weightslab 1.2.4

Paving the way between black-box and white-box modeling.

WeightsLab — Inspect, Edit, and Evolve Neural Networks By Graybx.

About WeightsLab

WeightsLab is a powerful tool for editing and inspecting data & AI models.

What Problems Does It Solve?

WeightsLab addresses critical AI research challenges:

• Dataset insights & optimization
• Overfitting and training plateau
• Over/Under parameterization

Key Capabilities

The granular statistics and interactive paradigm enable powerful workflows:

• Monitor granular insights on data samples, signals, and weight parameters
• Use the AI agent to:
  • Create slices of data and discard them for the next training iteration
  • Discard low-quality samples from training data
  • Iterative pruning or growing of the architectures (INCOMING feature)

Quick Start

Requirements

• Docker Desktop v4.77 or newer — required to deploy the Weights Studio UI (weightslab ui launch).
• Docker Compose v2 (the docker compose CLI plugin, bundled with Docker Desktop) — recommended. The legacy v1 standalone binary (docker-compose, ≥ 1.27) also works: weightslab ui launch auto-detects whichever is installed and uses it. Compose v1 below 1.27 is not supported.
• Python >=3.10, <3.15 — to install and run the weightslab framework.

[!OPTIONAL] We are fully compatible with Conda.

[!TIP] Check your Compose version with docker compose version (v2) or docker-compose --version (v1). If docker compose version prints a version string, you have v2.

Installation

[!OPTIONAL] Seting a clean environment: Define a Python environment:

python -m venv weightslab_venv
./weightslab_venv/Scripts/activate

Or install directly on your machine.

Install our framework:

pip install weightslab

Deploy our interface:

weightslab ui launch

weightslab ui launch removes any stale weightslab/weights_studio Docker resources that could break the launch, then starts the UI stack. By default it runs unsecured (HTTP, no gRPC auth) — no certificates are generated.

To run secured (HTTPS + gRPC auth), pass --certs:

weightslab ui launch --certs   # generates TLS certs + a gRPC auth token if missing, then launches secured

[!IMPORTANT] When using certs, set WEIGHTSLAB_CERTS_DIR so the training backend and any new terminal use the same certificates (it is the single source of truth). weightslab se and weightslab ui launch --certs print the exact export/setx command for your shell. You can also generate certs up front with weightslab se.

[!IMPORTANT] For a detailed installation guide and more advanced features, please see the Installation Documentation.

Quick Training Example

Step-by-Step Integration

1. Add the import at the top of your script:

   import weightslab as wl  # ← Include our SDK into your experiment
   

2. Wrap your parameters with WeightsLab tracking:

   model = wl.watch_or_edit(parameters, ...)  # ← Now WeightsLab monitors your parameters and allow you to update them from your UI
   

3. Wrap your model with WeightsLab tracking:

   model = wl.watch_or_edit(SimpleModel(...), ...)  # ← Now WeightsLab monitors your model state
   

4. Wrap your optimizer with WeightsLab tracking:

   optimizer = wl.watch_or_edit(optim.Adam(...), ...)  # ← Tracks optimizer state and update optimizer learning rate from your UI
   

5. Wrap your signal with WeightsLab tracking:

    criterion = wl.watch_or_edit(nn.CrossEntropyLoss(reduction="none"), ...)  # ← Tracks this signal and others (metrics, ..etc) from your UI
   

6. Wrap your dataset with WeightsLab tracking:

    train_loader = wl.watch_or_edit(dataset, ...)  # ← Tracks this dataset and others (validation, test) from your UI
   

7. Run your training script as usual:

   python train.py
   

8. Launch the UI in another terminal:

   weightslab ui launch
   

9. Open your browser to https://localhost:5173 to track experiment evoluation and results!

Details

Here's a complete example showing how to integrate WeightsLab into a basic PyTorch training script:

#!/usr/bin/env python3
"""
Basic PyTorch training script with WeightsLab integration
"""
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
import weightslab as wl  # ← Import WeightsLab (uses TLS certs from WEIGHTSLAB_CERTS_DIR if present)


# Define a simple model
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__(input_shape=12, output_shape=2)
        self.linear = nn.Linear(input_shape, 1)

    def forward(self, x):
        return self.linear(x)


# Create synthetic data
def create_data(n_samples=1000):
    X = torch.randn(n_samples, 10)
    y = X.sum(dim=1, keepdim=True) + 0.1 * torch.randn(n_samples, 1)
    return TensorDataset(X, y)


# Main training function
def main():
    # Initialize WeightsLab - this creates certificates automatically!
    print("🚀 Initializing WeightsLab...")

