stable-pretraining

PyTorch Lightning for foundation-model research. Dict-shaped state so any intermediate tensor is loggable, live evaluation probes that attach without touching the training loop (OnlineProbe, OnlineKNN, RankMe, LiDAR, …), SLURM-grade requeue + atomic checkpoints + queryable run registry, GPU-side batched augmentation, and 30+ ready recipes spanning SSL, supervised, and multi-modal pretraining (SimCLR, DINO/DINOv2, MAE, BYOL, VICReg, Barlow Twins, LeJEPA, CLIP, …). An experimental JAX / Flax-NNX backend mirrors the same design (forward-dict, callbacks, SLURM-grade Manager) for users who prefer JAX.

30-second tour ↓ · JAX backend ↓ · Built-in methods ↓ · Discord

Table of Contents

• How?
• Quick Setup
• Tutorial Notebook
• 30-second tour
• JAX backend (experimental)
• Core Structure
  • Data
  • Module
  • Callbacks
  • Trainer
• Global Configuration
• Output Directory (cache_dir)
• Run Registry
• Built-in Methods
• Backbones
• Optimizers & Schedulers
• Complete Example
• Quick Start with spt CLI
• Installation
• Contributing
• Citation
• Contributors
• Benchmarks

How?

To reach flexibility, scalability and stability, we rely on battle-tested third party libraries: PyTorch, Lightning, HuggingFace, TorchMetrics amongst a few others. Those dependencies allow us to focus on assembling everything into a powerful ML framework. stable-pretraining adopts a flexible and modular design for seamless integration of components from external libraries, including architectures, loss functions, evaluation metrics, and augmentations.

Quick setup

# Clone the repository
git clone https://github.com/galilai-group/stable-pretraining.git

# Install the framework
cd stable-pretraining
pip install -e .

For an interactive walkthrough — data loading, Module, callbacks, training, and evaluation all in one place — open the tutorial notebook:

jupyter notebook examples/simclr_cifar10_tutorial.ipynb

30-second tour

The whole framework is four components that pass dicts to each other. Once you see the shape, the rest of the README is reference.

%%{init: {'theme':'base', 'themeVariables': {
    'fontFamily':'-apple-system, BlinkMacSystemFont, system-ui, sans-serif',
    'fontSize':'14px',
    'primaryColor':'#fafafa',
    'primaryTextColor':'#1f2937',
    'primaryBorderColor':'#9ca3af',
    'lineColor':'#9ca3af',
    'edgeLabelBackground':'#ffffff'
}}}%%
flowchart LR
    classDef stack fill:#f9fafb,stroke:#9ca3af,color:#111827,stroke-width:1px
    classDef yours fill:#eef2ff,stroke:#6366f1,color:#3730a3,stroke-width:1px
    classDef hook  fill:#f0fdf4,stroke:#22c55e,color:#14532d,stroke-width:1px
    classDef orch  fill:#1f2937,stroke:#1f2937,color:#f9fafb,stroke-width:0px

    DM([DataModule])
    TR([Lightning Trainer])
    MOD([Module])
    CB([Callbacks])
    LOG([Loggers · Registry])
    MGR((Manager))

    DM -->|batch| TR --> MOD
    MOD -->|state| CB
    MOD --> LOG
    CB --> LOG
    MGR -.- TR

    class DM yours
    class MOD yours
    class CB hook
    class TR stack
    class LOG stack
    class MGR orch

Indigo nodes are what you write (data + forward). Green is the hook surface (callbacks). Slate is the orchestration that runs underneath.

A minimal end-to-end run looks like this (substitute your own loader / loss):

import lightning as pl
import stable_pretraining as spt

# 1. Data flows as dicts. CPU transforms decode + resize; random
#    augmentation can live on the GPU via `gpu_transform=` (optional).
train_ds = spt.data.HFDataset("cifar10", split="train", transform=...)
val_ds   = spt.data.HFDataset("cifar10", split="test",  transform=...)
dm = spt.data.DataModule(
    train=torch.utils.data.DataLoader(train_ds, batch_size=256, num_workers=8),
    val=torch.utils.data.DataLoader(val_ds, batch_size=256, num_workers=8),
)

# 2. Module = backbone + a forward function that returns a state dict.
#    `forward.simclr` (and friends) are pre-built; you can also
#    write your own — anything returning {"loss": ..., "embedding": ...}.
module = spt.Module(
    backbone=spt.backbone.from_timm("resnet18", num_classes=0),
    projector=spt.backbone.MLP(512, 512, 128),
    forward=spt.forward.simclr,
    simclr_loss=spt.losses.NTXEntLoss(temperature=0.5),
)

# 3. Callbacks watch the state dict and train online probes / log metrics
#    without touching the main loop.
trainer = pl.Trainer(max_epochs=100, precision="bf16-mixed", callbacks=[
    spt.callbacks.OnlineProbe(module, name="probe", input="embedding", target="label",
                              probe=nn.Linear(512, 10), loss=nn.CrossEntropyLoss()),
    spt.callbacks.OnlineKNN(name="knn", input="embedding", target="label",
                            queue_length=10_000, input_dim=512, k=20),
])

