clearn

Wrap once. Train forever.

Continual learning for PyTorch models.
Prevent catastrophic forgetting with one line of code.

---

When you fine-tune a neural network on new data, it catastrophically forgets what it learned before. clearn fixes this. Wrap any PyTorch model, train on sequential tasks, and your model remembers everything.

import clearn

model = clearn.wrap(your_model, strategy="ewc")

model.fit(task1_loader, optimizer, task_id="q1_fraud")
model.fit(task2_loader, optimizer, task_id="q2_fraud")

print(model.diff())

RetentionReport
├── q1_fraud: 94.2% retained  (-5.8%)
├── q2_fraud: 100.0% (current task)
├── plasticity_score: 0.87
├── stability_score: 0.94
└── recommendation: "stable — no action needed"

---

Installation

pip install clearn-ai

For HuggingFace integration:

pip install clearn-ai[hf]

---

Quickstart

import torch
import torch.nn as nn
from torch.utils.data import DataLoader, TensorDataset
import clearn

# 1. Your PyTorch model
model = nn.Sequential(nn.Linear(784, 128), nn.ReLU(), nn.Linear(128, 10))

# 2. Wrap it — one line
cl_model = clearn.wrap(model, strategy="ewc")

# 3. Train on sequential tasks
for i, task_data in enumerate(sequential_tasks):
    loader = DataLoader(task_data, batch_size=64)
    optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
    metrics = cl_model.fit(loader, optimizer, task_id=f"task_{i}")
    print(f"Task {i}: loss={metrics.final_loss:.4f}, acc={metrics.final_accuracy:.2%}")

# 4. See what was retained
print(cl_model.diff())

That's it. Four steps. Your model now remembers.

---

Why clearn?

Problem	Without clearn	With clearn
Train on Task 2	Task 1 accuracy: 8%	Task 1 accuracy: 94%
Train on 20 tasks	First task: destroyed	First task: preserved
Debug forgetting	Print loss, guess	model.diff() tells you exactly

---

Strategies

clearn ships five strategies:

EWC (Elastic Weight Consolidation)

Regularization-based. Identifies which weights matter most via the Fisher Information Matrix, then protects them during future training. No need to store past data.

model = clearn.wrap(net, strategy="ewc", lambda_=5000)

Parameter	Default	Description
lambda_	5000	Regularization strength. Higher = less forgetting, less plasticity
n_fisher_samples	200	Samples used to estimate weight importance

SI (Synaptic Intelligence)

Online importance estimation. Tracks per-parameter contribution to loss reduction during training, then penalizes changes to important weights. No separate computation pass needed — importance is accumulated during training.

model = clearn.wrap(net, strategy="si", c=1.0)

Parameter	Default	Description
c	1.0	Regularization strength (analogous to EWC's lambda)
epsilon	1e-3	Numerical stability constant

DER++ (Dark Experience Replay)

Replay-based. Stores a small buffer of past examples and replays them during training, matching original logits via KL divergence with temperature scaling. Best general-purpose performance.

model = clearn.wrap(net, strategy="der", buffer_size=500)

Parameter	Default	Description
buffer_size	200	Number of past samples to store
alpha	0.1	Weight for cross-entropy replay loss
beta	0.5	Weight for KL divergence logit-matching loss
temperature	2.0	Temperature for KL divergence softmax
buffer_device	"cpu"	Device to store buffer on ("cuda" avoids transfers)

GEM (Gradient Episodic Memory)

Constraint-based. Stores episodic memories from past tasks and projects gradients to avoid increasing loss on any previous task. Uses the efficient A-GEM variant.

model = clearn.wrap(net, strategy="gem", memory_size=256)

Parameter	Default	Description
memory_size	256	Samples to store per task

LoRA-EWC (Parameter-Efficient Continual Learning)

Combines LoRA adapters (via peft) with EWC regularization. Only the low-rank adapter weights are trained and protected — the base model stays frozen. Ideal for LLMs.

# Requires: pip install clearn-ai[hf]
model = clearn.from_pretrained("bert-base-uncased", strategy="lora-ewc", lora_r=8)

Parameter	Default	Description
lora_r	8	LoRA rank (lower = more efficient)
lora_alpha	16	LoRA alpha scaling
lambda_	5000	EWC regularization on LoRA weights

Which strategy should I use?

Using a large language model?
├── Yes  → LoRA-EWC (parameter-efficient + forgetting protection)
└── No   → Can you store past data?
           ├── Yes  → DER++ (best retention)
           └── No   → Do you need online tracking?
                      ├── Yes  → SI (no Fisher pass needed)
                      └── No   → Want hard constraints?
                                 ├── Yes  → GEM (gradient projection)
                                 └── No   → EWC (classic, reliable)

---

The diff() Report

The key feature. Like git diff, but for model knowledge.

report = model.diff()
print(report)

RetentionReport
├── task_a: 94.2% retained  (-5.8%)
├── task_b: 88.1% retained  (-11.9%)
├── task_c: 100.0% (current task)
├── plasticity_score: 0.91
├── stability_score: 0.91
└── recommendation: "stable — no action needed"

The report gives you:

• Per-task retention — exactly how much each task was preserved
• Plasticity score — how well the latest task was learned
• Stability score — average retention across all past tasks
• Recommendation — actionable advice ("increase lambda", "try DER++", etc.)