    # Load hyperparameters (from YAML if present)
    parameters = {}
    config_path = os.path.join(os.path.dirname(__file__), "config.yaml")
    if os.path.exists(config_path):
        with open(config_path, "r") as fh:
            parameters = yaml.safe_load(fh) or {}
    parameters = wl.watch_or_edit(
        parameters,
        flag="hyperparameters",
        defaults=parameters,
        poll_interval=1.0,
    ) or {}  # Wrap the hyperparameters

    # Wrap your model and optimizer with WeightsLab
    model = wl.watch_or_edit(
      SimpleModel(
        input_shape=parameters.get('model', {}).get('input_shape', 10),
        output_shape=parameters.get('model', {}).get('output_shape', 1)
      )
    )  # ← WeightsLab tracks your model
    optimizer = wl.watch_or_edit(
      optim.Adam(model.parameters(), lr=parameters.get('model', {}).get('optimizer', {}).get('lr', 0.01)),
      flag='optimizer'
    )  # ← WeightsLab tracks optimizer

    # Create and wrap criterion
    criterion = wl.watch_or_edit(
        nn.CrossEntropyLoss(reduction="none"),
        flag="loss",
        signal_name="train-loss-CE",
        log=True  # If log is False, only save per sample value, not plot criterion
    )

    # Create data and dataloader
    dataset = create_data()
    train_loader = wl.watch_or_edit(
        dataset,
        flag="data",
        loader_name="loader",
        batch_size=parameters.get('data', {}).get('train_loader', {}).get('batch_size', 8),
        shuffle=parameters.get('data', {}).get('train_loader', {}).get('shuffle', False),
        is_training=True,  # Is it the training dataloader ?
        compute_hash=parameters.get('data', {}).get('train_loader', {}).get('compute_hash', True),  # Compute hash for train loader to allow dynamic augmentations and dataset sanity check
        preload_labels=parameters.get('data', {}).get('train_loader', {}).get('preload_labels', True),
        preload_metadata=parameters.get('data', {}).get('train_loader', {}).get('preload_metadata', True),
        enable_h5_persistence=parameters.get('data', {}).get('train_loader', {}).get('enable_h5_persistence', True),
        num_workers=parameters.get('data', {}).get('train_loader', {}).get('num_workers', 4)
    )

    # Training loop
    print("🏃 Starting training...")
    print("💡 Launch the UI with: weightslab ui launch")
    print("🌐 Open browser to: https://localhost:5173")
    n_epochs = parameters.get('n_epochs')
    pbar = tqdm.tqdm(range(n_epochs), desc='Training..') if parameters.get('tqdm_display', False) else range(n_epochs)
    for epoch in pbar:  # Train for 5 epochs
        total_loss = 0

        for batch_X, batch_y in dataloader:
            # Forward pass
            predictions = model(batch_X)
            loss = criterion(predictions, batch_y)

            # Backward pass
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            total_loss += loss.item()

        avg_loss = total_loss / len(dataloader)
        print(f"Epoch {epoch+1}/5 - Loss: {avg_loss:.4f}")

    print("✅ Training complete!")


if __name__ == "__main__":
    main()

Migrating from wandb? See the diff:

--- train_baseline.py
+++ train_wl.py
@@ -1,11 +1,12 @@
 import argparse
 import torch
 import torch.nn as nn
-from torch.utils.data import DataLoader
 from torchvision import datasets, transforms, models
 from torchmetrics.classification import MulticlassAccuracy

-import wandb
+import weightslab as wl
+from weightslab.components.global_monitoring import (
+    guard_training_context, guard_testing_context)
+
[e-mail](name="byte_adjusted_loss", subscribe_to="loss/CE")
+def byte_adjusted_loss(ctx): return ctx.subscribed_value / ctx.image_bytes  # chains on image_bytes
+
 def main():
@@ -15,29 +16,38 @@

     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     parameters = {"batch_size": 128, "lr": 1e-3}

-    wandb.init(project="cifar10")
-
     transform = transforms.Compose([
         transforms.ToTensor(),
         transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
     ])
     train_set = datasets.CIFAR10("./data", train=True,  download=True, transform=transform)
     test_set  = datasets.CIFAR10("./data", train=False, download=True, transform=transform)
-    train_loader = DataLoader(train_set, batch_size=parameters["batch_size"], shuffle=True, num_workers=2)
-    test_loader  = DataLoader(test_set,  batch_size=256,                              num_workers=2)
+    wl.watch_or_edit(parameters, flag="hyperparameters")  # live-editable in UI
+
+    train_loader = wl.watch_or_edit(
+        train_set, flag="data", loader_name="train_loader",
+        batch_size=parameters["batch_size"], shuffle=True, is_training=True)
+    test_loader  = wl.watch_or_edit(
+        test_set,  flag="data", loader_name="test_loader",
+        batch_size=256, shuffle=False, is_training=False)

     model = models.resnet18(weights=None)
     model.fc = nn.Linear(model.fc.in_features, 10)
     model = model.to(device)
     optimizer = torch.optim.Adam(model.parameters(), lr=parameters["lr"])

