rv-iters RV_ITERS, -rvi RV_ITERS
Number of iteration to do on last memory after
training.
Example of possible commands used to reproduce results are written in the commands.sh file.
While running experiment, the loss it plotted to tensorboard graphs. You can view them by running a tensorboard session with the corresponding logdir. This can be simply done using the following command.
tensorboard --logdir=tb_runs
A sample code can be found in graph.ipynb to read load aggregated results in a dataframe.
docker run -it --rm -u $(id -u):$(id -g) --name ucl -v $PWD:/home/micheln -v /data/dataset/torchvision:/data --gpus all --shm-size=8G ucl
tensorboard --logdir=runs --host 0.0.0.0
@inproceedings{michel2022contrastive,
title={Contrastive learning for online semi-supervised general continual learning},
author={Michel, Nicolas and Negrel, Romain and Chierchia, Giovanni and Bercher, Jean-Fmn{\c{c}}ois},
booktitle={2022 IEEE International Conference on Image Processing (ICIP)},
pages={1896--1900},
year={2022},
organization={IEEE}
}