# 4. Manager wraps fit() with SLURM-requeue, atomic checkpoints, and
#    the run registry. On a workstation it's just a thin call; on a
#    cluster it adds preempt/resume + run tracking for free.
spt.Manager(trainer=trainer, module=module, data=dm)()

A single sample's journey through the stack:

%%{init: {'theme':'base', 'themeVariables': {
    'fontFamily':'-apple-system, BlinkMacSystemFont, system-ui, sans-serif',
    'fontSize':'13px',
    'primaryColor':'#fafafa',
    'primaryTextColor':'#1f2937',
    'primaryBorderColor':'#9ca3af',
    'lineColor':'#9ca3af',
    'edgeLabelBackground':'#ffffff'
}}}%%
flowchart LR
    classDef cpu fill:#f9fafb,stroke:#9ca3af,color:#111827,stroke-width:1px
    classDef gpu fill:#eef2ff,stroke:#6366f1,color:#3730a3,stroke-width:1px
    classDef out fill:#f0fdf4,stroke:#22c55e,color:#14532d,stroke-width:1px

    SAMP([sample])
    CPU([CPU prep])
    H2D([H2D])
    AUG([GPU aug])
    FWD([forward])
    BWD([backward])
    LOG([loggers])

    SAMP --> CPU --> H2D --> AUG --> FWD --> BWD --> LOG
    FWD -. callbacks .-> LOG

    class SAMP cpu
    class CPU cpu
    class H2D cpu
    class AUG gpu
    class FWD gpu
    class BWD gpu
    class LOG out

Read left-to-right. Slate runs on CPU (decode, resize, pinned H2D). Indigo runs on GPU — including the augmentation that traditionally lived on CPU. The dotted arrow shows callbacks tapping the same state dict the loss is computed from, without touching the loop.

The dict-everywhere design means callbacks attach without modifying the training loop, GPU augmentation slots in via dataset.gpu_transform= (see §GPU-side batched augmentation), and any intermediate quantity in the state dict is automatically available to loggers.

JAX backend (experimental)

A parallel JAX / Flax-NNX backend lives under stable_pretraining.jax (install with pip install -e ".[jax]"). It is opt-in and isolated — import stable_pretraining never imports JAX — and it mirrors the torch design: a stateless forward(self, batch, stage) that returns a state dict, callbacks that read that dict under the same Lightning-style hooks, and a Manager with the same cache_dir run-dir layout

• SLURM preempt/requeue + atomic checkpoints. The engine underneath is JAX-native (nnx.value_and_grad + optax, nnx.jit), and metrics are plain jnp/NumPy — no torchmetrics.

The same minimal SimCLR run, side by side:

# ---- PyTorch / Lightning -------------------     # ---- JAX / Flax-NNX ----------------------------
import stable_pretraining as spt                    import stable_pretraining.jax as spj
import torch, lightning as pl                       from flax import nnx
                                                     rngs = nnx.Rngs(0)
module = spt.Module(                                 module = spj.Module(
    forward=spt.forward.simclr,                          forward=spj.forward.simclr,
    backbone=spt.backbone.from_torchvision(              backbone=spj.backbone.resnet18(
        "resnet18", low_resolution=True),                    rngs=rngs, low_resolution=True),
    projector=torch.nn.Linear(512, 128),                 projector=spj.backbone.MLP(512, [128], rngs=rngs),
    simclr_loss=spt.losses.NTXEntLoss(0.5),              simclr_loss=spj.losses.NTXEntLoss(0.5),
    optim={"optimizer": {"type": "LARS",                 optim={"type": "lars",
                         "lr": 5}},                                 "learning_rate": 5.0},
)                                                    )
probe = spt.OnlineProbe(module, name="probe",        probe = spj.OnlineProbe("probe",
    input="embedding", target="label",                   probe=nnx.Linear(512, 10, rngs=rngs))
    probe=torch.nn.Linear(512, 10),
    loss=torch.nn.CrossEntropyLoss())
trainer = pl.Trainer(max_epochs=100,                 trainer = spj.Trainer(max_epochs=100,
                     callbacks=[probe])                                    callbacks=[probe])
spt.Manager(trainer=trainer, module=module,          spj.Manager(trainer, module,
            data=dm)()                                           train_loader, val_loader)()

Multi-GPU is one flag — spj.Trainer(..., data_parallel=True) shards each batch across all visible devices (SPMD); bf16 mixed precision is dtype=jnp.bfloat16 on the backbone.

What's available today: methods SimCLR / VICReg / Barlow Twins / SimSiam / BYOL; backbones MLP / ResNet-{9,18,34,50,101,152} / ConvMixer / ViT-{tiny,small, base,large}; losses NT-Xent / InfoNCE / BYOL / VICReg / Barlow Twins / SwAV-Sinkhorn / neg-cosine; callbacks OnlineProbe, OnlineKNN, RankMe, LiDAR, OnlineQueue, OnlineWriter, EarlyStopping, EMA TeacherStudentCallback; checkpointing + exact resume; on-device augmentation (spj.augment). Every torch↔JAX numerical claim above is covered by parity regression tests.

