CLHive is a codebase on top of PyTorch for Continual Learning research. It provides the components necessary to run CL experiments, for both task-incremental and class-incremental settings. It is designed to be readable and easily extensible, to allow users to quickly run and experiment with their own ideas.

• LwF, 2016
• iCaRL, 2016
• EWC, 2017
• AGEM/AGEM++, 2018
• ER, 2019
• MIR, 2019 (Coming soon)
• DER/DER++, 2020
• SS-IL, 2020 (Coming soon)
• ER-ACE/ER-AML, 2022

CLHive requires Python 3.6+.

• fvcore>=0.1.5
• hydra-core>=1.0.0
• numpy>=1.22.4
• pytorch>=1.12.0
• torchvision>=0.13.0
• wandb>=0.12.19

It is strongly recommend that you install CLHive in a dedicated virtualenv, to avoid conflicting with your system packages.

git clone https://github.com/naderAsadi/CLHive.git
cd CLHive
pip install -e .

With clhive you can use latest continual learning methods in a modular way using the full power of PyTorch. Experiment with different backbones, models and loss functions. The framework has been designed to be easy to use from the ground up.

from clhive.data import SplitCIFAR10
from clhive.scenarios import ClassIncremental
from clhive.models import ContinualModel
from clhive.methods import auto_method

train_dataset = SplitCIFAR10(root="path/to/data/", train=True)
train_scenario = ClassIncremental(dataset=dataset, n_tasks=5, batch_size=32)

print(
  f"Number of tasks: {train_scenario.n_tasks} | Number of classes: {train_scenario.n_classes}"
)

model = ContinualModel.auto_model("resnet18", train_scenario, image_size=32)
agent = auto_method(
    name="finetuning", model=model, optim=SGD(model.parameters(), lr=0.01)
)

for task_id, train_loader in enumerate(train_scenario):
    for x, y, t in train_loader:
        loss = agent.observe(x, y, t)
        ...

To create a replay buffer for rehearsal-based methods, e.g. ER, you can use clhive.ReplayBuffer class.

from clhive import ReplayBuffer

device = torch.device("cuda")
buffer = ReplayBuffer(capacity=20 * 10, device=device)

agent = auto_method(
    name="er",
    model=model,
    optim=SGD(model.parameters(), lr=0.01),
    buffer=buffer
)

Instead of iterating over all tasks manually, you can easily use clhive.Trainer to train the continual agent in any of the supported scenarios.

from clhive import Trainer

trainer = Trainer(method=agent, scenario=train_scenario, n_epochs=5, device=device)
trainer.fit()

Similar to the Trainer class, clhive.utils.evaluators package offers several evaluators, e.g. ContinualEvaluator and ProbeEvaluator.

from clhive.utils.evaluators import ContinualEvaluator, ProbeEvaluator

test_dataset = SplitCIFAR10(root="path/to/data/", train=False)
test_scenario = ClassIncremental(test_dataset, n_tasks=5, batch_size=32, n_workers=6)

evaluator = ContinualEvaluator(method=agent, scenario=test_scenario, device=device)
evaluator.fit()

Evaluators can also be passed to clhive.Trainer for automatic evaluation after each task.

trainer = Trainer(
    method=agent, scenario=scenario, n_epochs=5, evaluator=evaluator, device=device
)
trainer.fit()

CLHive is accessible also through a command-line interface (CLI). To train a ER model on Tiny-ImageNet you can simply run the following command:

python main.py ...

python main.py data=coco model/vision_model=vit-b  model/text_model=vit-b

Below you can see a schematic overview of the different concepts present in the clhive Python package.

Here is a reference to what each emoji in the commits means:

• 📎 : Some basic updates.
• ♻️ : Refactoring.
• 💩 : Bad code, needs to be revised!
• 🐛 : Bug fix.
• 💡 : New feature.
• ⚡ : Performance Improvement.