Trace training signals back to the exact samples causing them

Pause training, mine live loss signals to surface mislabels, class imbalance & outliers,
then curate your image, video & LiDAR data, without restarting.

What it does

WeightsLab is an open-source PyTorch tool for dataset debugging, data quality monitoring, mislabel detection, and mid-training data curation for computer vision datasets: images, video & LiDAR point clouds. Longer-term, we're building toward bringing dataset management, training, fine-tuning, and validation together in a single, unified workflow.

Most data problems are invisible until your model tells you: through loss spikes, poor generalization, or silent underperformance. WeightsLab connects those training signals back to the exact samples causing them.

Wrap your training script with the SDK to capture per-sample signals live. Open Studio to inspect, filter, and curate your dataset mid-training, without restarting.

• Detect - Surface mislabels, outliers & class imbalance using live loss signals
• Curate - Discard bad samples, create data subsets, rebalance distributions
• Continue - Resume training on your cleaned dataset, no restart required

Quickstart

1. Install

pip install weightslab

2. Wrap your training script

# wrap the objects in your training script

import weightslab as wl
...
model  = wl.watch_or_edit(model, flag='model')
optim  = wl.watch_or_edit(optim, flag='opt')
loss   = wl.watch_or_edit(loss, flag='signal', name="loss", per_sample=True, log=True)
loader = wl.watch_or_edit(dataset, flag='data', loader_name="train")
...
wl.serve(serving_grpc=True, serving_cli=False)
...

3. Launch Studio

weightslab ui launch  # then open https://localhost:5173 🚀

For a detailed installation guide and advanced configuration → Installation Documentation.

Tip

Quick examples to get started

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

Resources & Community

Training script with Weightslab - Step-by-Step Integration

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

   import weightslab as wl

2. Wrap your parameters, model, optimizer, signals, and dataset:

   parameters      = wl.watch_or_edit(parameters, flag='hp',     ...) # ← WeightsLab monitors your parameters and lets you update them from the UI
   model           = wl.watch_or_edit(model, flag='model', ...) # ← WeightsLab monitors your model state
   optimizer       = wl.watch_or_edit(optim.Adam(...),                         flag='opt',    ...) # ← Tracks optimizer state and lets you update the learning rate from the UI

   train_criterion = wl.watch_or_edit(nn.CrossEntropyLoss(reduction="none"),  flag='signal', name="train_loss/sample", per_sample=True, log=True)   # ← Wrap and plot your signals on the UI
   test_criterion  = wl.watch_or_edit(nn.CrossEntropyLoss(reduction="none"),  flag='signal', name="test_loss/sample",  per_sample=True, log=False)  # ← Per-sample only, plot disabled

   train_loader    = wl.watch_or_edit(train_dataset, flag='data', loader_name="train_loader", ...)  # ← Track your training dataset
   val_loader      = wl.watch_or_edit(val_dataset,   flag='data', loader_name="val_loader",   ...)  # ← Track your validation dataset

3. Run your script, then launch the UI in a separate terminal:

   python train.py
   weightslab ui launch

4. Open your browser https://localhost:5173 and inspect your training in real time.

Training script with Weightslab - Full Example

#!/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


class SimpleModel(nn.Module):
    def __init__(self, input_shape=10, output_shape=1):
        super().__init__()
        self.linear = nn.Linear(input_shape, output_shape)

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


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)


def main():
    parameters = wl.watch_or_edit({}, flag="hyperparameters", poll_interval=1.0) or {}

    model     = wl.watch_or_edit(SimpleModel(), flag='model')
    optimizer = wl.watch_or_edit(optim.Adam(model.parameters(), lr=0.01), flag='optimizer')
    criterion = wl.watch_or_edit(nn.CrossEntropyLoss(reduction="none"), flag="loss", signal_name="train-loss-CE", log=True)
    loader    = wl.watch_or_edit(create_data(), flag="data", loader_name="loader", batch_size=8, is_training=True)

    for epoch in range(parameters.get('n_epochs', 5)):
        total_loss = 0
        for batch_X, batch_y in loader:
            predictions = model(batch_X)
            loss = criterion(predictions, batch_y)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            total_loss += loss.item()

            # Write the history of these samples every x steps
            if model.get_age() % 100 == 0:
                print(f'Dump signals history and dataframe at age {model.get_age()}')
                wl.write_history(
                    # path=None,  # Use root_log_dir by default, filename generated from parameters md5 hash
                    type_of_history="all",
                    graph_name=[
                        'train/clsf_instance',
                        'val/clsf_instance'
                    ],
                    # experiment_hash=None,  Default is 'last', i.e., current experiment hash
                    sample_id=['11', '29', '28', '27', '22'],
                    instance_id=[1, 2, 3]
                )

                # Dump the sample dataframe: all signals plus the loss_shape categorical tag,
                wl.write_dataframe(
                    columns=["signals", "tag:loss_shape"],
                    format='csv'
                    # sample_id=['0', '28']
                    # instance_id=[1, 2],
                )

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

    print("✅ Training complete!")


if __name__ == "__main__":
    main()

Migrating from Weights & Biases?

WeightsLab vs Weights & Biases

Weights & Biases (wandb) tracks experiments. WeightsLab connects training signals back to the exact samples causing them — so you can fix your data, not just log it.

--- 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)
+
+@wl.signal(name="byte_adjusted_loss", subscribe_to="loss/CE")
+def byte_adjusted_loss(ctx): return ctx.subscribed_value / ctx.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([...])
     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)
     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()
         for x, y in train_loader:
+            with guard_training_context:
                 logits = model(x.to(device))
                 loss   = criterion(logits, y.to(device))
                 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()
         with torch.no_grad():
             for x, y in test_loader:
+                with guard_testing_context:
                     accuracy.update(model(x.to(device)), 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()

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

We're building a community of ML engineers around data-centric training tooling. Interested in contributing or just want to say hi? → hello [at] graybx [dot] com