Logging is identical to the torch path — not reimplemented. The JAX Trainer drives the same logger classes (RegistryLogger, WandB, Trackio, SwanLab) via logger=, Module.log(name, value) matches, and metric keys are the same (fit/loss, train/<p>_loss, eval/<p>_acc, bare rankme/lidar, …). spj.Manager auto-attaches the RegistryLogger, so JAX runs write the same sidecar.json and show up in spt registry ls/best/... exactly like torch runs. Metrics use plain jnp/NumPy — no torchmetrics.

Still pending (vs the torch path): DINO/DINOv2/MAE/I-JEPA methods, the full ViT feature set (SwiGLU/QK-norm/RoPE), video backbones, and a few niche callbacks. Contributions welcome.

Core Structure

stable-pretraining simplifies complex ML workflows into 4 intuitive components:

1 - Data

Your dataset must follow a dictionary-structured format where each sample is a dictionary with named fields (e.g., {"image": ..., "label": ...}). This ensures consistent behavior across all components. You have multiple options for creating datasets:

• HuggingFace datasets (if available on the Hub):

import stable_pretraining as spt
train_dataset = spt.data.HFDataset(
    path="frgfm/imagenette",
    name="160px",
    split="train",
    transform=train_transform,
)

• From PyTorch datasets:

train_dataset = spt.data.FromTorchDataset(
    torchvision_dataset,
    names=["image", "label"],  # Map tuple outputs to dictionary keys
    transform=train_transform,
)

• Custom datasets: Any dataset that returns dictionaries

datamodule = spt.data.DataModule(train=train_dataloader, val=val_dataloader)

GPU-side batched augmentation (gpu_transform)

SSL augmentation traditionally runs per-sample in DataLoader workers (PIL + torchvision). On modern accelerators that becomes the bottleneck: CPU workers can't keep up with the GPU on heavy multi-view recipes. Moving augmentation to the GPU via kornia inside Lightning's on_after_batch_transfer hook vectorises it across the batch and frees CPU workers to do more I/O in parallel.

We measure substantial end-to-end throughput improvements across model sizes and precisions — see benchmarks/imagenet10/RESULTS.md for the full sweep (ViT-S/16, ViT-L/16; fp16 / bf16 / FP8; torch.compile on/off; various batch sizes).

Before — torchvision augmentation in CPU workers

The historical SSL recipe: every random transform runs per-sample inside the DataLoader workers. MultiViewTransform produces both views on the CPU side.

from stable_pretraining.data import transforms
import stable_pretraining as spt

train_transform = transforms.MultiViewTransform([
    transforms.Compose(                        # view 1
        transforms.RGB(),
        transforms.RandomResizedCrop((224, 224), scale=(0.08, 1.0)),
        transforms.ColorJitter(0.4, 0.4, 0.2, 0.1, p=0.8),
        transforms.RandomGrayscale(p=0.2),
        transforms.PILGaussianBlur(p=1.0),
        transforms.RandomHorizontalFlip(p=0.5),
        transforms.ToImage(**spt.data.static.ImageNet),
    ),
    transforms.Compose(                        # view 2 — same recipe
        transforms.RGB(),
        transforms.RandomResizedCrop((224, 224), scale=(0.08, 1.0)),
        transforms.ColorJitter(0.4, 0.4, 0.2, 0.1, p=0.8),
        transforms.RandomGrayscale(p=0.2),
        transforms.PILGaussianBlur(p=1.0),
        transforms.RandomHorizontalFlip(p=0.5),
        transforms.ToImage(**spt.data.static.ImageNet),
    ),
])

train_ds = spt.data.HFDataset(
    "frgfm/imagenette", split="train",
    transform=train_transform,
)

After — same augs, batched on the GPU

The CPU transform shrinks to "decode + resize". The augmentation chain runs on the GPU, and StackedMultiView produces N views from one source tensor in a single chain call (symmetric SSL — Barlow Twins, SimCLR, VICReg, NNCLR, LeJEPA).

from stable_pretraining.data import transforms, gpu_transforms as gt
import stable_pretraining as spt

# CPU side: just decode + resize. rgb=True folds in the old RGB() call.
cpu_transform = transforms.Compose(
    transforms.Resize((256, 256)),
    transforms.ToImage(rgb=True, scale=True),
)

# GPU side: same six augs, batched on device.
train_aug = gt.StackedMultiView(
    gt.GPUCompose([
        gt.GPURandomResizedCrop(size=224, scale=(0.08, 1.0)),
        gt.GPURandomHorizontalFlip(p=0.5),
        gt.GPUColorJitter(0.4, 0.4, 0.2, 0.1, p=0.8),
        gt.GPURandomGrayscale(p=0.2),
        gt.GPUGaussianBlur(kernel_size=23, p=0.5),
        gt.GPUNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]),
    n_views=2,        # use 8 for LeJEPA multi-view; or use gt.MultiView([c1, c2, ...]) for asymmetric methods like BYOL / DINO
)

train_ds = spt.data.HFDataset(
    "frgfm/imagenette", split="train",
    transform=cpu_transform,
    gpu_transform=train_aug,   # ← new arg; everything else stays the same
)
# val_ds: no gpu_transform — validation is already normalized on CPU.