---

Training Metrics

Every fit() call returns detailed metrics:

metrics = model.fit(loader, optimizer, task_id="q1", epochs=5)
print(metrics)

TrainingMetrics(task='q1')
├── epochs: 5
├── final_loss: 0.3421
├── final_accuracy: 91.20%
└── wall_time: 2.15s

Access per-epoch data: metrics.epoch_losses, metrics.epoch_accuracies.

---

Strategy Diagnostics

Inspect the internals of your strategy at any time:

diag = model.diagnostics()
# EWC example:
# {'strategy': 'ewc', 'lambda': 5000, 'consolidated': True,
#  'fisher_mean': 0.0023, 'fisher_max': 10000.0, 'current_penalty': 42.5, ...}

# DER++ example:
# {'strategy': 'der++', 'buffer_used': 200, 'buffer_utilization': 1.0,
#  'buffer_class_distribution': {0: 45, 1: 38, ...}, ...}

---

Callbacks

Hook into training with the callback system:

from clearn import ContinualCallback

class LogCallback(ContinualCallback):
    def on_task_start(self, model, task_id):
        print(f"Starting {task_id}")

    def on_batch_end(self, model, loss):
        pass  # Log to wandb, etc.

    def on_task_end(self, model, task_id, metrics):
        print(f"Finished {task_id}: {metrics.final_accuracy:.2%}")

model.fit(loader, optimizer, callbacks=[LogCallback()])

Built-in: EarlyStoppingCallback(patience=50).

---

Gradient Clipping & Mixed Precision

# Gradient clipping
model.fit(loader, optimizer, grad_clip=1.0)

# Mixed precision (AMP) — requires CUDA
model.fit(loader, optimizer, use_amp=True)

# Both
model.fit(loader, optimizer, grad_clip=1.0, use_amp=True)

---

Save & Load

# Save full state (model + strategy + task history)
model.save("./checkpoints/my_model")

# Load it back — diff() works after load
model = clearn.load("./checkpoints/my_model", model=your_model)
print(model.diff())  # Retention report preserved

---

HuggingFace Integration

First-class support for HuggingFace Transformers.

# Load any HuggingFace model with continual learning
model = clearn.from_pretrained("bert-base-uncased", strategy="ewc", task="classification")
model = clearn.from_pretrained("gpt2", strategy="lora-ewc", task="causal-lm")

# Get the tokenizer too
model, tokenizer = clearn.from_pretrained(
    "bert-base-uncased", strategy="ewc", return_tokenizer=True
)

# Supported tasks: classification, token-classification, causal-lm, seq2seq-lm

ContinualTrainer — drop-in replacement for HuggingFace Trainer:

from clearn.integrations.huggingface import ContinualTrainer

trainer = ContinualTrainer(
    model=cl_model,
    args=training_args,
    train_dataset=dataset,
    task_id="sentiment_v1",
)
trainer.train()  # Automatically applies forgetting protection

Push to HuggingFace Hub:

model.push_to_hub("your-username/my-continual-model")

---

API Reference

import clearn

# Wrap any PyTorch model
model = clearn.wrap(model, strategy="ewc", **kwargs)

# Train on a task (returns TrainingMetrics)
metrics = model.fit(dataloader, optimizer, epochs=1, task_id=None,
                    loss_fn=None, grad_clip=None, callbacks=None, use_amp=False)

# Get retention report
report = model.diff()

# Get strategy diagnostics
diag = model.diagnostics()

# Save / Load (diff() works after load)
model.save("path/to/checkpoint")
model = clearn.load("path/to/checkpoint", model=your_model)

# HuggingFace (requires clearn-ai[hf])
model = clearn.from_pretrained("bert-base-uncased", strategy="ewc", task="classification")
model, tokenizer = clearn.from_pretrained("gpt2", strategy="lora-ewc",
                                           task="causal-lm", return_tokenizer=True)
model.push_to_hub("user/model-name")

---

Benchmark: CIFAR-100 Sequential

Split CIFAR-100 into 20 tasks. Train a ResNet-18 on each. Track Task 1 accuracy.

Method	Task 1 Accuracy (after 20 tasks)
Baseline (SGD)	~8%
clearn EWC	~82%
clearn DER++	~88%

Run the benchmark yourself:

---

Project Structure

clearn/
├── clearn/
│   ├── core.py              # ContinualModel — the main wrapper
│   ├── strategies/
│   │   ├── base.py           # Abstract strategy interface
│   │   ├── ewc.py            # Elastic Weight Consolidation
│   │   ├── si.py             # Synaptic Intelligence
│   │   ├── der.py            # Dark Experience Replay++
│   │   ├── gem.py            # Gradient Episodic Memory (A-GEM)
│   │   └── lora_ewc.py       # LoRA + EWC hybrid
│   ├── metrics.py            # RetentionReport, TrainingMetrics, diff() logic
│   ├── callbacks.py          # ContinualCallback, EarlyStoppingCallback
│   └── integrations/
│       └── huggingface.py    # from_pretrained(), ContinualTrainer, push_to_hub
├── tests/                    # 114 tests, all passing
├── examples/                 # Runnable demo scripts
└── benchmarks/               # CIFAR-100 notebook

---

Contributing

git clone https://github.com/itisrmk/clearn.git
cd clearn
pip install -e ".[dev]"
pytest tests/ -v

---

License

MIT

---

Built by Rahul Kashyap
_{Continual learning infrastructure for production ML}