-    criterion = nn.CrossEntropyLoss()
-    accuracy  = MulticlassAccuracy(num_classes=10).to(device)
+    criterion = wl.watch_or_edit(
+        nn.CrossEntropyLoss(), flag="loss", signal_name="loss/CE")
+    accuracy  = wl.watch_or_edit(
+        MulticlassAccuracy(num_classes=10).to(device),
+        flag="metric", signal_name="acc")
+
+    wl.serve(serving_grpc=True)

     for epoch in range(1, args.epochs + 1):
         model.train()
         accuracy.reset()
         for x, y in train_loader:
+            with guard_training_context:
                 x, y = x.to(device), y.to(device)
                 logits = model(x)
                 loss = criterion(logits, y)
                 optimizer.zero_grad()
                 loss.backward()
                 optimizer.step()
                 accuracy.update(logits, y)
-            wandb.log({"train/loss": loss.item()})
-        wandb.log({"train/acc": accuracy.compute().item(), "epoch": epoch})
+            wl.save_signals(preds_raw=logits, targets=y,
+                            signals={"metric/accuracy": accuracy.compute().item()})

         model.eval()
         accuracy.reset()
         with torch.no_grad():
             for x, y in test_loader:
+                with guard_testing_context:
                     x, y = x.to(device), y.to(device)
                     accuracy.update(model(x), y)
-        wandb.log({"test/acc": accuracy.compute().item(), "epoch": epoch})
+                wl.save_signals(preds_raw=logits, targets=y,
+                                signals={"metric/accuracy": accuracy.compute().item()})

-    wandb.finish()
+    wl.keep_serving()

What WeightsLab Does Automatically

• 📊 Experiment tracking for reproducibility
• 📈 Provides live metrics and visualization in the web UI
• 🔄 Enables data supervision during training and experiment hyperparameter tuning through the UI

Examples

Local examples

After starting the UI, launch a local experiment with the command:

weightslab start example          # classification (default)
# weightslab start example --cls    # classification
# weightslab start example --seg    # segmentation
# weightslab start example --det    # detection
# weightslab start example --clus   # clustering
# weightslab start example --gen    # generation

Cloud examples

Find our demos online. The password is graybx.

DEMOS

AI Agent

WeightsLab can run its data agent in two modes:

• Local provider with Ollama
• Cloud provider with OpenRouter

Use local Ollama when you want a fully local setup and do not need cloud-hosted models (see more details installation from the documentation). Use OpenRouter when you want larger hosted models and model selection directly from Weightslab UI.

Cloud OpenRouter

You can either preconfigure OpenRouter in agent_config.yaml / .env, or initialize it interactively from Weightslab UI.

SDK Configuration

Example static configuration:

agent:
  provider: openrouter
  openrouter_model: meta-llama/llama-3.3-70b-instruct
  fallback_to_local: false
  # openrouter_api_key: ${OPENROUTER_API_KEY}

Environment variable:

export OPENROUTER_API_KEY=your_key_here

Interactive setup from Weightslab UI

OpenRouter models can be initialized and set directly from the UI:

1. Click in the agent bar or double-click to expand the agent window.
2. Type /init.
3. Choose either:
   • A Enter your OpenRouter API key manually
   • B Use the OpenRouter OAuth flow
4. Select a model from the fetched list, then confirm.

The default OpenRouter model, as recommended by Graybx, is meta-llama/llama-3.3-70b-instruct.

Typical Usage Flow

1. Start your WeightsLab backend (e.g., "main.py").
2. Start Weightslab UI.
3. If you use Ollama, query the agent directly.
4. If you use OpenRouter and the agent is not configured yet, type /init.
5. Ask natural-language data operations such as sorting, filtering, slicing, and inspection requests. You can also ask questions about the data.
6. Use /model to try another cloud model without re-entering the key.
7. Use /reset if you want to clear the current connection and start over.

Documentation (API + SDK)

Find our documentation online.

Contributing & onboarding

New here (human or AI coding agent)? Start with AGENTS.md — it captures the cross-repo architecture (weightslab backend ↔ weights_studio frontend via the shared proto), the module maps, the wl.watch_or_edit integration pattern, where tests live, and the gotchas that aren't obvious from any single file. It's the fastest way to orient before a first change.

Community

Graybx is building a wonderful community of AI researchers and engineers. Are you interested in joining our project? Contact us at hello [at] graybx [dot] com