What changed

Module.on_after_batch_transfer discovers gpu_transform through the active DataLoader. Resolution order: dataset.gpu_transform (1) → datamodule.gpu_transform (2, callable or {"train": ..., "val": ...} dict for third-party datasets you can't modify). Setting it on the Module is rejected at on_train_start — two routes to the same thing is a footgun. The resolved transform is auto-moved to self.device on first use (DDP-safe) and dropped during dataset pickling so DataLoader workers never serialise the nn.Module.

2 - Module

The key differentiator from PyTorch Lightning - you only define the forward function, not training_step! This unified approach computes losses and generates useful quantities that can be retrieved for monitoring and analysis:

# Use the pre-built forward functions from stable_pretraining
from stable_pretraining import forward

# Simply use the appropriate forward for your method
module = spt.Module(
    backbone=backbone,
    projector=projector,
    forward=forward.simclr,  # Or byol, vicreg, etc.
    simclr_loss=spt.losses.NTXEntLoss(temperature=0.5),
    optim={
        "optimizer": {"type": "Adam", "lr": 0.001},
        "scheduler": {"type": "CosineAnnealingLR"},
        "interval": "epoch"
    }
)

Or define your own custom forward:

def forward(self, batch, stage):
    out = {}

    if isinstance(batch, list):
        # Multi-view training - batch is a list of view dicts
        embeddings = [self.backbone(view["image"]) for view in batch]
        out["embedding"] = torch.cat(embeddings, dim=0)

        if self.training:
            projections = [self.projector(emb) for emb in embeddings]
            out["loss"] = self.simclr_loss(projections[0], projections[1])
    else:
        # Single-view validation
        out["embedding"] = self.backbone(batch["image"])

    return out

Key points:

• The forward method defines both the loss and any quantities to monitor
• No need to override training_step, validation_step, etc.
• Return a dictionary with a "loss" key for training
• All model components are passed as kwargs to spt.Module

3 - Callbacks

Monitor and evaluate your models in real-time during training. Callbacks are key ingredients of stable-pretraining, providing rich insights without interrupting your training flow.

Evaluation & Monitoring

Training Utilities

Checkpointing & Export

System & Logging

Intelligent Queue System

Callbacks that need rolling feature stores (OnlineKNN, RankMe, LiDAR, LatentViz) share memory through an automatic queue management system. If two callbacks monitor the same key with different queue lengths, a single queue is allocated at the maximum length and shared, eliminating redundant computation.

Why callbacks matter: Get real-time feedback on representation quality, catch issues like collapse early, and track multiple metrics simultaneously. For detailed usage and practical considerations, see the Callback guide.

Example:

# Monitor SSL representations with a linear classifier
linear_probe = spt.callbacks.OnlineProbe(
    module,
    name="linear_probe",
    input="embedding",
    target="label",
    probe=torch.nn.Linear(512, 10),
    loss_fn=torch.nn.CrossEntropyLoss(),
    metrics={
        "top1": torchmetrics.classification.MulticlassAccuracy(10),
        "top5": torchmetrics.classification.MulticlassAccuracy(10, top_k=5),
    },
)

# Track representation quality with KNN evaluation
knn_probe = spt.callbacks.OnlineKNN(
    name="knn_probe",
    input="embedding",
    target="label",
    queue_length=20000,
    k=10,
)

4 - Trainer

Orchestrate everything together with PyTorch Lightning's Trainer:

trainer = pl.Trainer(
    max_epochs=10,
    num_sanity_val_steps=1,
    callbacks=[linear_probe, knn_probe, rankme],  # Your monitoring callbacks
    precision="16-mixed",
    logger=False,
    enable_checkpointing=False,
)
manager = spt.Manager(trainer=trainer, module=module, data=data)
manager()

Once configured, the Manager connects all components and handles the training loop with precise logging and monitoring (optional).

Global Configuration

Instead of scattering options across environment variables, callback constructors, and factory functions, stable-pretraining provides a single entry-point to configure library-wide behavior:

import stable_pretraining as spt

spt.set(
    verbose="WARNING",                          # Global log level (also controls callback verbosity)
    progress_bar="rich",                        # "auto", "rich", "simple", or "none"
    cleanup={"checkpoints": False, "slurm": False},  # What CleanUpCallback removes
    log_rank="all",                             # Which distributed rank(s) may log (default: 0)
    default_callbacks={"env_dump": False},       # Toggle individual default callbacks on/off
)

# Inspect the current configuration
print(spt.get_config())

Settings apply immediately and persist for the process lifetime. spt.set() can be called multiple times; only the settings you pass are updated.

(output-directory-cache-dir)=

Output Directory (cache_dir)

By default, Lightning and Hydra scatter training outputs (checkpoints, logs, wandb data) based on the current working directory or Hydra's run.dir. This causes collisions when multiple sweep jobs start at the same time and resolve to the same path.

stable-pretraining solves this with a centralized cache_dir. When set, every run gets its own unique directory and all outputs are routed there automatically:

{cache_dir}/runs/{YYYYMMDD}/{HHMMSS}/{run_id}/
├── checkpoints/last.ckpt
├── wandb_resume.json
├── run_meta.json
├── environment.json
└── ...

Enabling cache_dir

import stable_pretraining as spt

# Option 1: in Python
spt.set(cache_dir="~/.cache/stable_pretraining")

# Option 2: via environment variable (e.g. in ~/.bashrc)
# export SPT_CACHE_DIR=~/.cache/stable_pretraining

When cache_dir is set, the Manager:

1. Creates a unique run directory under cache_dir/runs/.
2. Sets the Trainer's default_root_dir to that directory (before instantiation).
3. Redirects all ModelCheckpoint callbacks to run_dir/checkpoints/ (preserving their filename, monitor, and other settings).
4. Adds a requeue checkpoint (last.ckpt, saved every epoch) for seamless SLURM preemption recovery. You never need to add one yourself.
5. Routes all callback outputs (environment dumps, latent visualizations, HuggingFace exports, etc.) there.

If preemption is not a concern and you want to skip the requeue checkpoint overhead:

spt.set(cache_dir="/scratch/runs", requeue_checkpoint=False)

When cache_dir is not set (None, the default), the library behaves exactly as before.

How the run ID is generated

The run ID is chosen to be deterministic across all ranks of the same job, so multi-GPU training always agrees on a single directory:

ckpt_path is for loading only

ckpt_path and cache_dir serve different purposes:

• ckpt_path = where to load weights from (one-time, read-only).
• cache_dir = where to save everything going forward.

manager = spt.Manager(
    trainer=trainer_cfg,
    module=module,
    data=data,
    ckpt_path="/old/run/pretrained.ckpt",  # Load from here once
)
# New checkpoints, logs, wandb data → cache_dir/runs/.../

If you don't pass ckpt_path, the system checks run_dir/checkpoints/last.ckpt automatically. This means SLURM requeue works without any user configuration: the job is preempted, restarted with the same SLURM_JOB_ID, finds its previous run directory, and resumes from the last checkpoint.

Hydra compatibility

When cache_dir is active, Hydra's run.dir, sweep.dir, and job.chdir settings are ignored for trainer outputs (a warning is logged). Hydra still manages its own .hydra/ config dumps as usual. Note that SLURM .out/.err files are created by the scheduler before Python starts and cannot be redirected into the run directory.

Run Registry

When cache_dir is set, stable-pretraining automatically maintains a filesystem-backed run registry that indexes every run. Think of it as a local, offline, instant-query alternative to the wandb dashboard — designed for large sweeps on HPC clusters.

There is nothing to configure — if cache_dir is set, the registry is active:

import stable_pretraining as spt

spt.set(cache_dir="/scratch/runs")
# That's it. Every run writes a sidecar.json into its run directory.

Architecture

The registry uses a filesystem-first sidecar pattern. During training, each run writes only plain files — no SQLite, no network I/O:

{run_dir}/
  sidecar.json    ← atomic JSON snapshot (hparams, summary, status, tags)
  heartbeat       ← empty file, mtime touched every flush (liveness signal)
  metrics.csv     ← CSVLogger per-step time series
  hparams.yaml    ← CSVLogger hparams
  checkpoints/    ← Lightning checkpoints

The sidecar.json is the source of truth for each run. It is atomically rewritten (tmp + fsync + rename) so readers never see a partial file. A separate scanner (spt registry scan) walks {cache_dir}/runs/**/sidecar.json and builds a SQLite cache (registry.db) for fast querying. This cache is fully derived — deleting it is harmless; run spt registry scan --full to rebuild.

This design eliminates all SQLite contention during training: thousands of concurrent SLURM jobs write only to their own run directory and never touch a shared database.

What gets stored

The registry captures three categories of data per run, all automatically:

• Config / hparams — the full Hydra config (trainer, module, data) is flattened into dot-separated keys (e.g. module.optim.optimizer.lr, trainer.max_epochs) and stored as both config and hparams. This works the same way as wandb's config: the Manager flattens the Hydra DictConfig and injects it into module.save_hyperparameters() before trainer.fit(), so Lightning's built-in _log_hyperparams sends it to all loggers (wandb, CSV, TensorBoard, and the registry) automatically.
• Summary — every self.log() call in your LightningModule accumulates into a wandb-style summary dict (last value per metric key). At the end of training, the final summary (e.g. {"val_acc": 0.85, "train_loss": 0.12}) is written to the sidecar.
• Metadata — run ID, status (running/completed/failed/orphaned), liveness (heartbeat-based), tags, notes, run_dir path, and best checkpoint path.

Grouping with tags

All grouping is done through tags — a flat list of strings attached to each run. There is no separate "project" or "group" concept; cache_dir already acts as the project, and tags handle everything else.

For SLURM array jobs, a "sweep:<SLURM_ARRAY_JOB_ID>" tag is automatically added so that all tasks in the same array are queryable as a group. You can add your own tags in YAML:

logger:
  - _target_: stable_pretraining.registry.RegistryLogger
    tags: [resnet50, simclr, ablation-v2]
    notes: "Testing higher learning rates"

Querying runs

open_registry() triggers an incremental scan of the filesystem before returning, so the cache reflects the current state. A short in-process TTL ensures back-to-back queries don't re-scan.

import stable_pretraining as spt

spt.set(cache_dir="/scratch/runs")
reg = spt.open_registry()

# All completed runs from a SLURM array sweep
best = reg.query(tag="sweep:12345", status="completed", sort_by="summary.val_acc", limit=5)
for r in best:
    print(f"{r.run_id}: val_acc={r.summary['val_acc']:.3f}  lr={r.hparams['module.optim.optimizer.lr']}")

# Load the best checkpoint directly
import torch
ckpt = torch.load(best[0].checkpoint_path)

# Filter by any Hydra config key (deeply nested keys work)
lars_runs = reg.query(hparams={"module.optim.optimizer.type": "LARS"})

# Check which runs are still alive (heartbeat-based)
active = reg.query(alive=True)

# All resnet runs as a pandas DataFrame
df = reg.to_dataframe(tag="resnet50")
# Columns include flattened hparams and summary:
#   run_id, status, alive, tags, notes, checkpoint_path,
#   hparams.module.optim.optimizer.lr, hparams.trainer.max_epochs,
#   summary.val_acc, summary.train_loss, ...

# Quick analysis
df[["run_id", "hparams.module.optim.optimizer.lr", "summary.val_acc"]].sort_values(
    "summary.val_acc", ascending=False
).head(10)

Concurrency and SLURM requeue

Since training jobs only write to their own run directory (plain files, no shared database), there is zero SQLite contention — thousands of concurrent SLURM jobs run safely with no coordination. The scanner is the sole SQLite writer and uses WAL mode so concurrent readers never block it. On SLURM requeue, the run ID is deterministic (derived from SLURM_JOB_ID), so the requeued job finds its previous sidecar and resumes seamlessly — the same mechanism used for wandb run resumption and checkpoint recovery.

Liveness detection

The registry tracks whether a run is alive via a heartbeat file (an empty file whose mtime is touched on every log_metrics call). The scanner considers a run alive if its heartbeat is newer than 180 seconds and its status is not terminal. This lets you distinguish running, stalled, and dead jobs without contacting SLURM:

# Only alive runs
spt registry ls --alive

# Only dead/finished runs
spt registry ls --dead

CLI

The spt registry command lets you query runs from the terminal:

# List all runs
spt registry ls

# Filter by tag, status, or liveness
spt registry ls --tag resnet50 --status completed
spt registry ls --alive

# Top 5 runs by a metric (use --asc for losses)
spt registry best val_acc
spt registry best train_loss --asc -n 10

# Show full details for a run (config, summary, tags)
spt registry show <run_id>

# Export to CSV or Parquet
spt registry export sweep_results.csv --tag sweep:12345

# Manually refresh the SQLite cache from sidecars
spt registry scan
spt registry scan --full  # re-ingest everything (schema migration, DB rebuild)

# Migrate from a legacy server-backed registry.db
spt registry migrate /path/to/old/registry.db --cache-dir /scratch/runs

By default the CLI uses $SPT_CACHE_DIR/registry.db (or the spt.set(cache_dir=...) value). Pass --db /path/to/registry.db and/or --cache-dir to override.

Disabling the registry

spt.set(default_loggers={"registry": False})

Built-in Methods

stable-pretraining ships with ready-to-use forward functions and matching loss functions for popular self-supervised learning methods:

The table above covers forward functions for use with spt.Module. For 30 full LightningModule implementations (BEiT, CMAE, Data2Vec, iBOT, iGPT, IJEPA, LeJEPA, MAE, MaskFeat, MIMRefiner, MoCov2, MoCov3, MSN, PIRL, SimMIM, SimSiam, TiCO, VICRegL, WMSE, and more), see METHODS.md and stable_pretraining/methods/.

Backbones

Load architectures from popular libraries or use built-in components:

# From torchvision
backbone = spt.backbone.from_torchvision("resnet50")

# From timm (thousands of pretrained models)
backbone = spt.backbone.from_timm("vit_base_patch16_224")

# From HuggingFace
backbone = spt.backbone.vit_hf("google/vit-base-patch16-224")

Additional building blocks: MLP, ConvMixer, Resnet9, TeacherStudentWrapper (EMA), MAEDecoder, MaskedEncoder, FlexibleTransformer, PatchMasking, IJEPAMasking, MultiBlockMasking, LinearProbe, AutoLinearClassifier, AutoTuneMLP, and more.

Optimizers & Schedulers

module = spt.Module(
    ...,
    optim={
        "optimizer": {"type": "LARS", "lr": 5, "weight_decay": 1e-6},
        "scheduler": {"type": "LinearWarmupCosineAnnealing"},
        "interval": "epoch",
    },
)

Complete Example

import lightning as pl
import torch
import torchmetrics
import torchvision
from torch import nn
from lightning.pytorch.loggers import WandbLogger

import stable_pretraining as spt
from stable_pretraining import forward
from stable_pretraining.data import transforms

# Define augmentations for SimCLR (creates 2 views of each image)
simclr_transform = transforms.MultiViewTransform(
    [
        transforms.Compose(
            transforms.RGB(),
            transforms.RandomResizedCrop((32, 32), scale=(0.2, 1.0)),
            transforms.RandomHorizontalFlip(p=0.5),
            transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1, p=0.8),
            transforms.RandomGrayscale(p=0.2),
            transforms.ToImage(**spt.data.static.CIFAR10),
        ),
        # Second view with slightly different augmentations
        transforms.Compose(
            transforms.RGB(),
            transforms.RandomResizedCrop((32, 32), scale=(0.08, 1.0)),
            transforms.RandomHorizontalFlip(p=0.5),
            transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1, p=0.8),
            transforms.RandomGrayscale(p=0.2),
            transforms.RandomSolarize(threshold=0.5, p=0.2),
            transforms.ToImage(**spt.data.static.CIFAR10),
        ),
    ]
)

# Load CIFAR-10 and wrap in dictionary format
cifar_train = torchvision.datasets.CIFAR10(root="./data", train=True, download=True)
cifar_val = torchvision.datasets.CIFAR10(root="./data", train=False, download=True)

train_dataset = spt.data.FromTorchDataset(
    cifar_train,
    names=["image", "label"],  # Convert tuple to dictionary
    transform=simclr_transform,
)

val_dataset = spt.data.FromTorchDataset(
    cifar_val,
    names=["image", "label"],
    transform=transforms.Compose(
        transforms.RGB(),
        transforms.Resize((32, 32)),
        transforms.ToImage(**spt.data.static.CIFAR10),
    ),
)

# Create dataloaders - MultiViewTransform handles the view creation
train_dataloader = torch.utils.data.DataLoader(
    dataset=train_dataset,
    batch_size=256,
    num_workers=8,
    drop_last=True,
    shuffle=True,  # Simple shuffle, no RepeatedRandomSampler needed
)

val_dataloader = torch.utils.data.DataLoader(
    dataset=val_dataset,
    batch_size=256,
    num_workers=10,
)

data = spt.data.DataModule(train=train_dataloader, val=val_dataloader)

# Build model components
backbone = spt.backbone.from_torchvision("resnet18", low_resolution=True)
backbone.fc = torch.nn.Identity()  # Remove classification head

projector = nn.Sequential(
    nn.Linear(512, 2048),
    nn.BatchNorm1d(2048),
    nn.ReLU(inplace=True),
    nn.Linear(2048, 2048),
    nn.BatchNorm1d(2048),
    nn.ReLU(inplace=True),
    nn.Linear(2048, 256),
)

# Create the module using the built-in SimCLR forward function
module = spt.Module(
    backbone=backbone,
    projector=projector,
    forward=forward.simclr,  # Use the built-in forward function
    simclr_loss=spt.losses.NTXEntLoss(temperature=0.5),
    optim={
        "optimizer": {"type": "LARS", "lr": 5, "weight_decay": 1e-6},
        "scheduler": {"type": "LinearWarmupCosineAnnealing"},
        "interval": "epoch",
    },
)

# Add callbacks for monitoring performance during training
linear_probe = spt.callbacks.OnlineProbe(
    module,
    name="linear_probe",
    input="embedding",
    target="label",
    probe=torch.nn.Linear(512, 10),
    loss_fn=torch.nn.CrossEntropyLoss(),
    metrics={
        "top1": torchmetrics.classification.MulticlassAccuracy(10),
        "top5": torchmetrics.classification.MulticlassAccuracy(10, top_k=5),
    },
)

knn_probe = spt.callbacks.OnlineKNN(
    name="knn_probe",
    input="embedding",
    target="label",
    queue_length=20000,
    metrics={"accuracy": torchmetrics.classification.MulticlassAccuracy(10)},
    input_dim=512,
    k=10,
)

# Configure training
trainer = pl.Trainer(
    max_epochs=1000,
    callbacks=[knn_probe, linear_probe],  # Monitor SSL quality in real-time
    precision="16-mixed",
    logger=WandbLogger(project="cifar10-simclr"),
)

# Launch training
manager = spt.Manager(trainer=trainer, module=module, data=data)
manager()

Quick Start with spt CLI

The spt command launches training from YAML configuration files using Hydra.

Note: spt requires YAML configs. If you have Python-based configs, you can:

• Convert them to YAML format where each component uses _target_ to specify the importable class/function
• See examples/simclr_cifar10_config.yaml for the structure and syntax

Local Training

# Run with a config file
spt examples/simclr_cifar10_config.yaml

# With parameter overrides
spt examples/simclr_cifar10_config.yaml trainer.max_epochs=50 module.optim.lr=0.01

# Run from any directory - supports absolute and relative paths
spt ../configs/my_config.yaml
spt /path/to/config.yaml

SLURM Cluster Training

For training on SLURM clusters, use the -m flag to enable multirun mode:

# Use the provided SLURM template (customize partition/QOS in the file)
spt examples/simclr_cifar10_slurm.yaml -m

# Override SLURM parameters via command line
spt examples/simclr_cifar10_slurm.yaml -m \
    hydra.launcher.partition=gpu \
    hydra.launcher.qos=normal \
    hydra.launcher.timeout_min=720

The SLURM template (examples/simclr_cifar10_slurm.yaml) includes placeholders for cluster-specific settings. Either modify the file directly or override values via command line.

Installation

The library is not yet available on PyPI. You can install it from the source code, as follows.

1. conda (optional)

   First use your favorite environment manager and install your favorite pytorch version, we provide an example with conda

   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
   bash Miniconda3-latest-Linux-x86_64.sh
   

   follow installation instructions... once completed, create your environment

   conda create -n my_env python=3.11
   

   with your environment name (here my_env) and your favorite Python version (here, 3.11). Once completed, make sure to activate your environment (conda activate my_env) before proceeding to the next steps!

2. Pytorch and our library (we recommend using uv for quicker package management):

   pip3 install uv
   uv pip install torch torchvision torchaudio
   uv pip install -e .  # Core dependencies only
   

   For optional features (vision models, experiment tracking, cluster support, etc.):

   uv pip install -e ".[vision,tracking]"  # Example: add vision models and wandb
   uv pip install -e ".[all]"  # Or install all optional dependencies
   

   See pyproject.toml for available dependency groups (vision, tracking, cluster, visualization, datasets, extras, dev, doc).

   If you do not want to use uv, simply remove it from the above commands.

3. API login (optional)

   wandb login
   huggingface-cli login
   

4. LaTeX support in Matplotlib (optional)

   Click to expand setup instructions

   Install TeX Live (minimal, no sudo):

   cd /tmp
   wget https://mirror.ctan.org/systems/texlive/tlnet/install-tl-unx.tar.gz
   tar xzf install-tl-unx.tar.gz
   cd install-tl-*/
   ./install-tl --texdir ~/texlive --no-interaction --scheme=scheme-basic
   

   Add to your ~/.bashrc (or equivalent):

   export PATH="$HOME/texlive/bin/x86_64-linux:$PATH"
   

   Install required LaTeX packages. Pin a known-good CTAN mirror first (the default redirector occasionally serves corrupt files or unreachable hosts), then install one package per line so any failure is visible:

   tlmgr option repository https://ctan.math.illinois.edu/systems/texlive/tlnet
   for pkg in type1cm cm-super dvipng collection-fontsrecommended \
              amsmath amsfonts tools underscore xcolor iftex epstopdf-pkg; do
     tlmgr install "$pkg" || echo "FAILED: $pkg"
   done
   

   Notes on the package list:

   • amsfonts provides amssymb.sty; tools provides bm.sty; iftex provides ifvtex.sty. Naming these as amssymb/bm/ifvtex directly will fail with "package not present in repository".
   • If tlmgr reports checksum mismatches, switch to a different mirror (e.g. https://mirror.ox.ac.uk/sites/ctan.org/systems/texlive/tlnet) and rerun.

   Computer Modern TTF fonts (only if you also use usetex=False): with usetex=True, all matplotlib text is rendered by LaTeX using TeX Live's bundled fonts and this step is unnecessary. Install the TTFs only if you want CMU for matplotlib's own text rendering (no LaTeX roundtrip):

   mkdir -p ~/.local/share/fonts
   cp assets/cm-unicode-0.7.0\ 2/*.ttf ~/.local/share/fonts/
   fc-cache -f -v   # requires fontconfig; skip if unavailable
   # verify: fc-list | grep cmu
   

   Clear matplotlib font cache (after either step above):

   python -c "import shutil, matplotlib; shutil.rmtree(matplotlib.get_cachedir(), ignore_errors=True)"
   

   Verify (heredoc with single-quoted 'PY' so the shell does not touch $):

   python - <<'PY'
   import matplotlib; matplotlib.use('Agg')
   import matplotlib.pyplot as plt
   plt.rc('text', usetex=True)
   plt.rc('font', family='serif')
   plt.figure(); plt.title(r'$\sum_{i=1}^n x_i$')
   plt.savefig('/tmp/tex_test.png')
   print('Success!')
   PY
   

Ways You Can Contribute:

• If you'd like to contribute new features, bug fixes, or improvements to the documentation, please refer to our contributing guide for detailed instructions on how to get started.

• You can also contribute by adding new methods, datasets, or configurations that improve the current performance of a method in the benchmark section.

Citation

If you use stable-pretraining in your research, please cite:

@article{balestriero2025stable,
  title={stable-pretraining-v1: Foundation Model Research Made Simple},
  author={Balestriero, Randall and Van Assel, Hugues and BuGhanem, Sami and Maes, Lucas},
  journal={arXiv preprint arXiv:2511.19484},
  year={2025}
}

Contributors

Core contributors (in order of joining the project):

• Randall Balestriero
• Hugues Van Assel
• Sami BuGhanem
• Lucas Maes

Benchmarks — Imagenette ViT-S/16, 200 epochs

Caveat. These are out-of-the-box results on Imagenette (10-class subset of ImageNet, 224×224) at 200 epochs with ViT-S/16, using each method's paper-default ImageNet-1k hyperparameters scaled to the batch size used here. No per-method hyperparameter tuning was performed. Several methods (e.g. MIM-family at short schedules, whitening methods at small batch) are known to need longer training or larger batches to reach their headline accuracy. Treat the table as a sanity-check sweep over the 27 method classes, not a SOTA comparison.

Online linear probe and 20-NN probe top-1 accuracy on Imagenette, single A100, no W&B. Full details, hyperparameters per method, and reproduction commands: benchmarks/imagenet10/RESULTS.md.

* = run still climbing at time of writing.